© Copyright 2025 

Rui Chen 

  



 

 

Programmable nano-photonics with electrically controlled phase-change materials 

 

 

Rui Chen 

 

 

 

 

A dissertation 

 

submitted in partial fulfillment of the 

 

requirements for the degree of 

 

 

 

Doctor of Philosophy 

 

 

 

 

University of Washington 

 

2025 

 

 

 

 

Reading Committee: 

 

Arka Majumdar, Chair 

Mo Li 

Sajjad Moazeni 

 

 

 

 

Program Authorized to Offer Degree:  

 

Electrical and Computer Engineering 

  



 

 

 

University of Washington 

 

Abstract 

 

 

 Programmable nano-photonics with electrically controlled phase-change materials 

 

 

Rui Chen 

 

 

Chair of the Supervisory Committee: 

Arka Majumdar 

Department of Electrical and Computer Engineering 

 

Programmable nano-photonics have the potential to completely transform a range of emerging 

applications, including optical computing, optical signal processing, light detecting and ranging, and 

quantum applications. However, implementing energy-efficient and large-scale systems remains elusive 

because commonly used programmable photonic approaches are volatile and energy-hungry. Recent results 

on non-volatile phase-change material integrated photonics present a promising opportunity to create truly 

programmable nano-photonics. The ability to drastically change the refractive index of the PCMs in a non-

volatile fashion allows creating programmable units with zero-static energy. By taking advantage of the 

electrical control, non-volatile reconfiguration, and zero crosstalk between each unit, PCMs can enable both 

extremely large-scale integrated photonics and metasurfaces.  

In this dissertation, we present our main progress in PCM nano-photonics and discuss the challenges 

and limitations of this emerging technology. We first demonstrated 2 × 2 electrically programmable units 

using a well-studied, prototypical PCM Ge2Sb2Te5 and scalable doped silicon PIN heaters. These 

components exhibit low insertion loss (~2 dB), high extinction ratio (8 dB), and large endurance (> 2,800 

cycles), and are critical for large-scale photonic networks and photonic field programmable arrays. We also 



 

 

designed a three-waveguide directional coupler multiple operation levels by adding a detuning parameter 

between the center and the other waveguides. To further reduce the loss, we explored an emerging PCM 

Sb2S3 with lower absorption loss and demonstrated phase shifters in both micro-ring resonators and Mach-

Zehnder interferometers. An asymmetric directional coupler with two waveguides is also experimentally 

demonstrated, showing better loss (~1 dB) and extinction ratio (~ 15 dB) than the GST one. Interestingly, 

the device also supported a stepwise multi-level switching behavior, which offered unique opportunity for 

deterministic tuning, and we showed 32 levels with excellent repeatability. With this multilevel operation, 

we demonstrate an application of post-fabrication trimming to correct the phase error in a balanced MZI. 

Beyond the university level, in-house fabrication, which relies on electron beam lithography and has limited 

scalability, we also cooperated with Intel and showed a scalable process to put PCMs on commercial 300-

mm silicon photonic wafers.  

Active metasurfaces are another emerging field in nanophotonics. However, current electrically 

controlled PCM-based metasurfaces are limited to global amplitude modulation, insufficient for SLMs. We 

demonstrated an individual-pixel addressable, transmissive metasurface using the low-loss PCM Sb2Se3 

and doped silicon nanowire heaters. The nanowires simultaneously form a diatomic metasurface, 

supporting a high quality-factor (~406) quasi-bound-state-in-the-continuum mode. Global phase-only 

modulation of ~0.25π (~0.2π) in simulation (experiment) is achieved, showing ten times enhancement. 2π 

phase shift is further obtained using a guided-mode resonance with enhanced light-Sb2Se3 interaction. 

Finally, individual-pixel addressability and SLM functionality are demonstrated through deterministic 

multilevel switching (ten levels) and tunable far-field beam shaping. Our work presents zero-static power 

transmissive phase-only SLMs, enabled by electrically controlled low-loss PCMs and individual meta-

molecule addressable metasurfaces. 

As a perspective, we argue that energy efficiency is a more critical parameter than the operating speed 

for programmable photonics, making PCMs an ideal candidate. This has the potential for a disruptive 

paradigm shift in the reconfigurable photonics research philosophy, as slow but energy-efficient and large 

index modulation can provide a better solution for extremely large-scale integrated photonics than fast but 



 

 

power-hungry, small index tuning methods. We also highlight the exciting opportunities to leverage wide 

bandgap PCMs for visible-wavelength applications, such as quantum photonics and optogenetics, and for 

rewritable photonic integrated circuits using nanosecond pulsed lasers. The latter can dramatically reduce 

the fabrication cost of PICs and democratize the PIC manufacturing process for rapid prototyping. 

  



 

 

TABLE OF CONTENTS 

List of Figures ................................................................................................................................. 9 

List of Tables ................................................................................................................................ 15 

ACKNOWLEDGEMENTS .......................................................................................................... 16 

Chapter 1. Introduction ................................................................................................................. 18 

Chapter 2. Broadband electrically controlled SOI switches based on GST ................................. 26 

2.1 Optical refractive index characteristics of GST ............................................................ 26 

2.2 Design of programmable 1 × 2 photonic programmable units using GST ................... 27 

2.3 Design of the 2 × 2 programmable photonic units based on GST ................................ 30 

2.4 Fabrication variation tolerance analysis........................................................................ 33 

2.5 Heat transfer simulation ................................................................................................ 34 

2.6 Device fabrication ......................................................................................................... 36 

2.7 Device characterization ................................................................................................. 39 

2.7.1 Characterization setups ......................................................................................... 39 

2.7.2 Results for 1 × 1 waveguide switch ...................................................................... 40 

2.7.3 Results for 1 × 2 silicon photonic programmable units ........................................ 42 

2.7.4 Results for 2 × 2 silicon photonic programmable units ........................................ 43 

2.7.5 Transient response of the 2 × 2  programmable unit ............................................ 47 

2.7.6 Thermal stability test results ................................................................................. 48 

2.8 Deterministic multilevel operation with multiple GST segments................................. 49 

2.8.1 1 × 2 programmable units ..................................................................................... 49 

2.8.2 2 × 2 programmable unit ....................................................................................... 51 



 

 

2.9 System-level applications envision ............................................................................... 58 

Chapter 3. Electrically programmable multi-level PICs with Sb2S3............................................. 61 

3.1 Characterization of Sb2S3 thin films ............................................................................. 61 

3.2 Simulation and design of Sb2S3-based silicon photonic components ........................... 63 

3.2.1 Sb2S3-based silicon phase shifters ........................................................................ 63 

3.2.2 Sb2S3-based silicon asymmetric directional coupler............................................. 65 

3.3 Fabricated devices and characterization results ............................................................ 68 

3.3.1 Trench effect of PCM liftoff ................................................................................. 69 

3.3.2 Nonvolatile micro-ring switch integrated with Sb2S3 phase shifter...................... 69 

3.3.3 Nonvolatile Mach-Zehnder switch integrated with Sb2S3 phase shifter ............... 72 

3.3.4 Asymmetric directional coupler silicon switch with Sb2S3 .................................. 76 

3.4 Multilevel 5-bit operation with dynamic electrical control .......................................... 84 

3.5 Random phase error correction in balanced MZIs exploiting multilevel operation ..... 89 

3.6 Discussion ..................................................................................................................... 92 

Chapter 4. PCM integration in high-volume silicon photonics .................................................... 94 

4.1 Reproducible zero-change integration of PCMs on silicon photonics from a 300-mm 

fab 94 

4.2 Characterization results ................................................................................................. 99 

4.2.1 Experimental measurement results ....................................................................... 99 

4.3 Asymmetric directional couplers with multiple Sb2S3 segments ................................ 105 

4.3.1 Improved design procedure................................................................................. 105 

4.4 Quasi-continuously tunable MZIs with equal and unequal-length multi-segments 

schemes ................................................................................................................................... 112 



 

 

4.5 Discussion ................................................................................................................... 115 

Chapter 5. Nonvolatile transmissive spatial light modulator with PCMs ................................... 119 

5.1 High-Q silicon diatomic metasurfaces ........................................................................ 121 

5.2 A nonvolatile electrically reconfigurable metasurface for phase-only control ........... 124 

5.3 Individually addressing meta-gratings ........................................................................ 131 

Chapter 6. Challenges and opportunities .................................................................................... 147 

6.1 High loss of traditional PCMs .................................................................................... 147 

6.2 Reliable multi-level operation with electrical control ................................................ 148 

6.3 Limited PCM endurance ............................................................................................. 150 

6.4 Difficulty identifying phase change conditions .......................................................... 152 

6.5 opportunities toward large-scale systems ................................................................... 153 

6.5.1 Quasi-arbitrary unitary transformation ............................................................... 154 

6.5.2 Adding optical redundancy with non-blocking switching fabric ........................ 155 

6.5.3 Optical neural networks ...................................................................................... 156 

6.5.4 Non-volatile electro-optical programmable gate array ....................................... 158 

6.6 Visible phase-change material photonics .................................................................... 159 

6.7 Laser rewriteable phase-change material integrated photonics .................................. 161 

 

 

  



 

 

LIST OF FIGURES 

Figure 1.1. A PCM optical property map. ................................................................................. 20 

Figure 1.2: PCM-based reconfigurable integrated photonics (top)49,52,80–83 and meta-optics 

(bottom)53,61–64 with optical (left) or electrical (right) controls. .............................................. 22 

Figure 2.1: The complex refractive index of a/c-GST measured by ellipsometry. ................. 27 

Figure 2.2: Broadband nonvolatile electrically controlled 1 × 2 programmable unit 

schematic and images ................................................................................................................. 28 

Figure 2.3: Mode simulation results for the 1 × 2 programmable unit. .................................. 30 

Figure 2.4: Mode simulation results for the 2 × 2 programmable units. ................................ 31 

Figure 2.5: FDTD simulation results for both 1 × 2 and 2 × 2 programmable units. ............ 32 

Figure 2.6: Insertion loss vs. fabrication variation study on the 2 × 2 programmable unit. . 33 

Figure 2.7: Extinction ratio vs. fabrication variation study on the 2 × 2 programmable unit.

....................................................................................................................................................... 34 

Figure 2.8: Transient heat transfer simulation results. ............................................................ 35 

Figure 2.9: Broadband non-volatile electrically controlled 2 × 2 programmable unit in 

silicon photonics. ......................................................................................................................... 38 

Figure 2.10: Measurement setup schematics. ............................................................................ 40 

Figure 2.11: Measurement results for the waveguide switch. .................................................. 41 

Figure 2.12: Optical transmission spectra measurement results for the 1 × 2 programmable 

unit. ............................................................................................................................................... 42 

Figure 2.13: The endurance test results for the 1 × 2 programmable unit. ............................ 43 

Figure 2.14: Characterizations of the electrically controlled 2 × 2 programmable unit. ...... 44 

Figure 2.15: Cyclability test of the electrically controlled 2 × 2 programmable unit. ........... 46 



 

 

Figure 2.16: SEM of the devices after multiple cycles. ............................................................. 47 

Figure 2.17: Transient response measurement results for electrically controlled 2 × 2 

programmable unit. .................................................................................................................... 48 

Figure 2.18: Thermal stability test measurement. .................................................................... 49 

Figure 2.19 FDTD simulation for a deterministic three-level 1 × 2 programmable unit using GST.

....................................................................................................................................................... 50 

Figure 2.20: Phase-matched three-WG 2 × 2 directional couplers have high loss for 

intermediate levels. ..................................................................................................................... 52 

Figure 2.21: Analytical model reveals that phase-matched three-WG couplers can never 

have low-loss intermediate levels. .............................................................................................. 53 

Figure 2.22: Detuning the middle waveguide provides rich optical intensity dynamics, 

enabling low-loss intermediate levels. ....................................................................................... 55 

Figure 2.23: Optimized deterministic three-level 2 × 2 programmable units with phase-detuned 

TDC design (𝛿 = 1.5𝜅). ............................................................................................................... 56 

Figure 2.24: Simulated transmission spectrum of the designed three-level 2 × 2 TDC 

programmable unit (𝜹 = 𝟏. 𝟓𝜿). ................................................................................................ 58 

Figure 2.25: PIC schematics for various applications. ............................................................. 60 

Figure 3.1: Sb2S3 material characterization. ............................................................................. 62 

Figure 3.2: Non-uniform surface forms after the annealing due to the density difference 

between amorphous and crystalline Sb2S3. ............................................................................... 63 

Figure 3.3: Simulations for Sb2S3-on-SOI phase shifters. ........................................................ 64 

Figure 3.4: Simulations for Sb2S3-on-SOI 2 × 2 tunable directional couplers with low 

insertion loss and large extinction ratios. ................................................................................. 66 



 

 

Figure 3.5: Phase matching a-Sb2S3-SOI waveguide with bare SOI waveguide results in an 

asymmetric (input port dependent) bar state when switching the Sb2S3 to its crystalline 

phase. ............................................................................................................................................ 67 

Figure 3.6: A high-Q ring resonator loaded with 10 μm long 20-nm-thick Sb2S3. ................. 71 

Figure 3.7: Single SET (RESET) pulses could not reliably trigger complete crystallization or 

amorphization. ............................................................................................................................ 72 

Figure 3.8: Low-loss 50:50 multimode interferometer (MMI) at 1310 nm. ........................... 73 

Figure 3.9: A Mach-Zehnder interferometer with both arms covered with 20-nm-thick 

Sb2S3. ............................................................................................................................................ 75 

Figure 3.10: Extra measurement results on MZIs. ................................................................... 76 

Figure 3.11: An asymmetric directional coupler with Sb2S3-Si hybrid waveguide. ............... 78 

Figure 3.12: Simulated transmission spectrum for asymmetric directional coupler. ............ 79 

Figure 3.13: Cyclability test for a-Sb2S3-based asymmetric directional coupler. ................... 80 

Figure 3.14: The slow crystallization nature of Sb2S3 limits the pulse duration longer than 

100 𝝁𝒔. .......................................................................................................................................... 81 

Figure 3.15: Complete amorphization was triggered with relatively long, 10-𝝁𝒔 pulses 

because of the slow crystallization nature of Sb2S3. ................................................................. 83 

Figure 3.16: Only slight device degradation occurred after 1,150 cycles (2,300 switching 

events), and the SEM images suggest it is due to thermal reflowing. .................................... 84 

Figure 3.17: A quasi-continuously tunable directional coupler based on multilevel Sb2S3. .. 86 

Figure 3.18: Stepwise partial crystallization using identical, low amplitude, and long 

duration pulses with a one-second interval. ............................................................................. 88 

Figure 3.19: Micrograph images for directional couplers in different intermediate levels... 89 



 

 

Figure 3.20: Random phase error correction in a balanced MZI based on the multilevel 

operation. ..................................................................................................................................... 91 

Figure 4.1: In-house fabrication workflow diagram................................................................. 95 

Figure 4.2: Schematic and photograph of the fabricated wafers and reticles. ....................... 98 

Figure 4.3: Measured optical transmission spectra for a micro-ring resonator with 20-μm-

long oxide window. ...................................................................................................................... 98 

Figure 4.4: Reversible reconfiguration of micro-ring resonators and Mach-Zehnder 

interferometers through electrically controlled 20-nm-thick Sb2S3 thin films. ................... 100 

Figure 4.5: Numerical simulation results for 𝝅 phase shift length 𝑳𝝅 and loss per 𝝅 when 

varying waveguide width and Sb2S3 thickness. ...................................................................... 101 

Figure 4.6: Numerical simulation results for 𝝅 phase shift length 𝑳𝝅 when varying Sb2S3 

thickness and waveguide height. .............................................................................................. 102 

Figure 4.7: Measured I-V curve of the micro-ring resonator and Mach-Zehnder 

interferometer. .......................................................................................................................... 103 

Figure 4.8: Endurance test of an unbalanced MZI. ............................................................... 104 

Figure 4.9: Schematic of the Sb2S3-based asymmetric directional coupler and main results 

during simulation. ..................................................................................................................... 107 

Figure 4.10: Simulated electric field distribution and transmission spectra for quasi-

continuously tunable beam splitters with 2 Sb2S3 segments. ................................................. 107 

Figure 4.11: Quasi-continuously tunable asymmetric directional coupler with two 

individually controlled Sb2S3 segments. ...................................................................................110 

Figure 4.12: The I-V curve before and after the PIN diode heater was switched for 26,300 

cycles............................................................................................................................................ 111 



 

 

Figure 4.13: The scanning electron microscope images after switching 50,000 switching 

events. ..........................................................................................................................................112 

Figure 4.14: Quasi-continuously tunable Mach-Zehnder interferometers with four equal 

and unequal Sb2S3 segments. ....................................................................................................114 

Figure 5.1: High-Q silicon diatomic metasurfaces. ................................................................. 123 

Figure 5.2: Electrical field amplitude of the diatomic resonance. ......................................... 124 

Figure 5.3: A nonvolatile electrically reconfigurable metasurface based on Sb2Se3. ........... 126 

Figure 5.4: Optical measurement setup for in situ electrical control of metasurface. ......... 127 

Figure 5.5: Reversible switching of the diatomic metasurface and nonvolatile phase-only 

modulation. ................................................................................................................................ 130 

Figure 5.6: Reversible switching of GST blanket film using doped Si heater. ..................... 131 

Figure 5.7: Transverse-electric polarized guided-mode resonance mode with strong field 

overlap with Sb2Se3, achieving larger resonance shift and a full 2π phase modulation. .... 132 

Figure 5.8: Spectrum of the Fianium broadband source combined with a grating. ........... 133 

Figure 5.9: Metal wire engineering to achieve equal resistance across 17 channels. ........... 134 

Figure 5.10: Electrically addressing individual meta-atoms. ................................................ 137 

Figure 5.11: Normalized transmission spectra showing evidence of thermal crosstalk-free 

performance during amorphization process. ......................................................................... 139 

Figure 5.12: Joule heating simulation in COMSOL for understanding of the thermal 

crosstalk-free performance during amorphization. ............................................................... 140 

Figure 5.13: Simulated resonance shift of the metasurface when different meta-atoms are 

switched. ..................................................................................................................................... 141 



 

 

Figure 5.14: Fitted Q-factor of the metasurface versus the number of switched meta-

molecules. ................................................................................................................................... 142 

Figure 5.15: Simulated energy-momentum spectrum for both resonance modes and a 

flatband metasurface design for wide input FOV. ................................................................. 143 

Figure 5.16: Tunable free-space beam focusing. ..................................................................... 144 

Figure 6.1: Quasi-continuously programmable phase shifters, photonic switches, and beam 

splitters with PCMs and PIN diode-based doped silicon micro-heaters on a silicon-on-

insulator (SOI) chip. ................................................................................................................. 149 

Figure 6.2: Schematic of non-volatile arbitrary unitary transformation and switching 

fabrics. ........................................................................................................................................ 156 

Figure 6.3: Schematic of PCM-based non-volatile platform to perform linear operations in 

an optical neural network. ....................................................................................................... 157 

Figure 6.4: Non-volatile programmable gate arrays based on PCMs. .................................. 158 

Figure 6.5: Non-volatile active photonics in the visible employing emerging wide bandgap 

PCM Sb2S3. ................................................................................................................................ 161 

Figure 6.6: Lithography-free rewritable PICs using ns-lasers and PCMs. .......................... 162 

  



 

 

LIST OF TABLES 

Table 1: FOM for different PCMs in 640nm and 1550nm ........................................................... 21 

Table 2: Designed parameters for the multi-level 2 × 2 GST programmable unit with 𝜹 =

𝟏. 𝟓𝜿. Unit: nm except for 𝒍𝒄. ..................................................................................................... 57 

Table 3: Performances of Sb2S3 directional couplers using different Sb2S3 thicknesses .............. 68 

Table 4: Sb2S3 thickness reduces after liftoff measured by AFM (Unit: nm) ............................... 69 

Table 5: PCM integrated photonic device performance comparison ...........................................117 

Table 6: Comparison of our Sb2Se3-based transmissive SLMs with other SLMs ...................... 145 

Table 7: PCM integrated photonic device performance comparison .......................................... 151 

Table 8: State-of-the-art PIC systems comparison ...................................................................... 154 

  



 

 

ACKNOWLEDGEMENTS 

I would like to extend my heartfelt gratitude to all the committee members for attending my dissertation 

defense. In particular, I want to express my deepest appreciation to my advisor, Prof. Arka Majumdar, for 

his unwavering support for my research. Our discussions about new research ideas and perspectives have 

always been inspiring and intellectually stimulating. He has cultivated a free and encouraging atmosphere 

for all group members to explore innovative ideas, while also providing exceptional experimental resources 

and interdisciplinary collaborations to realize and test them. He is particularly inspiring in terms of abstract 

ideas and philosophy, which has gradually become an important part of my methodology of logic thinking.  

I am sincerely thankful to Prof. Mo Li, whose photonics class laid the theoretical foundation for my 

research. Additionally, I appreciate the assistance of his group members, Dr. Changming Wu and Dr. 

Bingzhao Li, who have significantly supported my work in nanofabrication and device characterization. I 

also wish to thank Prof. Sajjad Moazeni for teaching EE 538B: Integrated Photonic Systems, which 

provided profound insights into state-of-the-art electro-optics integrated systems. The knowledge and skills 

I gained from this class have become increasingly important as my research focus shifts from device-level 

to system-level demonstrations. I am also grateful to Prof. David Bergsman for serving as the GSR on my 

committee and for providing meaningful feedback throughout the process. 

I would also like to thank my lab mates and friends from other research groups for their camaraderie, 

collaboration, and fruitful discussions. Special thanks to Dr. Roger (Zhuoran) Fang, Dr. Jiajiu Zheng, Mr. 

Virat Tara, Dr. Johannes Fröch, Dr. Jie Fang, Mr. Forrest Miller, Dr. Abhi Saxena, Dr. Quentin Tanguy, 

Dr. Yueyang Chen, Dr. Changming Wu, Dr. Bingzhao Li, Ms. Jayita Dutta, Mr. Andrew Tang, Dr. Cheng 

Chang, Dr. Zhendong Wang, Mr. Xi Wang, Mr. Yijie Wang, Mr. Songli Wang, Prof. Keren Bergman, Prof. 

Alex Gaeta, Prof. Michal Lipson and Prof. Alex (Xiang) Meng for their enormous support in my research 

and life. In particular, I want to acknowledge Dr. Roger Fang and Dr. Jiajiu Zheng for their countless and 

invaluable assistance and for their insightful discussions that have greatly enriched my research journey. I 

want to also express my deepest thanks to Prof. Alex Meng at Columbia University, who have been 

continuously teaching me both technical knowledge as well as a better view of our life and the world. 



 

 

I was honored to get an internship opportunity in Meta Platforms, Inc. during the summer of my 4th 

year and got to work together with several excellent peers and researchers, such as Dr. Zhujun Shi, Dr. 

Giuseppe Calafiore, Mr. Hanfeng Wang, Mr. Arnab Manna. I truly miss and value the time we spent 

discussing interesting yet practical ideas. Special thanks to my intern manager Dr. Zhujun Shi, who is 

extremely inspiring and very informative during my internship.  

I am grateful to the University of Washington for providing world-class facilities that made my research 

possible. A significant portion of my work was conducted at the Molecular Analysis Facility (MAF) and 

the Washington Nanofabrication Facility (WNF), and I greatly appreciate the resources and expertise the 

staff has offered, special thanks to Dr. N Shane Patrick, Dr. Darick Baker, and Dr. Mark Brunson for their 

help. 

Finally, I want to express my deepest thanks to my parents and my beloved partner, Ms. Zhaoyi Chen. 

Their unwavering understanding and support always recharge me and give me the inner strength to 

overcome seemingly impossible challenges. New journey is ahead of me. Surly they are the ones who are 

always by my side, and so am I. 



 

 

Chapter 1.  Introduction1  

Nano-photonics is a field that studies light-matter interaction where the structure is of nanometer scales, 

i.e., typically smaller than the optical wavelengths2,3. Study on such an extreme scale is only made possible 

in recent decades thanks to the advancements in electromagnetic simulation software4, as well as the rapid 

development of nanofabrication techniques. With nano-photonics technologies, a lot of applications have 

emerged, including optical imaging5–9, optical communication10,11, optical information processing in both 

classic12–16 and quantum regime17–21, light detection and ranging22–25, hyperspectral imaging26–28, 

miniaturized spectrometers29–32, just to name a few. At the heart of modern nano-photonics are photonic 

integrated circuits (PICs)33 and metasurfaces3. PICs are optical analogy of electronic integrated circuits 

(EICs), where it is light, instead of electrical current, that is directed on a small chip through various 

photonic components, such as integrated lasers, waveguides, couplers, phase shifters, amplitude 

modulators, and on-chip photodetectors. Instead of manipulating light on-chip, metasurfaces are for free-

space light wave manipulation, which are made of an array of carefully engineered nanoscale scatterers. 

One critical aspect of both PICs and metasurfaces is their programmability for their generic usage for optical 

information processing and optical networking applications. The goal of this dissertation is to provide some 

advancements in programmability for both PICs and metasurfaces. 

PICs are typically application-specific34, i.e., every fabricated chip serves only one particular function. 

In contrast, recent advancements in silicon photonics urgently call for universally programmable PICs35, 

adaptable to a wide range of tasks, including optical neural networks36, quantum information processing37, 

and light detection and ranging38. The fundamental building block constituting such programmable PICs is 

an electrically controlled 2 × 2 programmable unit or optical switch, that diverts light to one of two ports. 

In silicon photonics, the operation of such elements is commonly based on either the thermo-optic39 or free-

carrier effect40, which are, however, plagued by large footprints (typically > 100 µm) and/ or low energy 

efficiency due to their large static power consumption. Although micro-ring resonators can significantly 

reduce the footprint (down to the order of tens of µm), the operational optical bandwidth becomes limited 



 

 

to typically less than 1 nm41. Moreover, their thermal instability requires an even larger static power 

consumption, as a constant feedback signal is needed to lock the resonance11. Emerging technologies such 

as microelectromechanical systems (MEMS) 42, electro-optic polymer43, and integrated lithium niobate44 

can potentially provide either a smaller footprint or lower programming energy. Still, they are all based on 

volatile effects with limited CMOS compatibility. Phase-change materials (PCMs) provide an attractive 

solution towards compact and energy-efficient programmable units with zero static power45–48, thanks to a 

non-volatile phase transition49, large refractive index contrast (Δn≥1)50, and CMOS compatibility47. PCMs 

have two stable states in ambient environments: amorphous (a-PCM) and crystalline (c-PCM). The optical 

properties of these two states are drastically different. Therefore, by switching between them, the photonic 

device functionality can be changed. Notably, these two states can be reversibly switched with proper 

thermal pulses. In general, to change a PCM from the crystalline to the amorphous state, the PCM needs to 

be heated above its melting temperature Tm and quenched (cooled down with rates higher than 109𝐾 ⋅

𝑠−151). To change from amorphous to crystalline state requires heating the PCM above the glass transition 

temperature Tg but below Tm. The temperature must be held for a long time (hundreds of ns or longer) to 

allow crystal nucleation and growth.  

We compare different PCMs for three crucial material properties in Figure 1.1, and identify a tradeoff 

between larger index contrast Δn and increasing loss. For most photonic applications, we need high 

refractive index contrast Δn but low material absorption (𝜅𝑎  for amorphous and 𝜅𝑐  for crystalline). 

Ge2Sb2Te5 (GST), one of the most heavily investigated PCMs, is especially suitable for amplitude 

modulation in the near-infrared (NIR) or mid-wave infrared (MWIR) wavelengths thanks to the drastic 

material loss contrast and negligible loss in the amorphous state. On the other hand, emerging transparent 

PCMs in the NIR, such as Sb2Se3 and Sb2S3, are excellent candidates for phase-only modulation, because 

of the drastic index contrast and near zero absorption loss in both phases.  



 

 

 

Figure 1.1. A PCM optical property map.  

Plots for a wavelength of (a) 1310nm and (b) 1550nm. PCMs with a small extinction coefficient or 

low loss are circled in red. Different optical PCMs, GST52, GSST53, Sb2S3
54, Sb2Se3

55,GeTe45 and 

GSSe56 are represented with distinct symbols. The amorphous and crystalline phases are shown in 

blue and orange, respectively. The dashed trendlines suggest that the PCMs generally exhibit larger 

absorption losses as the refractive index contrast increases. 

 

To understand how material properties can determine the appropriate application, we define three 

optical figures of merit (FOM): (1) the refractive index change over the loss in the amorphous state 

(𝐹𝑂𝑀1 =
Δ𝑛

𝜅𝑎
), (2) the refractive index change over the loss in the crystalline state (𝐹𝑂𝑀2 =

Δ𝑛

𝜅𝑐
) and (3) the 

loss ratio between the two phases (𝐹𝑂𝑀3 =
𝜅𝑐

𝜅𝑎
). Phase-only modulation ideally requires zero loss. Since 

crystalline state loss is generally higher than amorphous, FOM2 is suitable for assessing the worst-case 

insertion loss and phase tunability of the device. For amplitude modulation applications, FOM3 can be used 

to evaluate the tradeoff between extinction ratio and insertion loss. For some devices, e.g., non-volatile 

beam splitters57–59, it is possible to mitigate the material loss in one of the phases by engineering device 

geometry. In such cases, FOM1 should be considered. Table 1 lists three FOM for various PCMs in the 

visible and telecommunication wavelengths, with red bold highlighting the highest. 

 



 

 

Table 1: FOM for different PCMs in 640 nm and 1550 nm 

 640 nm 1550nm 

 GST GSST Sb2S3 Sb2Se3 GeTe GSSe GST GSST Sb2S3 Sb2Se3 GeTe GSSe 

FOM1 0.10 0.77 ∞ 1.95 1.26 0.03 109.72 116.73 ∞ ∞ 73.12 78.92 

FOM2 0.04 0.24 3.05 0.77 0.47 0.01 2.52 4.17 190.33 ∞ 6.90 3.95 

FOM3 2.47 3.18 ∞ 2.53 2.67 2.32 43.56 28.00 ∞ ∞ 10.60 20 

 

PCMs have already been widely used in electronics, specifically for memory applications60, and was 

first proposed as optical memories around ten years ago61. Since then, tremendous efforts have been made 

in integrating PCMs into nanophotonics62,63, summarized in Figure 1.2. Using fast pulsed lasers (typically 

fs-lasers64), the PCMs can be switched between their crystalline, amorphous or mixed states. Different PCM 

states then represent distinct bits in optical intensity. Recent works have demonstrated reliable multi-bit 

operations, large capacity, and reliability62,65,66. Exploiting the reliable multi-bit operation, researchers also 

used PCMs to realize neuromorphic optical computing67–72, potentially overcoming the von Neumann 

bottleneck in modern computer architectures. Other than applications in integrated photonics, recent years 

have also seen significant advancements in using PCMs for active meta-optics64,73–81. Meta-optics can 

implement arbitrary phase profiles through engineering sub-wavelength scatterers (or meta-atoms). By 

incorporating PCMs in these meta-atoms, the phase profile can change significantly once the micro-

structural phase is switched. Such active tuning capability can enable many demanding applications, e.g., 

varifocal lens78,82–85, beam steering86,87, spatial light modulator22,88,89, and tunable holography90,91.  



 

 

 

Figure 1.2: PCM-based reconfigurable integrated photonics (top)59,62,92–95 and meta-optics 

(bottom)64,73–76 with optical (left) or electrical (right) controls.  

The device functionality depends on the material phase of the PCM. In optically controlled devices, 

laser pulses are absorbed to heat the PCM and trigger a phase transition. Complex optical alignment 

is generally required, limiting this approach to only small-scale applications. A more scalable 

approach is via electrical control. Short electrical pulses are sent to on-chip heaters to generate heat 

and subsequently actuate the phase transition. (Reprint from Ref [59, 62, 92-95, 64, 73-76] with 

permission) 

 

While remarkable progress has been made in optically induced switching of PCMs, the electrical tuning 

for photonic devices is still in its early stage (the first work was reported only five years back96), despite 

the promised scalability of electrical control. Early works reported limited cyclability (around ten cycles)96 

and low optical contrast (< 1dB)94. Recently, high endurance (over half a million cycles97) and large contrast 

switching (> 10 dB contrast) have been achieved using external heaters55,95,97,98 with high-quality 

encapsulation.  

 



 

 

Yet, broadband electrically controlled 2 × 2 programmable unit based on GST was still missing at the 

beginning of my study. While such functionality is ultimately indispensable for large-scale reprogrammable 

photonic applications, the inherent high absorptive loss of crystalline GST (c-GST) 99 becomes a severe 

issue for most designs. Existing electrically tunable GST devices are either designed only for 1 × 1 

switches100,101 or rely on ring resonators101,102 to circumvent the loss issue. The former is clearly not suitable 

for encoding information in different optical paths, or spatial information encoding. Thus, it has limited 

applications in programmable photonics for applications such as optical switch fabric and information 

processing35,36,39,103. While the latter is suitable for spatial information encoding in a narrow band, a 

broadband design is preferred to access a wider optical wavelength range.  

One direction to achieve the broadband operation is further engineering and optimization of the device 

geometry. In particular, a three-waveguide directional coupler switch structure was proposed in previous 

works45,57, which can mitigate the loss in the crystalline GST. However, there was no reversible switching 

demonstration of such devices, and the demonstrations were limited to either simulations or rapid thermal 

annealing for crystallization.  

The other approach, which provides a more generic solution to the loss problem, is to explore new 

wide-bandgap phase change materials with lower loss. In that direction, emerging wide-bandgap PCMs, 

such as GeSbSeTe (GSST)53, antimony selenide (Sb2Se3)104, and antimony sulfide (Sb2S3)105 have recently 

generated strong interest in the community78,106–109. In particular, Sb2S3 shows the widest bandgap among 

all these PCMs, allowing transparency down to ~ 600 nm in the amorphous phase105. Moreover, the lack of 

selenium in Sb2S3 makes it less toxic110 and less prone to causing chamber contamination during sputtering 

or evaporation processes. Thus, Sb2S3 is much more amenable to be adapted in a commercial foundry. 

Despite these promises, high-endurance electrical control of Sb2S3 remained unsolved. In Chapter 2 and 

Chapter 3, we show electrical control of both prototypical PCM GST and emerging low-loss PCM Sb2S3 

with high endurance (~ 1,000 cycles), low loss (< 1.0 dB), high extinction ratio (> 10 dB) and broadband 

operation (> 35 nm). The unique material property of Sb2S3 further enables a multi-bit operation (up to 5 

bits or 32 levels), which is of paramount importance to optical computing applications.  



 

 

Another challenge in PCM-based programmable PICs is the CMOS compatibility of Chalcogenide 

PCMs and the system’s scalability. Currently, the electrically controlled PCM-based PICs are still either in 

a single-device level59,106,107,111–114, or a very small system scale115. This can primarily be attributed to the 

low yield of the in-house fabrication due to the more sophisticated fabrication steps compared to optical 

actuation schemes, such as multi-stage overlay lithography, doping, annealing, and growth of metal vias. 

This incurs a significantly prolonged designing and testing cycle. Moreover, inconsistencies in the in-house 

fabrication process causes difficulty in building a very large-scale PCM-based PIC system. While 

commercial foundries provide much more reliable silicon photonic components, PCMs are not directly 

available there yet. In Chapter 4, we present the effort on scaling up the PCM-based PICs, where we 

developed a simple fabrication process to integrate the PCMs onto Intel large-volume silicon photonic 

platforms.   

On the metasurface side, free-space modulation of light is widely exploited in optical communications, 

holography, light detection and ranging, and virtual or augmented reality. Traditional spatial light 

modulators (SLMs) based on liquid crystals (LCs) and micro-electro-mechanical systems (MEMS) are 

slow, bulky, require large driving voltage, suffer from phase jitters, and only operate in a volatile manner. 

To address these limitations, recently there has been a lot of effort to realize free-space light control based 

on metasurfaces integrated with active materials116–122.The metasurfaces support resonances that allow 

reduction of the optical path length to attain certain phase or amplitude modulation.  For example, high-Q 

metasurfaces based on EO polymers have achieved GHz modulation speed119; LCs combined with 

Huygen’s metasurface can significantly reduce the pixel size and LC thickness required to attain a certain 

phase shift116; metasurfaces based on plasmonic nano-resonators coupled to epsilon-near-zero materials has 

achieved full 2π modulation with independent control of amplitude and phase117; large phase-modulation 

of 237° can be achieved by tuning a plasmonic metasurface using graphene123. Nevertheless, these 

approaches are all based on volatile changes such as Pockels or carrier depletion/accumulation effects, 

necessitating a constant power supply at static state. LC-based phase-only SLMs also suffer from temporal 

phase fluctuations due to constant refreshing124,125.  In contrast, nonvolatile control of free-space light can 



 

 

significantly improve energy efficiency and avoid phase flickering, since it holds steady once reconfigured, 

leading to a true ‘set-and-forget’ modulation. Particularly, chalcogenide-based phase change materials 

PCMs can afford a feasible solution126, thanks to the nonvolatile microstructural phase transition49, large 

contrast in complex refractive index (typically Δn ≥ 1)50, and CMOS compatibility47. In fact, PCMs have 

recently attracted considerable attention to create tunable metasurfaces for applications such as 

lensing122,127, beam steering128,129, intensity switching120,121,130,131, and spectral filtering132. 

Despite the progress, nonvolatile phase-only modulation - a highly desirable feature for SLMs – 

remains elusive. This is because previous works on nonvolatile tunable metasurfaces have used lossy PCMs 

such as GST120,131 or GSST121 which concomitantly induces large absorption upon phase transition, 

prohibiting phase-only modulation. The use of metallic heaters further leads to ohmic losses and makes 

transmissive operation impossible. Although nonvolatile phase-only control has been demonstrated in mid 

infrared where GST becomes transparent133, laser is used to switch the PCMs which is unscalable47 and 

impractical for SLMs that are normally electrically controlled. In Chapter 5, beyond on-chip light 

manipulation, we further demonstrate control of free space light using a high-Q BIC resonance and low-

loss PCM Sb2Se3. Importantly, this was the first attempt to control individual PCM meta-molecules to 

realize spatial light modulation.  

Finally, we end our dissertation with discussions about the challenges and perspectives about PCM-

based programmable PICs in Chapter 6.  

  



 

 

Chapter 2. Broadband electrically controlled silicon photonic switches based 

on GST59,134 

Programmable PICs have recently gained significant interest due to their potential in creating next-

generation technologies ranging from artificial neural networks and microwave photonics to quantum 

information processing. The fundamental building block of such programmable PICs is a 2 × 2 

programmable unit, traditionally controlled by the thermo-optic or free-carrier dispersion. Yet, these 

implementations are power-hungry, volatile, and have a large footprint (typically > 100 µm). Therefore, a 

truly 'set-and-forget' type 2 × 2 programmable unit with zero static power consumption is highly desirable 

for large-scale PICs. Here, we report a broadband non-volatile electrically controlled 2 × 2 programmable 

unit in silicon photonics based on the PCM GST. The directional-coupler-type programmable unit exhibits 

a compact coupling length (64 µm), small insertion loss (~2 dB), and minimal crosstalk (<-8 dB) across the 

entire telecommunication C-band while maintaining a record-high endurance of over 2,800 switching 

cycles without significant performance degradation. This non-volatile programmable unit constitutes a 

critical component for realizing future generic programmable silicon photonic systems. 

2.1 Optical refractive index characteristics of GST 

GST has a large loss in the crystalline state at 1550 nm in   

Figure 2.1, which is measured by ellipsometry (Woollam E-2000) and fitted with Tauc-Lorentz models. 

Note that the extinction coefficient 𝜅 is proportional to the optical loss 𝛼(𝑑𝐵/𝑚), given by  

 
α = 10 lg(e) ⋅

4πκ

λ
 (2.1) 

We simulate the fundamental TE mode of a hybrid GST silicon rib waveguide (HW) using Lumerical 

MODE. The results are shown in Figure 2.3(d), where a high absorption loss of 5.5 dB/µm when the GST 

is in the crystalline state is estimated. Directly using GST in photonic switches will introduce huge insertion 

loss. For example, if GST is used in a Mach-Zehnder interferometer, then to achieve a π-phase shift, a 



 

 

length of 𝐿 =
𝜆0

2Δ𝑛
= 4.2𝜇𝑚 would be required. This will then give a very high loss of 23.1 dB compared 

to <1 dB loss using the three waveguide structures in the following sections. 

  

Figure 2.1: The complex refractive index of a/c-GST measured by ellipsometry.  

 

2.2 Design of programmable 1 × 2 photonic programmable units using GST 

We design the 1 × 2 programmable unit so that the light can only be coupled from the input bare silicon 

waveguide (SW) to the hybrid (HW) when the GST is in its amorphous state, where only a slight loss is 

introduced, as shown in Figure 2.2.  

 



 

 

Figure 2.2: Broadband nonvolatile electrically controlled 1 × 2 programmable unit schematic and images.  

(a) device schematic (b) The microscope and (c) SEM images of the device. In the SEM image, the GST thin 

film, p++, and n++ regions are indicated by orange, blue, and green fake colors. 

 

The design procedure of the programmable units is similar to Ref 45 as follows. The HW width was 

optimized so that when GST is in an amorphous state, the HW and SW can couple to each other most 

strongly, which happens when the two share the same effective index. As shown in Figure 2.3(a), a 408 nm 

a-GST HW has the same effective index with a 450 nm SW. Due to the significant refractive index change 

when the GST is switched into a crystalline state, a large phase mismatch would be induced. Thus, the 

coupling would be weakened.  

Then the supermodes of the HW and SW with a gap g are simulated and optimized. The coupling length 

Lc and the loss when GST is in the amorphous/crystalline state are determined when sweeping g, where the 

coupling length is determined with coupled mode theory135, using the following formula: 

 
𝐿𝑐
𝑎 =

𝜆0
2(𝑛𝑒𝑣𝑒𝑛 − 𝑛𝑜𝑑𝑑)

 (2.2) 

Where 𝜆0 is the free-space wavelength, 𝑛𝑒𝑣𝑒𝑛 and 𝑛𝑜𝑑𝑑 are effective index for even and odd mode and 

𝐿𝑐
𝑎 is the coupling length with a-GST. We note that the loss is determined by decomposing the input field 

into the supermodes and calculating the loss separately, as below: 

 

𝑇𝑎,𝑐 =∑ 𝑇𝑗
𝑎,𝑐

𝑗
=∑ (𝐴𝑗

𝑎,𝑐 exp (
𝑖2𝜋𝜅𝑗

𝑎,𝑐

𝜆
𝐿𝑐
𝑎))

2

𝑗
 (2.3) 

 𝑙𝑜𝑠𝑠𝑎,𝑐 = 10 log𝑇𝑎,𝑐 (2.4) 

 Where 𝑇𝑎,𝑐  is the transmission and the superscript indicates the same formula applies to both 

amorphous and crystalline GST, 𝑇𝑗
𝑎,𝑐

 is the transmission of the jth supermode, 𝐴𝑗
𝑎,𝑐

 is the expansion 

coefficient of the input field into the jth supermode field, 𝜅𝑗
𝑎,𝑐

 is the imaginary part of the effective index of 

the jth supermode. The crosstalk of the device can be calculated with similar formula as well. This formula 

was implemented in the Lumerical MODE solver. The result is shown in Figure 2.3(c). As g increases, the 



 

 

coupling length increases, the loss in c-state decreases because of a weaker mode interaction. However, 

since the amorphous state GST still has a finite amount of loss, a longer coupling length leads to a larger 

insertion loss. We selected g of 450nm to get an insertion loss smaller than 0.5 dB and a compact device 

footprint of 𝐿𝑐 = 42 µ𝑚. The supermode simulation results are shown in Figure 2.3(e). With a-GST, the 

supermode extends in both waveguides, indicating a strong coupling. While with c-GST, the supermodes 

are highly localized in each of the waveguides, indicating weak coupling. Thus, the light bypasses the high 

absorption loss of c-GST. 

 



 

 

Figure 2.3: Mode simulation results for the 1 × 2 programmable unit.  

(a) Effective indices of a bare silicon waveguide, a-GST HW, and c-GST HW changing with the waveguide 

width. (b) Effective loss of a/c-GST HW, the inset shows the zoomed-in loss of a-GST HW. (c) Mode 

simulation result of 1 × 2 programmable unit coupling length and insertion loss with a/c-GST changing with 

gap. (d) Mode simulation results for (i)a-GST and (ii)c-GST HW. (e) Mode simulation results for 1 × 2 

programmable unit supermodes. (i)(ii) First and second order mode when GST is in amorphous state. (iii)(iv) 

First and second order mode when GST is in crystalline state.  

 

2.3 Design of the 2 × 2 programmable photonic units based on GST 

The first three TE supermodes are considered for the 2 × 2 programmable unit design. To achieve the 

best coupling between the three waveguides, a phase-matching condition needs to be satisfied: 

 2𝑛2 = 𝑛1 + 𝑛3 (2.5) 

 By fixing the width of SWs to 450 nm and sweeping the width of the middle waveguide, an optimal 

HW width of 409 nm is determined, as shown in Figure 2.4(a). The minor difference from the previously 

determined HW width for 1 × 2 programmable unit can be attributed to numerical error. 

The coupling length and the insertion loss are calculated with different gap g, as shown in Figure 2.4(b). 

An optimal gap of 450 nm is selected for a short coupling length, as well as a small insertion loss of 

0.33/0.65dB for a/c-GST, respectively. The coupling length of a three waveguide coupler is calculated by 

the following formula when the phase-matching condition is satisfied: 

 
𝐿𝑐
𝑎 =

𝜆

2(𝑛1 − 𝑛2)
 (2.6) 

Finally, we show the optical modes in Figure 2.4(c). One can see a strong coupling between the HW 

and SW when GST is in amorphous state, as in Figure 2.4(c-i, c-iii). However, when the GST is switched 

into a crystalline state, the modes are localized in SW or HW, as in Figure 2.4(c-iv, c-v,  c-vi), indicating a 

very weak coupling. Although the effective index of mode 1 (cry) has a large imaginary part, since light 

can’t get coupled into mode 1 if excited from SW, this high loss mode is bypassed. 

 



 

 

 

Figure 2.4: Mode simulation results for the 2 × 2 programmable units.  

(a) supermode effective indices changing with HW width. (b) coupling length and loss with a/c-GST changing 

with the gap. (c) Mode simulation results. (i)(ii)(iii) First, second, and third-order modes when GST is in the 

amorphous state, (iv)(v)(vi) First, second, and third-order modes when GST is in the crystalline state. 

 

The designs of the 1 × 2 and 2 × 2 programmable units are verified with Lumerical FDTD as in Figure 

2.5. A larger insertion loss of around 1.0 dB compared to the 0.5 dB loss in the MODE solver can be 

attributed to extra loss introduced by the S-bends and can be further reduced by a better S-bend design. 

 

 



 

 

 

Figure 2.5: FDTD simulation results for both 1 × 2 and 2 × 2 programmable units.  

(a, b) Transmission spectrum and field distribution of a-GST 1 × 2 programmable unit, (c, d) Transmission 

spectrum and field distribution of c-GST 1 × 2 programmable unit, (e, f) a-GST 2 × 2 programmable unit, (g, 

h) c-GST 2 × 2 programmable unit. 

 

 



 

 

2.4 Fabrication variation tolerance analysis 

In this section, we simulate how the insertion loss and extinction ratio in both amorphous and crystalline 

state changes with respect to various device parameters, such as the waveguide width, the GST width, the 

GST misalignment, the etch depth and the gap. The results are summarized in Figure 2.6 and Figure 2.7. 

The fabrication variation in waveguide width, etching depth and GST thickness can lead to significant 

higher insertion loss and lower extinction ratio. 

 

Figure 2.6: Insertion loss vs. fabrication variation study on the 2 × 2 programmable unit.  

(a) Waveguide width deviation (b) etch depth deviation (c) GST height deviation (d) gap deviation (e) GST 

width deviation (f) GST position deviation from the nominal values.  



 

 

 

Figure 2.7: Extinction ratio vs. fabrication variation study on the 2 × 2 programmable unit.  

(a) Waveguide width deviation; (b) etch depth deviation; (c) GST height deviation; (d) gap deviation; (e) GST 

width deviation; (f) GST position deviation from the nominal values. 

 

2.5 Heat transfer simulation 

The time-dependent heat transfer performance was studied with a cross-coupled electro-thermal model 

in COMSOL Multiphysics, which includes the semiconductor model for PIN electrical simulation as well 

as heat transfer in solid model for heating simulation. The detailed description of the model can be found 

in101,136. For amorphization, a short sharp pulse with high peak voltage is favorable for the melting and 

quenching process, where the temperature of GST increases above its melting point (Tm ~ 616°C) and 

decreases rapidly below the glass transition temperature (Tg ~ 155°C) to form the amorphous phase. The 

amorphous GST can be crystallized through nucleation and growth process, by heating it above the Tg but 

below Tm, and gradually cooling it down. The amorphization condition was identified as pulse duration of 

100 ns, and peak amplitude of 6.8(8.4) V for the 1 × 2 (2 × 2) programmable units. To crystallize the GST, 

a pulse with a duration of 50 µs and peak amplitude of 2.4 (2.8) V. The larger peak amplitude required for 



 

 

the 2 × 2 programmable unit is due to the wider intrinsic region, thus a larger resistance and heating volume. 

Figure 2.8(a) and (d) show the time dependent temperature result when the amorphization or crystallization 

pulses are applied. The experimental pulses have higher voltages because of nonideal contact and load 

impedance mismatch. 

 

Figure 2.8: Transient heat transfer simulation results.  



 

 

(a, d) The time-dependent temperature of the GST. The blue and orange lines are amorphization and 

crystallization pulses, respectively. The insets show zoom-in figures of the amorphization pulse. (b, e) 

Temperature distribution at 50 µs for crystallization. (c, f) Temperature distribution at 100 ns for amorphization. 

(a-c) 1 × 2 unit; (d-f) 2 × 2 unit. 

 

2.6 Device fabrication  

Here we describe the fabrication of our electrically controlled programmable units using GST. The 2 × 

2 programmable unit leverages an asymmetrical three-waveguide directional coupler structure45,137,138 

fabricated in a Silicon on Insulator (SOI) platform combined with an on-chip P-I-N (p++-doped-intrinsic-

n++-doped) heater, as shown in Figure 2.9(a). Compared to previous works, where rapid thermal 

annealing45,139 was used to change the PCM phase, our devices can be reliably and reversibly switched with 

short electrical pulses via on-chip PIN micro-heaters for more than 2,800 cycles. The design consists of 

two Si transmission waveguides with a coupling region separated by a truncated transition waveguide. To 

circumvent the high absorption loss of c-GST, only the transition waveguide is loaded with a 20 nm-thick 

GST, encapsulated by 40 nm Al2O3 to prevent GST oxidation. By design, the phase-matching condition is 

satisfied only when the GST is switched to the amorphous state (a-GST), which allows light to couple into 

the transition waveguide and further to the cross port. However, if the GST is in the crystalline state, light 

barely couples into the transition waveguide due to a significant index mismatch, thus bypassing the high 

loss of c-GST. The gap between the waveguides is designed to achieve a good tradeoff between insertion 

loss and device footprint. Our simulations show an insertion loss of less than 1 dB, an extinction ratio of 

larger than 13 dB at 1520 nm and a 1 dB bandwidth larger than 50 nm. While similar designs have been 

reported before45,137, it is noteworthy that electrical tunability was not demonstrated in such programmable 

unit designs, primarily due to the lack of suitable heater designs for large-volume GST switching. A PIN 

heating element implements the electrical control over the phase switching, which is realized by selectively 

doping the adjacent regions to the waveguide. We note that this concept directly extends our prior works, 

where we implemented this concept101 for a 1 × 1 switch. However, in contrast to the previous 1 × 1 



 

 

programmable unit work, this work demonstrates a 2 × 2 programmable unit which can be used in a lot 

more applications thanks to its suitability for spatial information encoding. Moreover, the volume of the 

GST being electrically switched experimentally is ~0.4 µm3
 in the present work, around ten times 

larger98,100,101,140. Therefore, prior to fabrication, we optimized the design and operation of the PIN 

microheater and device transmission using finite element simulations101,136, considering the device's 

electrical and heat transfer transient performance. We also stress that, compared with resistor-type heaters 

such as metal, P-I-P, or N-I-N heaters, PIN heaters are more energy efficient. This can be intuitively 

understood by recognizing the electric current in the PIN diode increases exponentially with voltage, while 

that of resistor-type heaters only grows linearly.  

Our electrically controlled 2 × 2 programmable unit were fabricated on a commercial SOI Wafer with 

220nm-thick Silicon on 3µm-thick SiO2 (SOITECH). The rib waveguides and grating couplers were 

defined with electron-beam lithography (EBL, JEOL JBX-6300FS) using a positive tone E-beam resist 

(200-nm-thick ZEP-520A) and partially etched by 120 nm in a fluorine-based inductively coupled plasma 

etcher (ICP, Oxford PlasmaLab 100 ICP-18) using a mixed gas of SF6 and C4F8. The etching rate was 

calibrated each time before the etching to ensure a correct etch depth. The doping regions were defined by 

two additional EBL steps with 600-nm-thick poly (methyl methacrylate) (PMMA) resist and implanted by 

boron (phosphorus) ions for p++(n++) with a dosage of 2 × 1015 ions per cm2 and ion energy of 14 keV (40 

keV). A tilt angle of 7° while conducting ion implantation was used to misalign with the silicon lattice and 

thus achieve uniform deep doping. To activate the dopants, the chips were annealed at 950 °C for 10 min 

(Expertech CRT200 Anneal Furnace). Before the metal contact deposition, to ensure ideal Ohmic contact, 

the surface native oxide was removed by immersing the chips in 10:1 buffered oxide etchant (BOE) for 10 

seconds. The metal contacts were then immediately patterned by a fourth EBL step using PMMA and 

formed by electron-beam evaporation (CHA SEC-600) and lift-off of Ti/Pd (5 nm/180 nm) layers. After a 

fifth EBL defining the GST window, a 20-nm GST thin film was deposited using a GST target (AJA 

International) in a magnetron sputtering system (Lesker Lab 18), followed by a lift-off process. The 

deposition rate was calibrated on a silicon chip using Ellipsometer (Woollam alpha-SE). The GST is then 



 

 

encapsulated by 40-nm-thick Al2O3 through ALD (Oxford Plasmalab 80PLUS OpAL ALD) at 150 °C. To 

ensure good contact between the electric probe and metal pads while measuring, the Al2O3 on the metal 

contacts were removed by defining a window using a sixth EBL with 600nm PMMA, then etching in a 

chlorine-based inductively coupled plasma etcher (ICP-RIE, Oxford PlasmaLab 100 ICP-18). Finally, rapid 

thermal annealing (RTA) at 200 °C for 10 min was conducted to completely crystallize the GST. 

The fabricated device's optical and scanning electron microscope (SEM) images are shown in Figure 

2.9(b) and (c), respectively. In detail, the PCM phase is switched by an on-chip P-I-N heater, defined by 

doping the 100-nm-thick silicon slab via ion implantation while leaving the three waveguides intrinsic101 to 

reduce the insertion loss. The doped regions and metal pads are designed to maintain a distance of 200 nm 

and 1 µm, respectively, from the directional coupler to reduce excess loss from the free carriers.  

We also fabricated 10-µm-long GST on silicon waveguide switches, as shown in Figure 2.11(a) and 

(b). 

 

Figure 2.9: Broadband non-volatile electrically controlled 2 × 2 programmable unit in silicon photonics.  

(a) Device schematics. Inset: the cross-sectional view at the center of the device. The 40 nm encapsulation 

Al2O3 is not shown. (b) The optical micrograph and (c) SEM images of the device. In the SEM image, the GST 

thin film, p++, and n++ regions are indicated by false colors.  



 

 

2.7 Device characterization 

2.7.1 Characterization setups 

The fabricated devices were characterized in a customized vertical fiber coupling setup. The 

programmable units were measured with a vertical fiber-coupling setup using a coupling angle of 25°, 

shown in Figure 2.10(a). The stage temperature was fixed at 26 °C controlled by a thermoelectric controller 

(TEC, TE Technology TC-720). A tunable continuous-wave laser (Santec TSL-510) provided the input 

light. The polarization was controlled by a manual fiber polarization controller (Thorlabs FPC526) to match 

the TE mode of the rib waveguides. A low-noise power meter (Keysight 81634B) was used to measure the 

static optical transmission from the output grating couplers. The transmission spectra of the devices were 

obtained by normalizing to the closest reference waveguide spectra. For the I–V characterization and on-

chip electrical switching, electrical pulses were applied to the on-chip metal contacts via a pair of electrical 

probes controlled by two probe positioners (Cascade Microtech DPP105-M-AI-S). In particular, the I-V 

curve measurement was performed using a source meter (Keithley 2450), which is later used to estimate 

the power of the applied pulses. The crystallization and amorphization pulses were generated from a pulse 

function arbitrary generator (Keysight 81160A). The tunable laser, power meter, thermal controller, source 

meter, and pulse function arbitrary generator are controlled by a laptop with a National Instrument interface.  

The transient response of the programmable units was measured under the same vertical fiber coupling 

setup but with a faster light detection instrument as shown in Figure 2.10(b). Instead of the power meter, 

we used a switchable gain balanced amp photodetector (Thorlabs, PDB450C) with 150MHz bandwidth to 

obtain the fast time response. An optical amplifier (Amonics AEDFA-30-B-FA) was used to amplify the 

optical signal to ~1 mW before the photodetector so that the photodetector is operating in its linear range. 

The photodetector response was captured by an oscilloscope, triggered by the applied electrical signal.  



 

 

  

Figure 2.10: Measurement setup schematics.  

Setup for (a) optical transmission measurement setup; (b) transient response measurement.  

2.7.2 Results for 1 × 1 waveguide switch 

By switching the GST between amorphous/crystalline (low-loss/high-loss) states, the optical switch is 

controlled on/off, with an extinction ratio larger than 15 dB. We show that this device can be switched to 

ten distinct transmission levels. This was accomplished by first complete crystallizing (2.2 V peak voltage, 

30 µs duration and 10 µs falling edge) the GST and then applying gradually increased voltage (peak volage 

starts from 6.5V, with an increment step of 0.1V, and 10 total steps, 200 ns pulse duration and 8 ns falling 

edge) for partial amorphization, as in Figure 2.11(c). We repeated the multilevel switching test three times 

at a wavelength of 1550 nm, and each time a multilevel performance can be observed, shown in Figure 

2.11(e). We further demonstrate more than 2000 switching events via sending in amorphization (2.2V) and 

crystallization (7.4V) pulses alternately, and the results are shown in Figure 2.11(d). The waveguide switch 

exhibits a large extinction ratio between the ON and OFF state after more than 2000 switching events, 

showing excellent endurance. We note that the transmission level becomes stable after around 600 

switching events due to the larger crystalline domain forming from the initial ‘conditioning’ steps62. 



 

 

 

Figure 2.11: Measurement results for the waveguide switch.  

(a) Microscope and (b) SEM image of the waveguide, where the GST thin film is indicated by the orange fake 

color. (c) Multilevel switching results. (d) Cyclability test results. (e) Three rounds of multilevel switching 

results. 

 

 



 

 

2.7.3 Results for 1 × 2 silicon photonic programmable units 

We first show the characterization results for the 1 × 2 programmable unit. It has an extinction ratio of 

9 dB, a larger insertion loss of 4 dB than simulation, nonvolatility, and large endurance of 8,000 switching 

events, as in Figure 2.12 and Figure 2.13. The high loss in this design can also be attributed to fabrication 

imperfection. We also note that 1 × 2 design inherently has a higher overall loss than 2 × 2 design, as shown 

in Figure 2.3(c) and Figure 2.4(b).  

 

Figure 2.12: Optical transmission spectra measurement results for the 1 × 2 programmable unit.  

(a) The transmission spectrum (i) initially when GST is in the crystalline state, (ii) after sending an 

amorphization pulse, (iii) after sending a crystallization pulse. (b) Transmission spectrum of 3 cycles 

amorphization and crystallization at the bar port. The spectra overlap well, showing a reliable and reversible 

switching performance. (c, d) The continuous-time measurement result for (c) cross and (d) bar ports. 



 

 

 

Figure 2.13: The endurance test results for the 1 × 2 programmable unit.  

(a) cross and (b) bar port. 8000 switching events are shown without significant performance degradation at 

1550nm. The degradation in extinction ratio and insertion loss compared to Figure 2.12 is mainly due to 

material ablation when testing the switching pulse conditions.  

 

2.7.4 Results for 2 × 2 silicon photonic programmable units 

The reversible operation of the electrically controlled 2 × 2 programmable unit is demonstrated by 

measuring the transmission of a tunable laser through the bar and cross port at two different phases of GST, 

summarized in Figure 2.14(a). Initially, the GST is prepared in the crystalline state (c-GST), for which the 

light passes through the bar port with an extinction ratio over 10 dB at 1550 nm and over 8 dB across the 

entire C-band in Figure 2.14(a-i). After applying a short, high amplitude pulse (13.6 V, 8 ns leading/falling 

edge, 200 ns pulse width with 380 nJ pulse energy) to induce amorphization (a-GST), the transmission 

spectrum flips, resulting in a high transmission through the cross port with over 10 dB extinction ratio at 

1550 nm and over 8 dB across the entire C-band in Figure 2.14(a-ii). Subsequently, we applied a longer but 

lower amplitude pulse (3.2 V, 8 ns leading edge, 50 µs pulse width with 6.83 µJ pulse energy) to trigger 

crystallization, returning the switch to the original high bar transmission state in Figure 2.14(a-iii). Crosstalk 

less than -8 dB is observed across the entire telecommunication C-band (1530 to 1565 nm) in cross and bar 



 

 

states. We estimate the switching energy density to be 0.95/1.63 fJ/nm3 for amorphization/crystallization. 

This is comparable with 101, where an energy density of 0.2/1.95 fJ/nm3 for the phase transition was 

demonstrated. The cycle-to-cycle reproducibility was verified over three consecutive switching cycles, as 

shown in Figure 2.14(b), where transmission spectra for a-GST and c-GST at bar port overlap nearly 

perfectly. We then established the long-term retention of the device states by performing several 

measurements over a period of one month, as shown in Figure 2.14(c). A good match between the 

transmission spectra for devices in the a-GST configuration after putting the device in an ambient 

environment for 132 hours and 720 hours indicates that the switching is indeed non-volatile. Slight 

variations among the spectra are likely due to minor differences in the optical alignment during the different 

measurements. 

 

Figure 2.14: Characterizations of the electrically controlled 2 × 2 programmable unit.  

(a) Transmission spectrum of the device (i) initially, (ii) after applying an amorphization pulse (iii) after 

applying a crystallization pulse. The shaded areas are measured raw spectra, and the solid lines are spectra after 

a low pass filter. IL and ER are insertion loss and extinction ratio, respectively. (b) Transmission spectra 

measured over three cycles at bar port. Each cycle is indicated with different color saturations. (c) Device 



 

 

functionality over an extended period of one month with a-GST. The solid and dashed lines correspond to the 

bar and cross port transmission.  

 

Finally, we demonstrate reversible operation over 5,750 switching events measured at 1550 nm without 

significant performance degradation, presented in Figure 2.15, thanks to the optimized PIN heater design. 

We note that due to limitations in our measurement setup, the cross and bar ports are measured 

consecutively, as shown in Figure 2.15(top) and (bottom), respectively. The two distinct transmission levels 

arising from switching between the crystalline and amorphous state are indicated by the shaded regions, 

whereas transmission contrasts of 8 dB and 11 dB are obtained for the cross and bar ports, respectively. 

The measured insertion loss at beginning is around 1 dB higher compared to Figure 2.14, which can be 

attributed to fiber alignment variation as well as device performance degradation due to previous 

measurement steps. Interestingly, although the cross-port transmission contrast reduces from 8 dB to 6 dB 

at the 2,750th event, the bar port contrast remains around 11 dB throughout the measurement. We attribute 

this reduction in the cross-port transmission contrast to thermal reflowing and material degradation of GST 

after multiple cycles. The performance degradation is less pronounced for the bar port because even if the 

GST shape changes due to thermal reflowing, a significant index mismatch still exists between the transition 

waveguide and input waveguide so that light still does not couple to the transition waveguide. We note that 

some data points are removed around the 600th, 800th, and 1,900th events because the same amorphization 

condition failed to trigger the phase change, and larger voltages of 13.8, 14.1, and 14.4V were used after 

that, respectively. This could result from drifting contacts of electrical probes and/or GST degradation. The 

contact drift issue could be solved by wire-bonding the chip to a carrier.  



 

 

 

Figure 2.15: Cyclability test of the electrically controlled 2 × 2 programmable unit.  

(Top) Transmission measured at the cross port. (Bottom) Transmission measured at the bar port. The 

blue/orange dots represent the measured transmission at 1550 nm after sending in a 

crystallization/amorphization pulse (with the same pulse condition in Figure 2.14). The shaded regions 

correspond to variations of the two distinct transmission levels. A larger insertion loss of 3 dB at both ports at 

the beginning can be attributed to fiber alignment variation as well as the material degradation and thermal 

reflowing of the GST thin film in previous measurement steps, see Discussion Section.  

 

To understand the failure mechanisms after the cyclability test, we show the SEM image of the devices 

after the cyclability test. In Figure 2.16(a) and (c). The GST thin film shrinks and exhibits a wavey edge. 

This is likely due to the reflowing67,102 while GST was melted or due to non-uniform heating (as can be 

seen in Figure 2.16(c)). Some bubbles can also be observed, as in Figure 2.16(b) and (d), which indicates 

some of the phase change materials are damaged.  



 

 

 

Figure 2.16: SEM of the devices after multiple cycles.  

(a, b) 1 × 2 programmable unit reflowing and ablation. (c, d) 2 × 2 programmable unit reflow and ablation. All 

scale bars are 500 nm. 

 

2.7.5 Transient response of the 2 × 2  programmable unit 

The results are shown in Figure 2.17, and we analyze the transient performance. A high transmission 

is expected at the bar/cross port when GST is in the crystalline/amorphous state. Indeed, the measurement 

results in Figure 2.17(a) and (c) show that around 1 µs or 2 µs after sending in the 200 ns amorphization 

pulse, the transmission at bar port decreases while at cross port increases to a steady level. In this case, the 

longer transient response time than electrical pulse duration can be attributed to the finite response time of 

the PIN junction in the forward bias mode due to the slow process of minority carrier injection. We also 

found a sudden transmission drop right after the electrical pulse due to the carrier effect, which significantly 

increased the loss of the waveguides. From Figure 2.17(b) and (d), we can conclude that in the process of 

crystallization, the response time of the programmable unit is almost the same as the total pulse duration 

due to the small voltage and the relatively long electrical pulse applied (50 μs steady time and 30 μs falling 



 

 

edge, total pulse duration 80 μs). The bar-port changes into a high transmission level, while the cross-port 

goes to a low transmission level. 

Interestingly, in Figure 2.17(b), we observed a two-step transmission rise. We can attribute such a 

phenomenon to carrier injection into the middle transition waveguide, which introduces an effective index 

imaginary part mismatch. The light then stays in the input waveguide. Thus, a higher bar-port transmission 

can be observed141.  

 
Figure 2.17: Transient response measurement results for electrically controlled 2 × 2 programmable unit.  

(a) amorphization and (b) crystallization at the bar port, and (c) amorphization and (d) crystallization at the 

cross port. The blue curve shows optical transmission while the orange line provides the electrical pulse for 

pulse-change actuation. The amorphization response time is around 1 or 2.5 μs while the crystallization time 

is around 80μs. 

 

2.7.6 Thermal stability test results 

One advantage of a broadband device is its thermal stability. Here we measured the transmission 

spectrum of the 2 × 2 programmable unit after the cyclability test at 25, 30, and 35 °C when the GST is in 

the crystalline state as shown in Figure 2.18. The excellent match among the curves offers good thermal 

stability. 



 

 

 

Figure 2.18: Thermal stability test measurement.  

The temperature varies from 25°C to 35°C. No significant spectrum shift is observed. The cross and bar port 

transmission are represented by solid and dashed lines, respectively.  

 

2.8 Deterministic multilevel operation with multiple GST segments 

2.8.1 Multilevel 1 × 2 programmable units 

Figure 2.19 shows the simulation result for a 1 × 2 optical programmable unit, for which the optical 

design is similar to Ref. 59. One change is that two individually controlled GST segments, instead of one, 

are used to achieve multi-level operation. The achievable low-loss states are, however, not 2N as in the 1 × 

1 programmable unit case, but N + 1, since any a-GST segment before c-GST segment results in light 

coupling to the GST-loaded waveguide, which then gets completely absorbed by the highly lossy c-GST 

segments. Our simulation results agree with this intuition. We denote the phase sequence from left to right 

as {phase of segment 1, phase of segment 2} and show the electric field distribution and the transmission 

spectrum for all four possible phase sequences. Figure 2.19(b) and (c) shows that Sequence {a, c} exhibits 

a high insertion loss of ~ 3 dB due to high absorption of the c-GST segment 2. This is verified by the electric 

field distribution image, where the light first couples into the cross port along the a-GST segment 1, then 



 

 

gets completely absorbed by c-GST segment 2. Since the length of the first segment is half of the coupling 

length, exactly half of the light is coupled and absorbed, agreeing well with the simulation. 

 

Figure 2.19 FDTD simulation for a deterministic three-level 1 × 2 programmable unit using GST. 



 

 

(a) Schematic. (b) Intensity distribution and (c) the corresponding transmission spectra for different GST phase 

combinations. GST sequence {a, c} introduces an insertion loss of ~ 3 dB due to c-GST absorption loss, pointed 

by a white arrow in the field distribution map. 

 

2.8.2 Multilevel 2 × 2 programmable unit 

Previous GST-based low-loss 2 × 2 programmable unit leveraged a PM-TDC structure to circumvent 

the loss of c-GST.57–59 In these works, an extra transition WG is placed between two silicon waveguides 

with 20-nm-thick GST deposited on top for active switching. When the GST is in the a-phase, the transition 

WG is designed to phase match with other two WGs. Therefore, the input light can be coupled into the 

transition WG, and then to the cross port. On the contrary, once switching the GST to the c-phase, a large 

effective index difference is introduced, and the light can barely couple to the high loss transition WG. This 

bypasses the high loss of c-GST. Unfortunately, a straightforward extension of this PM-TDC method to 

multi-level operation does not work as we will show in the following analysis. 

Challenges of multiple segments switching using the phase-matched TDC approach 

Figure 2.20 shows a simulation using a similar design to Ref. 59 but with two GST segments. It shows 

that both {a, c} and {c, a} phase sequences have a high loss of ~ 3 dB. The former can be explained based 

on the same concept as in the 1 × 2 programmable unit case: no c-GST should be placed after an a-GST 

segment when finite light is still in the transition WG, defined as the Type I absorption loss. Besides, the 

transition WG, which is open at the end, adds another constraint for low-loss operation: no light can remain 

in the transition WG at the end of this transition waveguide. Otherwise, that portion of light will radiate and 

get lost, defined as Type II radiation loss. 



 

 

 

Figure 2.20: Phase-matched three-WG 2 × 2 directional couplers have high loss for intermediate levels.  

(a) Schematic. (b) the transmission spectra for different GST phase sequences and (c) the corresponding 

intensity distribution. Both sequences {a, c} and {c, a} introduce an insertion loss of ~ 3 dB. The former is due 

to c-GST absorption loss (Type I loss), while the latter is due to the radiation loss at the end of the transition 

WG (Type II loss), both indicated by white arrows. 

 

Analytical and numerical study of the three-waveguide directional coupler systems 

To tackle this loss issue, it is crucial to understand the intensity dynamics inside of each WG in this 

coupled-waveguide system. We describe the coupled three waveguide system using the following coupled 

partial differential equation142: 

 



 

 

 

{
 
 

 
 
𝑑𝑎0
𝑑𝑧

= 𝑖𝜅∗(𝑎1 + 𝑎2)

𝑑𝑎1
𝑑𝑧

= 𝑖𝜅𝑎0 − 𝑖𝛿𝑎1

𝑑𝑎2
𝑑𝑧

= 𝑖𝜅𝑎0 − 𝑖𝛿𝑎2

 (2.7) 

 where 𝑧 is the distance in light propagation direction, 𝑎𝑖,𝑖=0,1,2 is the amplitude in three WGs in Figure 

2.21(a), 𝜅 = Δ𝑛𝑒𝑓𝑓,𝑐 ∙
2𝜋

𝜆0
 and 𝛿 = Δ𝑛𝑒𝑓𝑓,𝑖 ∙

2𝜋

𝜆0
 are the coupling coefficient and detuning between the center 

and other WGs, Δ𝑛𝑒𝑓𝑓,𝑐 is the effective index difference between the super-modes formed by coupled WG0 

and WG1 (or WG2), Δ𝑛𝑒𝑓𝑓,𝑖 is the effective index difference between the modes in WG0 and WG1 (or 

WG2) assuming they are isolated, and 𝜆0 is the vacuum wavelength of light. We note that only the nearest 

coupling is considered in this model, which is usually a valid assumption in a 2D photonic WG array. We 

also assume that WG1 and WG2 have the same geometry to simplify the model and neglect the loss term 

for simplicity. To quickly capture physics, we show the numerical analysis below. 

 

Figure 2.21: Analytical model reveals that phase-matched three-WG couplers can never have low-loss 

intermediate levels. 

(a) Schematic for a generic GST three-waveguide coupler. The GST patches are denoted in white. In all the 

discussions, we assume the WG1 the input port. (b) Numerically calculated light intensity in three WGs. Blue, 

orange and green colors denote WG0, WG1, and WG2, respectively. In the case of two GST segments, phase 

sequence {c, a} would lead to a propagation distance of 𝐿𝑐/2, indicated by the gray dotted line. It is clear that 

finite energy still remains in WG0, causing Type II radiative loss. 

 



 

 

First, to analyze the case where WG0 is phase matched with WGs 1 and 2, we set 𝛿 to zero and solve 

this equation numerically. We found that the middle WG intensity is zero only when the light is completely 

in the cross or bar port in Figure 2.21(b).Therefore, Type I loss occurs in the {a, c} state, while Type II loss 

occurs in the {c, a} state. Therefore, we summarize the requirements for low-loss multilevel operation in a 

three-waveguide directional coupler as (1) no light can be in the middle waveguide at a c-GST segments to 

avoid Type-I loss, and (2) no light can be at the end of the middle waveguide to avoid Type-II loss, and (3) 

there must exist some intermediate splitting ratios (except from complete bar or cross) when the intensity 

of light in WG0 is zero. Obviously, the last requirement can never be fulfilled in a phase-matched TDC. 

New degree of freedom: phase-detuned three-waveguide directional couplers 

To solve this issue, a new degree of freedom must be introduced. We deliberately designed WG0 to 

have a phase detuning 𝛿 with WG1 and WG2, i.e., a phase-detuned TDC scheme. The phase detuning 

between WGs can be quantitatively represented by the detuning rate 𝛿, where 𝛽 = 2𝜋/𝜆 is the vacuum 

wavevector and Δ𝑛𝑒𝑓𝑓 is the effective index difference between WGs. To intuitively understand the wave 

dynamics in this system, we performed a numerical simulation as shown in Figure 2.22, where different 

plots show the intensity in the three WGs versus propagation length under different detuning rates 𝛿. It can 

be clearly seen that different splitting ratios between WG1 and WG2 can be obtained when zero light is in 

WG0. We note that Eqns. (2.7) are analytically solvable, and the solution is presented in Appendix I of 

ref134.  



 

 

 

Figure 2.22: Detuning the middle waveguide provides rich optical intensity dynamics, enabling low-loss 

intermediate levels.  

(a)-(i) Intensity dynamics in the three waveguides as the detuning 𝛿 varies from 0 to 4𝜅. As the detuning gets 

larger, the intensity in WG0 becomes smaller and that in the other two is smoother. As 𝛿 increases from zero, 

multiple zeros in blue curve emerge (such as pointed by red arrows in (b), overcoming Type II loss. Note that 

this simulation assumes a coupling length 𝐿𝑐 of 10 𝜇𝑚 between WG0 and WG1 (or WG2) when 𝛿 = 0. 

 

A tradeoff must be considered when choosing detuning rate 𝛿. A large detuning 𝛿 generally creates 

more zeros in the middle waveguide within the same coupling region length, hence more achievable 

intermediate levels. However, a large 𝛿 also leads to a longer device (for complete power transfer to the 

cross port) and hence larger overall loss when GST is in c-phase. To compensate for this, a thicker GST 

can be used to provide a larger 𝛥𝑛. However, the loss due to a-GST also increases. Furthermore, this also 

makes the reversible electrical switching more difficult.  



 

 

 

Figure 2.23: Optimized deterministic three-level 2 × 2 programmable units with phase-detuned TDC design 

(𝛿 = 1.5𝜅). 

(a) Schematic (left) and the cross-sectional view (right). Designed geometry parameters are listed in Table 1. 

(b) Simulated intensity distribution and (c) the corresponding transmission spectra for different GST phase 

sequences. The insertion loss is reduced to ~1.2 dB compared to ~ 3 dB in traditional phase-matched TDC 

designs.  



 

 

Based on the above discussion, we choose 𝛿 = 1.5𝜅 and design a multi-level 2 × 2 programmable unit 

with ellipsometry measured GST refractive index, (see Appendix II of 134 for detailed design procedures). 

Figure 2.23 shows the FDTD simulation results with a low insertion loss (~ 1.2 dB) and three achievable 

splitting ratios of 100:0, 50:50 and 0:100. The designed parameters are summarized in Table 2 below. 

 

Table 2: Designed parameters for the multi-level 2 × 2 GST programmable unit with 𝜹 = 𝟏. 𝟓𝜿. Unit: nm 

except for 𝒍𝒄. 

ℎ ℎ𝑠𝑙𝑎𝑏 𝑔𝑑 𝑤𝑏 𝑤ℎ 𝑔  ℎ𝑔𝑠𝑡 𝑤𝑔𝑠𝑡  𝑙𝑐 (𝜇m) 

220 100 200 630 500 315 30 450 115 

 

Figure 2.24 shows the insertion loss and extinction ratio of the optimized 2 × 2 programmable unit. The 

insertion loss is low (~1.2 dB) across a broad wavelength (60 nm) for all three levels (four configurations). 

We emphasize that this insertion loss is mainly due to the relatively thick GST film used in the design, 

which introduces absorption loss in its amorphous state. By replacing GST with lower loss PCMs such as 

GSST, we can potentially further reduce the loss. However, the index contrast Δ𝑛 of GSST is much smaller 

than GST, which may result in a thicker PCM and longer devices. The extinction ratios of the complete 

cross and bar states remain high across 60 nm, but the intermediate states are more wavelength-sensitive 

with a 3 dB variation bandwidth of ~ 22.6 nm.  



 

 

 

Figure 2.24: Simulated transmission spectrum of the designed three-level 2 × 2 TDC programmable unit 

(𝜹 = 𝟏. 𝟓𝜿). 

The device shows (a) a low loss (< 1.2 dB) for all four GST sequences and (b) a bandwidth of ~22.6 nm with 

less than 3 dB extinction ratio variation. 

 

2.9 System-level applications envision  

The presented nonvolatile 2 × 2 programmable unit, along with the 1 × 2 and 1 × 1 switches, render an 

energy-efficient and scalable solution to programmable integrated photonics. First, these are nonvolatile 

devices, and zero static energy is required to hold the state. This is very desirable for low-frequency 

programmable photonics, as shown in Table 1. Second, our device can be switched electrically. System-

level and automatic solutions can be envisioned by interfacing with mature electronic technologies. Lastly, 

the programmable units are broadband devices, which will enable full use of the parallelism operation of 

light by wavelength division multiplexing (WDM). Also, the device is relatively insensitive to thermal 

drifts and fabrication errors because of its broadband nature. 

The 2 × 2 programmable unit can be utilized in large-scale nonvolatile optical switching fabric137, as 

in Figure 2.25(a). An 8 × 8 non-blocking optical switching fabric using the classical Benes topology is 

presented. More applications can be envisioned if the 2 × 2 programmable unit can operate in a multilevel 

switching fashion, which can be potentially implemented by introducing a detuning between the SW and 



 

 

HW. For example, such devices are essential building blocks for a nonvolatile optical programmable gate 

array (OPGA). We propose to connect our 2 × 2 programmable unit with each other to form a mesh array, 

such as in Figure 2.25(b). By configuring the optical programmable unit into different states, diverse types 

of functions can be realized for optical on-chip information processing. Most OFGAs were demonstrated 

with thermal-optics effect39,41,143. But due to its volatile nature, a significant amount of energy is consumed 

to hold the configuration. PCM-based OPGA can significantly save energy thanks to its self-holding 

property and thus zero static energy consumption.  

Another possible application is optical computing. Recently there have been works on using phase 

change materials for on-chip non-Von Neumann optical computing68,69,67. These works mainly apply 

optical pulses to actuate the phase transition, limiting the device’s scalability due to complicated optical 

alignment. We propose the structure in Figure 2.25(c), which can function as an optical forward neural 

network. Although only reliable binary-level operations are feasible using the reported 2 × 2 programmable 

unit, this very similar design idea can be directly extended to using transparent PCMs such as Sb2Se3 and 

Sb2S3, which will support multi-bit operation144. The reliable operation of these devices can be achieved by 

a segmented heater and PCM design145.Also, similar to68, an optical convolutional neural network 

comprising 1 × 1 multilevel switches with electrical control can be envisioned, as shown in Figure 2.25(d). 

We expect a large-scale, nonvolatile, and in-memory optical computing platform using electrical controls. 

 

 



 

 

 

Figure 2.25: PIC schematics for various applications.  

(a) on-chip optical switching fabric. (b) multipurpose programmable PICs. (c) On-chip optical forward neural 

networks. (d) convolutional neural networks. 

 

  



 

 

Chapter 3. Electrically programmable multi-level PICs with Sb2S3
113 

Although the GST-based programmable unit in the previous Chapter was carefully designed, the loss 

in experiment was still high due to the slight fabrication variation. Moreover, the previous device is only 

suitable for binary operation due to the strict phase matching condition. In this Chapter, we aim to solve 

these longstanding problems through wide-bandgap PCMs such as antimony sulfide (Sb2S3) and antimony 

selenide (Sb2Se3). Because of their intrinsically lower optical absorption loss, a traditional two waveguide 

directional coupler design is suitable for the broadband optical switches, and we achieved low loss (<1.0 

dB), high extinction ratio (>10 dB), high cyclability (>1,600 switching events), and 5-bit operation. These 

Sb2S3-based devices are programmed via on-chip silicon PIN diode heaters within sub-ms timescale, with 

a programming energy density of ~10𝑓𝐽/𝑛𝑚3. Remarkably, Sb2S3 is programmed into fine intermediate 

states by applying multiple identical pulses, providing controllable multilevel operations. Through dynamic 

pulse control, we achieve 5-bit (32 levels) operations, rendering 0.50 ± 0.16 dB per step. Using this 

multilevel behavior, we further trim random phase error in a balanced MZI. Our work opens an attractive 

pathway toward large-scale energy-efficient programmable PICs with low-loss and multi-bit operations. 

3.1 Characterization of Sb2S3 thin films  

We started with characterizing the thin film Sb2S3 using ellipsometry, X-Ray diffraction, and Raman 

spectroscopy (Figure 3.1). A layer of 60-nm Sb2S3 was sputtered on a silicon wafer. The as-deposited a-

Sb2S3 was characterized first and then switched to the crystalline phase by annealing at 325◦C for 10 minutes 

under nitrogen flow. The Sb2S3 was characterized by ellipsometry, and the data is fitted with Cody-Lorentz 

models with low mean square error (~5). The fitted complex refractive indices in Figure 3.1(a) show a 

drastic change in the real part n (~0.7 at 1310 nm), while almost zero κ (0.0003 for amorphous and 0.03 for 

crystalline phases at 1310 nm) across the entire telecommunication O- and C-band. Therefore, switching 

the Sb2S3 produces a phase-only modulation. The micro-structural phase transition after annealing and the 

stoichiometry were verified under X-ray diffraction (XRD) and Raman spectroscopy146, as shown in Figure 



 

 

3.1(b) and (c), by the characteristic lattice constants in the XRD and wavenumber shifts in the Raman 

spectrum. These characteristics show good agreement with existing literature147.  

 

Figure 3.1: Sb2S3 material characterization.  

(a) Broadband complex refractive index fitted from ellipsometry measurement. (b) X-ray diffraction and (c) 

Raman spectroscopy measurement for the as-deposited (blue) and annealed (orange) Sb2S3 samples. The peak 

at 2θ = 65° for as-deposited Sb2S3 comes from the silicon substrate. The characteristic lattice constants and 

Raman shifts are marked. 

 

In Figure 3.2, we took a micrograph image of 60-nm-thick Sb2S3 on a blanket silicon chip, where the 

Sb2S3 films were annealed under 325°C and in the crystalline phase. The micrograph image shows a non-

uniform surface with a lot of local crystal grains and even flower-shaped patterns (circled in red), which 

could lead to unintended scattering loss for in-plane light propagation. We hypothesize that the crystal 

grains form because crystalline Sb2S3 has a density reduction of around 24% compared to its amorphous 

form105. Hence, a significant strain is induced during the phase transition, and results in a non-uniform 

surface. We also speculate that flower-like patterns occur near the nuclei.  

It is unclear what the c-Sb2S3 surface looks like on our devices because those 20-nm-thick, 450-nm-

wide Sb2S3 films are too small to image optically. The 40-nm-thick encapsulation alumina further prevents 

the accurate usage of SEM. Advanced material study is required to quantitatively understand the crystalline 

patterns formed. 



 

 

 

Figure 3.2: Non-uniform surface forms after the annealing due to the density difference between 

amorphous and crystalline Sb2S3.  

Some flower-shaped patterns are observed. We speculate the nonuniform surface is due to the large density 

difference between a- and c-Sb2S3, and the flower-shaped patterns occur near the nuclei.   

 

3.2 Simulation and design of Sb2S3-based silicon photonic components 

3.2.1 Sb2S3-based silicon phase shifters 

Figure 3.3 shows the simulated π phase shift length Lπ and insertion loss for the Sb2S3-based phase 

shifters. The refractive indices of Sb2S3 used here were obtained experimentally in the previous section. For 

a generic 500-nm-wide, 220-nm-high SOI waveguide with 20-nm-thick, 450-nm-wide Sb2S3 on the top, an 

effective index contrast of 0.018 is obtained at 1310 nm in Figure 3.3(a) and (b), indicating a 𝜋 phase shift 

length 𝐿𝜋 ≈ 38 𝜇𝑚. Figure 3.3(c) and (d) shows that a wider and thicker Sb2S3 film reduces Lπ and does 

not impact the insertion loss much (~0.23 dB/π). We adopted a generic waveguide width of 500 nm to avoid 

any extra taper structure, which renders 𝐿𝜋 ≈ 38 𝜇𝑚. The calculated mode coupling between the bare 



 

 

silicon waveguide mode and the c-Sb2S3-Si hybrid waveguide mode is 99.7% (or -0.013 dB), which is 

negligible. 

 

Figure 3.3: Simulations for Sb2S3-on-SOI phase shifters.  

(a, b) Mode simulations for (a) amorphous and (b) crystalline Sb2S3. An effective index contrast of 0.018 is 

obtained at 1310 nm, indicating a 𝜋 phase shift length 𝐿𝜋 ≈ 38 𝜇𝑚. (c, d) 𝐿𝜋 (blue) and excess loss for 𝜋 phase 

shift (orange) versus (c) the hybrid waveguide width and (d) Sb2S3 thickness. 𝐿𝜋 decreases as the increase of 

both waveguide width and Sb2S3 thickness, while the excess loss remains almost invariant. The dashed gray 

lines represent that we used a waveguide width of 500 nm and an Sb2S3 thickness of 20 nm in our experiments. 

 

We note that Lπ could be further reduced by exploiting a wider or thicker Sb2S3. Although thick Sb2S3 

gives a more compact device footprint, the complete phase transition is more difficult. The vertical 

temperature gradient at thick Sb2S3 films could lead to Sb2S3 ablation at the bottom while not high enough 

temperature for amorphization at the top. Unintentional re-amorphization could also happen with a thick 

Sb2S3 layers. Moreover, recrystallization could also occur because of the crystallization kinetics and finite 

heat diffusion time148. To accommodate this tradeoff, we pick 20 nm as the experimental thickness, 

providing both a complete, repeatable Sb2S3 phase transition and a relatively compact Lπ of 38μm. Since 



 

 

this is the first experiment to switch Sb2S3 electrically, we chose a conservative thickness of 20 nm. 

Considering the good switching behavior observed in our experiment, a thicker Sb2S3 could be used in 

future experiments to provide even more compact devices. The largest thickness of Sb2S3 films that could 

be switched completely require more experimental exploration.  

3.2.2 Sb2S3-based silicon asymmetric directional coupler 

The asymmetric directional coupler is designed using the Lumerical Mode simulator and the coupled-

mode theory. The idea has been depicted in the main text and the literature58,59,144. Figure 3.4(a) and (b) 

show two main optimization steps: (i). find the Sb2S3-loaded waveguide width with a fixed Sb2S3 thickness 

so that it is phase matched with a bare SOI waveguide; (ii). obtain a high transmission at the bar port when 

switching the Sb2S3 by optimizing the gap between two waveguides to allow the coupling length in the bar 

state to be an even multiple of the coupling length in the cross state144. Unlike the 1 × 2 coupler in ref144, 

here we consider a 2 × 2 directional coupler, where light comes in from both input ports. To achieve an 

input port independent performance, the Sb2S3-SOI hybrid waveguide and the bare SOI waveguide are 

phase matched when Sb2S3 is in the crystalline phase. Intuitively, the light interacts with the slight loss of 

the Sb2S3 waveguide half of the Lc regardless of the input port in this configuration. 

After accomplishing the optimization in the Lumerical MODE simulator, we ran a Finite Difference 

Time Domain (FDTD) simulation to verify the design, and the results are in Figure 3.4(c) - (f). The 

transmission spectra in Figure 3.4(c) and (e) show that a bar (cross) state is achieved by controlling the 

phase of Sb2S3 to amorphous (crystalline). At 1310 nm, the insertion loss is around 0.1 dB and 0.3 dB for 

the bar and cross state, respectively, and the extinction ratio is larger than 15 dB in both states. The higher 

insertion loss for the cross-state is mainly due to the slight material absorption of c-Sb2S3. The inset in 

Figure 3.4(e) shows that our device performance is symmetric and independent of the input port. Figure 3.4 

(d) and (f) shows the field propagation profile corresponding to Figure 3.4(c) and (e). Again, we can verify 

the input-port independent performance.  

 



 

 

 

Figure 3.4: Simulations for Sb2S3-on-SOI 2 × 2 tunable directional couplers with low insertion loss and 

large extinction ratios.  

(a) Phase-matching widths for c-Sb2S3-SOI and bare SOI waveguide. The dashed gray line shows that a 450-

nm-wide bare SOI waveguide is phase matched with a 409-nm-wide c-Sb2S3-SOI waveguide. Their mode 

profiles are shown in the insets with a similar effective index, indicating good phase matching. (b) The gap 

(blue) is optimized to maximize bar port transmission when the Sb2S3 is switched into its amorphous phase. 

The coupling length 𝐿𝑐  for the cross state (orange) is also obtained based on the MODE simulation and 

coupled-mode theory. An optimal gap of 420 nm is picked, corresponding to 𝐿𝑐 ≈ 79𝜇𝑚. (c - f) FDTD 

simulation results. (c, e) Transmission spectrum and (d, f) field profile for (c, d) amorphous- and (e, f) 

crystalline-Sb2S3. (i) and (ii) show the field profile when inputting light from the upper and lower port, 

respectively. The inset in (e) shows an input-port independent performance for the cross-state.  



 

 

 

On the contrary, if the waveguides are phase matched for a-Sb2S3, the loss would be much higher when 

light inputs from the Sb2S3 waveguide. Figure 3.5 shows the simulated spectrum of an optimized design, 

where 20-nm a-Sb2S3 is phase-matched with a bare SOI waveguide. As shown in the zoomed-in bar 

transmission plot in Figure 3.5(b), a difference of 0.3 dB between two input ports is observed when Sb2S3 

is in its crystalline phase. This seemingly slight asymmetry could lead to undesired, unbalanced optical 

response for different light paths in a large-scale PIC system. 

 

Figure 3.5: Phase matching a-Sb2S3-SOI waveguide with bare SOI waveguide results in an asymmetric 

(input port dependent) bar state when switching the Sb2S3 to its crystalline phase.  

(a) Transmission spectrum for c-Sb2S3. (b) Zoomed in transmission spectrum at the bar port shows a 

transmission difference of 0.3dB between two input ports.   

 

We adopted the relatively thin Sb2S3 with 20 nm thickness to ensure a complete phase transition. 

However, such a thin PCM layer provides limited tunability, resulting in a large device footprint. 

Fortunately, we note the slow crystallization speed of Sb2S3 could potentially enable a much thicker PCM 

layer to crystallize, which will shrink the device size. Using the same design methodology, we numerically 

designed asymmetric directional couplers with Sb2S3 thickness ranging from 30 nm to 50 nm in Table 3. 

Remarkably, the designed directional coupler with 50-nm-thick Sb2S3 shows a significantly reduced 

coupling length (~34 µm). 



 

 

Table 3: Performances of Sb2S3 directional couplers using different Sb2S3 thicknesses 

tSbS (nm) Wb (nm) Wh (nm) gap (nm) Lc (µm) 

20 450 409 420 79 

30 450 395 370 55 

40 450 383 330 41 

50 450 373 305 34 

Note: tSbS, Wb, and Wh represent the Sb2S3 thickness, and widths of bare SOI and hybrid Sb2S3-SOI waveguide, 

respectively. 

 

3.3 Fabricated devices and characterization results 

The silicon photonic devices were designed to operate at the telecommunication O-band (1260 – 1360 

nm) and fabricated on a standard silicon-on-insulator (SOI) wafer with 220 nm silicon and 2 μm buried 

oxide. The 500-nm-wide waveguides are fabricated by partially etching 120-nm silicon. We then deposited 

450-nm-wide Sb2S3 onto the SOI chip via sputtering. The slightly smaller width than the waveguide is to 

compensate for electron beam lithography (E-beam) overlay tolerance. The Sb2S3 films are electrically 

controlled via on-chip silicon PIN micro-heaters59,95. The p++ and n++ doping regions were designed 200 nm 

away from the waveguide to avoid free-carrier absorption loss95, indicated in the scanning electron 

microscope (SEM) images with false colors in Figure 3.6(c), Figure 3.9(c), and Figure 3.11(c). The Sb2S3 

stripes are encapsulated with 40 nm of Al2O3 grown by atomic layer deposition (ALD) under 150 °C. This 

conformal encapsulation is critical to prevent Sb2S3 from oxidation and thermal reflowing and is essential 

to attain high endurance. To show that our Sb2S3-clad silicon photonic platform is versatile and compatible 

with most PIC components, we demonstrate three widely used PIC components: (1) a micro-ring resonator 

to show low-loss tuning of cavities, (2) a balanced MZI to demonstrate a full π phase shift and a broadband 

operation, and (3) an asymmetric directional coupler to create a compact programmable unit.  

The fabrication process of the designed Sb2S3 on SOI devices is similar to what we described in Section 

2.6 and we shall not repeat it in too much detail here. 



 

 

3.3.1 Trench effect of PCM liftoff 

Notably, it was found that a narrow resist trench for Sb2S3 deposition and liftoff leads to a reduced 

Sb2S3 thickness than blanket Sb2S3 deposition. This could be attributed to that part of Sb2S3 was blocked by 

the resist due to a deep trench. We fabricated a chip with different patch width (300, 400, 500 nm) and 

measured the thickness after Sb2S3 liftoff. The measured results (Table 4) show a reduction factor of around 

0.5 compared to blanket deposition. We also measured the widths of the Sb2S3 widths which agree 

reasonably well with the Ebeam defined widths. Since this is not the main interest of this measurement, we 

only measured part of the Sb2S3 stripes, and the ones not measured are denoted as ’-’. 

Table 4: Sb2S3 thickness reduces after liftoff measured by AFM (Unit: nm) 

Deposited 

width  

Width after 

Liftoff  

Deposited 

thickness  

Thickness 

after liftoff 

300 373 30 13.8 

300 354 30 14 

400 471 30 16.9 

400 471 30 17 

300 - 40 16.4 

400 413 40 19.5 

300 - 50 21 

400 - 50 26 

400 471 50 25.9 

 

3.3.2 Nonvolatile micro-ring switch integrated with Sb2S3 phase shifter 

The characterization setup is similar to the one we described in Section 2.7.1, and we do not repeat it 

here. We deposited 10-μm-long, 20-nm-thick Sb2S3 on a micro-ring resonator with 30 µm radius, shown in 

Figure 3.6(a)-(c). The free spectral range (FSR) is ~2.42 nm in Figure 3.6(d), and the bus-ring gap is 280 

nm to achieve a near-critically coupled device. We switched the as-deposited a-Sb2S3 to c-Sb2S3 on the 

micro-ring resonator by applying three 1.6 V, 200-ms-long pulses with ~3.4 mJ energy (or SET pulses) 

separated by 1 second and then re-amorphized the material via three 7.5 V, 150-ns-long pulses with ~56 nJ 



 

 

energy also separated by 1 second (RESET pulses). The unit energy density (energy/total Sb2S3 volume) 

for switching Sb2S3 is then estimated as 0.6 𝑓𝐽/𝑛𝑚3 (38 𝑝𝐽/𝑛𝑚3) for amorphization (crystallization). We 

note that the 200-ms-long SET pulse is indeed significantly longer than other reported PCMs, such as GST 

(50 µs95 or 100 µs107) and Sb2Se3 (5 µs106 or 100 µs107), and causes a large crystallization energy and energy 

density. But we found that the SET pulse duration can be reduced to around 100 µs after the first few cycles. 

Therefore, the crystallization energy is reduced to ~1.7 𝜇J (19 𝑓𝐽/𝑛𝑚3) after the initial conditioning, 

comparable to other PCMs95,106. With the 100 𝜇𝑠  crystallization pulse condition, this device could be 

operated at ~ kHz speed for a complete SET/RESET cycle. The slow crystallization, however, can allow 

amorphizing a large volume of Sb2S3. Figure 3.6(d) shows a resonance shift of ~0.394 nm upon switching 

the 10 μm Sb2S3 from the a- (blue) to the c-phase (orange), corresponding to a π-phase shift length Lπ of 

~30.7 μm, significantly shorter than the 1-millimeter Lπ of ferroelectric non-volatile phase shifter149. Our 

simulation results show a change in effective index Δ𝑛𝑒𝑓𝑓 ≈ 0.018 between a- and c-Sb2S3 in Figure 3.3, 

which gives a theoretical 𝜋 phase shift length of 𝐿 𝜋 =
𝜆0

2Δ𝑛𝑒𝑓𝑓
= 36.4 𝜇𝑚. The slightly shorter 𝐿𝜋 in the 

experiment can be attributed to a slightly thicker Sb2S3 deposited. The SET and RESET processes were 

repeated for 10 cycles. The shift in resonance is highly repeatable, as suggested by the slight standard 

deviation in Figure 3.6(d). The excess loss from a-Sb2S3 is negligible146, and in c-Sb2S3-clad waveguides 

the loss is estimated to be 0.024 dB/μm (0.72 dB/π), which is three times larger than our simulation result 

(0.26 dB/𝜋). We attribute this excess loss to the scattering from the Sb2S3 thin film due to non-uniform 

deposition/liftoff and local crystal grains in c-Sb2S3. We verified that the loss due to mode mismatch at the 

transition between the bare silicon waveguide and the Sb2S3-loaded waveguide is small (~0.013 dB/facet), 

consistent with the fact that the thin Sb2S3 film should not significantly change the mode shape.  

 



 

 

 

Figure 3.6: A high-Q ring resonator loaded with 10 μm long 20-nm-thick Sb2S3.  

(a) Schematic of the device (the encapsulation ALD Al2O3 layer is not shown), (b) optical, and (c) Scanning-

Electron Microscope (SEM) images of the micro-ring resonator. The Sb2S3 thin film, n++, and p++ doped silicon 

regions are represented by false colors (blue, orange, and green, respectively). (d) Measured micro-ring spectra 

in two phases (SET: three 1.6 V, 200-ms-long, 3.4 mJ pulses to change Sb2S3 to the crystalline state; RESET: 

three 7.5 V, 150-ns-long, 56 nJ pulses to change Sb2S3 to the amorphous state). The spectra are averaged over 

ten cycles of reversible electrical switching, and the shaded area shows the standard deviation. Norm. Trans. 

Stands for normalized transmission, normalized to a reference waveguide on the same chip.  



 

 

 

In our experiment, single pulses could not trigger a reliable, complete phase transition. We sent in a 

single pulse for SET/RESET. The final state shows a significant variation in Figure 3.7(a). This could be 

attributed to multiple crystalline phases and incomplete thermal process due to the relatively thick Sb2S3. 

We tackled this issue by sending two or three pulses to trigger a complete phase transition. The resulting 

transmission spectra of another ring resonator is shown in Figure 3.7(b). This multi-pulse switching scheme 

achieved a much more consistent performance across ten cycles. We emphasize that this behavior is 

distinctly different compared to GST or Sb2Se3, where we were able to actuate the phase transition under a 

single shot, in a very similar PIC. While we provided a hypothesis for such behavior, more material studies 

are warranted to explain the behavior. 

 

Figure 3.7: Single SET (RESET) pulses could not reliably trigger complete crystallization or 

amorphization.  

(a, b) Transmission spectra of a ring resonator switched with (a) single SET(RESET) pulses and (b) three 

identical SET(RESET) pulses with one-second intervals. Different colors represent different experiments. 

 

3.3.3 Nonvolatile Mach-Zehnder switch integrated with Sb2S3 phase shifter 

A 120-nm partially etched MMI was designed using the Lumerical Eigen-Mode Expansion (EME) 

Method. The design was further tweaked and verified with 3D FDTD simulations. The optimized 

multimode device parameters are annotated in orange in Figure 3.8(a). Figure 3.8(a) and (b) show the 



 

 

simulated electric field propagation profile and the transmission spectra at both output ports. The optimized 

MMI exhibits an insertion loss < 0.1 dB at 1310 nm and a 1-dB bandwidth > 60 nm, as shown in Figure 

3.8(b). The measured transmission spectrum is presented in Figure 3.8(c). The slight spectral fluctuation 

(~0.1 dB) is attributed to the interference between back-reflected light at the interface between the single 

mode tapers and the multimode region. 

 

Figure 3.8: Low-loss 50:50 multimode interferometer (MMI) at 1310 nm.  

(a) Field profile and (b) transmission spectrum using FDTD simulation. The white dotted line outlines the 

device shape. (c) Measured transmission spectrum of the fabricated device. 

 

Figure 3.9(a)-(c) shows a balanced MZI with the 50:50 MMI in Figure 3.9(d) and both arms covered 

with 30-μm-long 20-nm-thick Sb2S3, which operates at wavelengths between 1,320 nm and 1,360 nm. 

Initially, the light comes out mainly from the bar port with an extinction ratio of ~13 dB in Figure 3.9(e). 

The light comes out from the bar port instead of the cross port in this balanced MZI, which can be explained 

by the random phase errors between two arms due to fabrication imperfections, especially in the S-bend 

and PCM segments. One can overcome these fabrication imperfections by exploiting a wider waveguide to 



 

 

improve fabrication robustness150. Alternatively, this random phase error can be post-fabrication trimmed 

using Sb2S3, as we show later in Section 3.5.  

We measure the current-voltage (IV) response on an MZI arm with a doped region length of 30 𝜇𝑚 by 

applying a voltage from -0.5 to 1.0 Volt in Figure 3.10(a). The measured IV-curve shows a typical rectified 

shape and has a turn-on voltage of around 0.8 V and resistance of 50 Ω. The Sb2S3 on one arm was switched 

by two 1.7 V, 200-ms SET electric pulses with an energy of 11.6 mJ (density: 42 𝑝𝐽/𝑛𝑚3) to provide a full 

π phase shift. An 8.1 V, 150-ns short RESET electric pulse with an energy of 197 nJ (density: 0.73 𝑓𝐽/𝑛𝑚3) 

switched the device back to the initial state. Figure 3.9(e) and (f) show the transmission spectra normalized 

to a reference waveguide when the Sb2S3 film is in the amorphous and crystalline phases, respectively. The 

c-Sb2S3 displays a complete spectrum flip, showing a bar state with an extinction ratio of 15 dB.  



 

 

 

Figure 3.9: A Mach-Zehnder interferometer with both arms covered with 20-nm-thick Sb2S3.  

(a) Sb2S3 phase shifter schematic (the encapsulation ALD Al2O3 layer is not shown). (b) Optical and (c) SEM 

images of the Sb2S3 phase shifter. (d) Optical micrograph of the 50:50 splitting ratio multi-mode interferometer. 

(e, f) Transmission spectra at both bar and cross ports for (e) a-Sb2S3 (RESET: an 8.1 V, 150 ns, 197 nJ pulse) 

and (f) c-Sb2S3 (SET: two 1.7 V, 200 ms, 12 mJ pulses). The green and red lines represent the measured 

transmission at cross and bar ports. The shaded region indicates the standard deviation of transmission for 5 

cycles of measurements. 

 



 

 

The Sb2S3 phase shifter in an MZI was switched reversibly by alternatively applying SET and RESET 

pulses. We note this cyclability test was performed with a single pulse scheme, so a significant variation 

was observed. In the later characterization for asymmetric directional coupler, we instead used three pulses 

that were used for each amorphization or crystallization, providing much better repeatability. Yet, we found 

100 consecutive switching events in Figure 3.10, where the device was switched with a high extinction ratio 

of ~15 dB. 

  

Figure 3.10: Extra measurement results on MZIs.  

(Left) I-V characteristics of the 30-𝜇𝑚-long PIN diode, showing a turn on voltage of 0.8 V and a resistance of 

50 Ω. (Right) The Sb2S3-based phase shifter in an MZI is experimentally switched between amorphous (blue) 

and crystalline (orange) phases for 100 switching events. Each switching event was triggered with a single 

pulse. 

 

3.3.4 Asymmetric directional coupler silicon switch with Sb2S3 

We also designed and fabricated a compact asymmetric directional coupler (coupling length Lc ≈ 79 

μm), as shown in Figure 3.11(a)-(c). The coupler consists of two waveguides with different widths. The 

narrower 409-nm-wide waveguide (hybrid waveguide) was capped with 20-nm-thick Sb2S3 and designed 

to allow phase match with the wider 450-nm-wide waveguide (bare waveguide) for c-Sb2S3
144. As such, the 

input light could completely couple to the cross port in one coupling length. Once Sb2S3 is switched to the 

amorphous state, the effective index of the hybrid waveguide changes while the bare waveguide remains 

the same. The resulting phase mismatch changes the coupling strength and coupling length. Then, co-



 

 

optimizing the gap and waveguide length permits a complete bar transmission. The unique c-Sb2S3 phase 

matching approach, instead of a-Sb2S3
57,59,144,151, allows a more symmetric performance regardless of the 

input port, crucial for a 2 × 2 device. If phase mismatch happens in the c-Sb2S3 state, the slight loss of c-

Sb2S3 on one of the waveguides will result in different bar state insertion loss when the light goes from 

different input ports. We note that the 79-μm coupling length can be potentially reduced (~ 34 μm) by 

depositing a thicker (50 nm) Sb2S3 to provide stronger refractive index modulation.  



 

 

 

Figure 3.11: An asymmetric directional coupler with Sb2S3-Si hybrid waveguide.  

(a) Schematic (the encapsulation ALD Al2O3 layer is not shown), (b) optical, and (c) SEM images of the 

asymmetric directional coupler. (d, e) Transmission spectra at bar and cross ports for (d) a-Sb2S3 (RESET: 

three 9.6 V, 500 ns, 922 nJ pulses) and (e) c-Sb2S3 (SET: three 2.7 V, 200 ms, 29.2 mJ pulses). The result was 

averaged over five measurements, and the shaded region indicates the standard deviation.  

 



 

 

Figure 3.11(d) and (e) show the transmission spectra for a- and c-Sb2S3, switched with three 9.6 V, 500 

ns, 922 nJ RESET, and 2.7 V, 200 ms, 29.2 mJ SET pulses, respectively. The energy density for 

amorphization (crystallization) is 1.28 𝑓𝐽/𝑛𝑚3 (40 𝑝𝐽/𝑛𝑚3). The insertion losses are 2 dB (0.5 dB), and 

the extinction ratios are around 10 dB (11 dB) for a (c)-Sb2S3). The unexpected high insertion loss when 

the Sb2S3 is in an amorphous state can be attributed to several factors, including deviation of the gap size 

from the design, Sb2S3 overlay alignment error, or the cross-port grating coupler fabrication imperfection. 

To estimate the actual loss of the device, we apply the c-Sb2S3 loss extracted from the ring resonator to the 

simulation, which includes the crystal grain scattering loss, by matching the loss measured in the high-Q 

resonator (0.76 dB/𝜇m) with MODE simulation. We obtained an extinction coefficient 𝜅′ ≈ 0.015 , 

compared to ellipsometry measurement result of 𝜅 = 0.005 at 1310 nm. We then estimate the actual loss 

of the asymmetric directional coupler using the new complex refractive index. Figure 3.12 shows the 

simulation results, where the insertion loss is around 0.85 dB, and the extinction ratio is 25 dB. This loss 

taking into consideration of the grain boundary scattering is similar to what we observed in device 

characterization, which emphasizes the importance of further optimizing the material for a smaller 

scattering loss by reducing grain size of crystalline PCMs152. We note that the amorphous Sb2S3 loss is very 

small in the experiment, so the initial simulation result still holds (0.1 dB insertion loss and 15 dB extinction 

ratio). 

 

Figure 3.12: Simulated transmission spectrum for asymmetric directional coupler.  



 

 

(a) Simulation results with the ring resonator measured c-Sb2S3 loss, showing high extinction ratio > 20 dB. 

(b) Zoomed-in plot to show the insertion loss. The insertion loss increases from 0.3 dB to 0.85 dB due to the 

high material absorption loss used.  

 

Figure 3.13 shows more than 1,600 switching events for the asymmetric directional coupler, a first 

demonstration of large-endurance electrical switching of Sb2S3 at the time of the work. We note that very 

few electrical controls of Sb2S3 had been experimentally demonstrated105,146 and no work had previously 

shown reversible electrical switching of Sb2S3. Limited by our measurement setup, we separately measured 

the cross and bar ports in Figure 3.13(a) and (b), respectively. The higher insertion loss (~1 dB) at around 

event 500 was due to optical fiber misalignment. We note that almost no performance degradation occurred 

at the end; hence, 1,600 switching events are not the limit of this device. We stopped the experiment at that 

point due to the long duration of the experiment. The cross-port transmission shows a relatively large 

variation for a-Sb2S3 (from ~ -15 dB to ~ -35 dB) in Figure 3.13(a), which was caused by incomplete 

amorphization or thermal reflow of the Sb2S3 film. Since the plot is on a logarithmic scale, such a large 

variation in a-Sb2S3 (due to higher transmission) is not visible in Figure 3.13(b). 

 

Figure 3.13: Cyclability test for a-Sb2S3-based asymmetric directional coupler.  

Measured transmission at the (a) cross and (b) bar ports. The blue and orange scatterers represent the 

normalized transmission when Sb2S3 is in the amorphous and crystalline phases. The phase change condition 

is the same as in Figure 3.11. The device was switched for over 1,600 events with no significant insertion loss 

and performance degradation. 



 

 

 

Experimentally, we observed a very slow Sb2S3 crystallization for the first time, where even 500 µs 

pulses could not trigger crystallization. The devices were then crystallized with DC voltage or pulses with 

very long durations such as 200 ms. This could be explained by the small enthalpy of fusion of Sb2S3 

(~40.64 kJ/mol153, compared to fusion enthalpy154 of GST 625 J/cm3  or 
625

𝜌
𝑀 =

625 𝐽/𝑐𝑚3

5.87 𝑔/𝑐𝑚3 ⋅

1026.8 𝑔/𝑚𝑜𝑙 = 109.3 𝑘𝐽/𝑚𝑜𝑙 , where 𝜌  is the density155 and 𝑀  is the molar mass of GST), which 

necessitates a large critical nuclei size to overcome the crystallization energy barrier. After the first 

crystallization, however, the successive crystal growth process happens without any requirement for 

overcoming the energy barrier. We were able to see partial crystallization using much shorter 100-µs pulses, 

as shown in Figure 3.14. We note this slow crystallization guarantees larger volumes of Sb2S3 amorphized 

again without recrystallization and thus could be beneficial for photonic applications. In the next section, 

we experimentally show that indeed the amorphization could be triggered with very long pulses. 

 

Figure 3.14: The slow crystallization nature of Sb2S3 limits the pulse duration longer than 100 𝝁𝒔.  

The 100-𝜇𝑠-long pulses were not able to trigger complete crystallization even after the first time. We note 

again that during the first few cycles of crystallization, only pulses with even much longer duration, such as 



 

 

200 ms, could trigger crystallization. This indicates the difficulty in forming initial Sb2S3 nuclei and the 

substantially faster rate of crystal growth compared with nucleation. 

 

One appealing aspect of using "slow" PCMs, such as Sb2S3, is that the slow crystallization process 

could avoid any unintentional recrystallization during the melt-quench process, hence could ensure the 

amorphization for a large volume of PCM. This capability of completely switching large volumes of PCMs 

is crucial in photonics, where the volume (typically orders of μm3) are much larger than in electronic 

applications (typically orders of (10 nm)3) and takes precedence over the phase transition speed1. 

Here, we verify Sb2S3’s resistance to unintentional recrystallization by comparing the longest 

amorphization time of an asymmetric directional coupler to other reported PCM devices. Figure 3.15 shows 

that the Sb2S3 on the directional coupler could be amorphized under different pulse durations, ranging from 

500 ns to 10 μs. A large degree of amorphization was observed even when we increased the pulse durations 

to 10 μs. This 10-μs pulse duration is much longer than in the reported GST59,107,111 (< 200 ns) or SbSe106 

(< 1 µs, but the thickness is 30 nm) switching experiments. This successful amorphization with long pulses 

supports that Sb2S3 is inherently more suitable for large volume amorphization because recrystallization is 

suppressed. Moreover, the long amorphization pulses relaxed the requirements of high voltage from ~10 V 

down to 5V, showing a potential path towards CMOS compatible voltage levels (1~2 V). 



 

 

 

Figure 3.15: Complete amorphization was triggered with relatively long, 10-𝝁𝒔 pulses because of the slow 

crystallization nature of Sb2S3.  

High-level port (bar port) transmission is consistent among all the pulse conditions. The low-level port (cross 

port) transmission variation could be attributed to a tiny portion of the Sb2S3 not being switched completely, 

which could be potentially overcome by optimizing the heater design. 

 

We switched the asymmetric directional coupler for 2,300 switching events. However, some drift in 

the optical measurement setup occurred and we only reported 1,600 cycles in Figure 3.13. Here, we compare 

the SEM image and the transmission spectrum before and after 2,300 switching events and explain the 

performance degradation or the transmission drifting. Figure 3.16(a) and (b) show the SEM images of Sb2S3 

directional couplers before and after the cyclability test. We note that the black dots on the lower waveguide 

in Figure 3.16(a) are likely the crystalline Sb2S3 nuclei formed during the rapid thermal annealing. The 

rough Sb2S3 edge in Figure 3.16(b) indicates that switching the relatively large volume of the Sb2S3 stripe 

many times results in a distorted Sb2S3. This is because the thin film always tends to minimize the surface 

energy during melt-quench process, hence leads to sinusoidal edge. This phenomena is known as Rayleigh 

instability of fluid motions156. The result of such surface-energy-induced instability could be discrete 

circular patterns, having been used for advanced nano-fabrication157. Therefore, to further improve the 



 

 

endurance of our devices, one possible approach is to pattern the long Sb2S3 stripes to subwavelength 

gratings to ensure a lower initial surface energy and a low excess loss. We note that this subwavelength 

grating idea has been experimentally demonstrated in other works using GST92,158,159, hence could be a 

future direction to improving the presented work. 

The transmission spectra before and after the cyclability test are shown in Figure 3.16(c). A minor drift 

in the spectra could be attributed to the thermal reflowing, which changes the Sb2S3 shape, hence breaking 

the phase matching condition. Nevertheless, the device still shows a low insertion loss (< 1 dB degradation) 

and high extinction ratio (~ 15 dB).  

 

Figure 3.16: Only slight device degradation occurred after 1,150 cycles (2,300 switching events), and the 

SEM images suggest it is due to thermal reflowing.  

(a, b) SEM images of a device after (a) a single rapid thermal annealing and (b) 1,150 cycles. Scale bar: 1 µm. 

(c) Transmission spectra before, in the middle, and at the end of the cyclability test. No significant performance 

degradation in the cross port (blue) was observed after switching the asymmetric directional coupler for more 

than 1,150 cycles. The variation in the bar port could be attributed to the change in Sb2S3 distribution due to 

the thermal reflowing or Rayleigh instability issue. 

 

3.4 Multilevel 5-bit operation with dynamic electrical control 

Our Sb2S3-Si integrated structures further show a stepwise multilevel operation up to 32 levels with 

dynamic pulse control. In Figure 3.17(a), we show the multilevel transmission at both cross and bar ports 



 

 

of an asymmetric directional coupler while sending in RESET 10 V, 550 ns, 1.1 𝜇J pulse every other second 

to amorphized the Sb2S3. We started the experiment with “coarse” tuning, where unoptimized, identical 

pulses were sent to partially amorphize the c-Sb2S3 device to demonstrate multiple levels. The asymmetric 

directional coupler was originally in the “cross state” (red region). After one partial amorphization pulse, it 

was reconfigured into an intermediate state (orange region), where light comes out from both cross and bar 

ports. After six pulses, a complete “bar state” (green region) was achieved. We repeated this experiment 

five times for each port and plotted the average transmission levels and the standard deviation. The variation 

is attributed to the stochastic phase change process using electrical controls160. Partial crystallization is 

briefly discussed in Figure 3.18. In the following experiments, we mainly focused on partial amorphization 

because of lower energy consumption and more operation levels with finer resolution.  

Such stepwise multilevel operation by applying identical pulses is distinctly different from previously 

reported multilevel operations in GST59,95 and Sb2Se3
106,107, where different voltage amplitudes or pulse 

duration were used to access multiple levels during amorphization. The pulse-number-dependent behavior 

is quite counterintuitive: one expects that after the first amorphization pulse, the thin Sb2S3 film would have 

reached its new equilibrium phase. Moreover, since the thermal processes relax within 10 μs, each pulse is 

independent due to the relatively long one-second interval. As a result, the subsequent identical and 

separated pulses should not further change the material phase. To understand the origin of the multi-level 

operation, we closely inspected four partially amorphized Sb2S3 devices under the microscope. We 

observed a few separate patches and a region that grew with more voltage pulses. As reported in some 

literature, one possibility could have been that a- and c-Sb2S3 have significantly different thermal 

conductivities and specific heat capacities. But we measured the thermal conductivities to be similar (a-

Sb2S3: 0.2 W/m/K; c-Sb2S3: 0.4 W/m/K), and hence, we ruled out this as a possible explanation. We 

hypothesize that this unique behavior comes from Sb2S3’s multiple crystalline phases. Sb2S3 has at least 

two distinct crystalline phases161, which may differ in the amorphization conditions. The partial 

amorphization pulse can cause amorphization in the hottest region, but at the lower temperature region, it 



 

 

may cause phase transition to the other crystalline phase. These regions get amorphized in subsequent 

pulses, resulting in multilevel operation. 

 

Figure 3.17: A quasi-continuously tunable directional coupler based on multilevel Sb2S3.  

(a) (Coarse tuning applying identical electrical pulses) Time trace measurement of the directional coupler when 

sending in an amorphization pulse (10 V, 550 ns, 1.1 𝜇J) each second. Green (red) curves represent cross (bar) 

transmission. The result is averaged over five experiments for each port, and the error bar shows the standard 

deviation. Depending on the pulses, a “bar”, “intermediate”, and “cross” state can be achieved as indicated by 

the red, orange, and green regions. (b) (Fine tuning using dynamically controlled near-identical electrical pulses) 

Normalized transmission at 1,340 nm at the bar port shows 5-bit operation (32 distinct levels), achieved by 

dynamically controlling the number, amplitude, and duration of pulses sent in (near-identical, 9.65 V ~ 9.85 

V, 550 ns, 1.02 𝜇J to 1.06 𝜇J). A precise transmission level of 0.50 ± 0.16 dB per step and 32 levels were 

simultaneously achieved. The only slight difference between the target and achieved transmission demonstrates 

an on-demand operation. The error bars represent the standard deviation over five experiments.  

 

An even finer multilevel operation was realized by monitoring the transmission level and dynamically 

changing the pulse conditions slightly. Here, we demonstrate on-demand 5-bit operation in a quasi-

continuously tunable directional coupler, as shown in Figure 3.17(b). We dynamically controlled the partial 

amorphization pulses to have slightly lower, near-identical voltages (ranging from 9.65 V to 9.85 V) and 

obtained up to 32 levels. Figure 3.17(b) demonstrates 5-bit operation (32 distinct levels) with a target 

resolution of 0.5 dB per level step at 1,340 nm. We emphasize that dynamic control is necessary to mitigate 

the stochastic nature of electrically controlled PCMs160, hence essential for a reliable many-level operation. 

Furthermore, the stepwise multi-level operation allows precise control by gradually approaching the desired 



 

 

operation level, whereas the voltage or pulse duration dependent approaches result in unrepeatable multi-

level operation, limited by PCM’s stochastic nature. In Figure 3.17(b), a linear fit shows a slope of -0.50 

dB per step and a standard deviation of 0.16 dB among five experiments, indicating a repeatable operation. 

While the 5-bit operation of GST was shown using laser pulses162, our demonstration is thus far the highest 

number of operating levels reported using electrical control in PCMs-based photonics. Moreover, our 

multilevel operation does not require sophisticated heater geometry engineering, such as the segmented 

doped silicon heater design106, and solely relies on the unique phase-change dynamics of Sb2S3. We note 

that the deviation from the ideal level mostly comes from the over-amorphization of the device, and the 

error can be reduced by incorporating a “backward” tuning using partial crystallization pulses. 

Implementing a more sophisticated heater geometry106,114 could also potentially increase the number of 

levels. 

We also experimentally tested the partial crystallization of the device and multilevel operation, which 

is based on the growth-dominant nature of Sb2S3 crystallization process144. By sending in relatively low 

voltage crystallization pulses (voltage 3.1 V, duration 200 ms, leading-edge 10 ms, falling edge 100 ms), 

we obtained partial crystallization performance in Figure 3.18. Such behavior could be explained by the 

growth-dominant nature of Sb2S3
144. The growth rate could be controlled by the temperature (pulse 

voltages); the total crystallization volume could be controlled with the duration of crystallization pulses. 

Further engineering the pulse parameters could potentially achieve more operation levels. 



 

 

 

Figure 3.18: Stepwise partial crystallization using identical, low amplitude, and long duration pulses with 

a one-second interval.  

Continuous time trace measurement at (a) bar and (b) cross port of an asymmetric directional coupler. The 

experiment was repeated 10 times at each output port, represented by different colors, and the multilevel 

characteristic occurred in all experiments. 

 

We inspected devices at different intermediate levels under the microscope in Figure 3.19. One can see 

a gradual increase of amorphous Sb2S3 as the degree of amorphization increases. The amorphous Sb2S3 is 

not located precisely in the center of the long strip, where the temperature is the highest. We attribute this 

to the possible non-uniform doping profile and recrystallization of Sb2S3 near the center. 



 

 

 

Figure 3.19: Micrograph images for directional couplers in different intermediate levels.  

(a)-(d) The levels are set using identical pulses with one-second intervals. The degree of amorphization 

increases from. The dark (light) green regions on the lower waveguide are a(c)-Sb2S3. Islands of a-Sb2S3 

(circled in red) are visible in (b) and (c) and continue to grow. (e) A zoom-in picture of the right-most circled 

region in (b). A clear boundary between a- and c-Sb2S3 is observed in the Sb2S3/Si hybrid waveguide (the lower 

waveguide). 

 

3.5 Random phase error correction in balanced MZIs exploiting multilevel 

operation 

Finally, exploiting the multilevel operation, we corrected random phase errors in a balanced MZI. A 

perfectly balanced MZI should initially be in an all-cross state. However, random phase errors due to 

fabrication imperfections can easily build up to a phase error of π, making the initial state unpredictable. 

Therefore, balanced MZIs usually require extra calibration150. For example, Figure 3.20(a) shows the 



 

 

transmission spectra of a phase error corrupted balanced MZI, indicating a high bar transmission. Both arms 

of the MZI are loaded with 40-µm-long Sb2S3 film to guarantee a phase tuning range of more than ±π. The 

corrected MZI spectra by multilevel tuning of Sb2S3 are demonstrated in Figure 3.20(b), showing a pure 

“cross state” with a high extinction ratio of 24 dB. The trimming process is shown in Figure 3.20(c). We 

sent in a partial amorphization pulse (8.8 V, 150 ns, 232 nJ) every other second, which gradually increased 

the portion of amorphous Sb2S3, resulting in quasi-continuous changes of the bar (blue) and cross (orange) 

transmission (at 1,340 nm). The correction finishes once a bar transmission minimum is reached, indicated 

by the red arrows in Figure 3.20(c), and the spectra in Figure 3.20(b) were then measured. Further pulses 

increase bar transmission because of the over-compensated phase. We performed the same experiment three 

times, indicated by different colored regions in Figure 3.20(c). Complete phase error correction was 

observed in all three instances, exhibiting excellent repeatability of our trimming process. Note that binary 

tuning cannot accomplish this task due to the random initial phase error. Even multilevel operations with 

limited discrete-level resolution can cause over- or under-corrected phase error, ultimately determining the 

trimming resolution. We highlight that this method requires zero static energy supply once the phase error 

is corrected, as the phase transition in Sb2S3 is non-volatile. Thanks to the relatively fine operation levels, 

slight over-tuning does not significantly affect the performance. The trimming resolution can be further 

improved using dynamically changed, near-identical electrical pulses, as shown in the previous Section. 

Moreover, if the phase error is over compensated, the device could be tuned back with partial crystallization 

pulses. In the future, our trimming process can potentially be fully automated by real-time adjusting the 

pulse numbers according to the measured transmission.  



 

 

 

Figure 3.20: Random phase error correction in a balanced MZI based on the multilevel operation.  

(a) Transmission spectrum of a balanced MZI with phase error due to fabrication imperfections. The red (green) 

line represents bar (cross) transmission. Light transmits mostly from the bar port instead of the ideal cross port. 

(b) Transmission spectrum after the correction with 8.8 V, 150 ns, 232 nJ pulses. The device is in a “cross 

state” with a high extinction ratio of ~24 dB, indicating a phase-error-free device. (c) Time trace measurement 

for three experiments shows the phase error correction step. The RESET pulses are sent at a one-second interval. 

As more pulses were sent in, the bar transmission continuously decreased until it reached a minimum of ~-24 

dB (red arrow), and then it went up. In all the experiments, the optimal error correction was obtained. The 

slight volatile increase (only visible when transmission is smaller than -12 dB) after each amorphization pulse 

happens over a much longer timescale than 10 µs thermal relaxation time and could be attributed to a weak 

persistent photocurrent effect and requires more study. However, the transmission stabilizes before the next 

pulse. The blue (orange) represents bar- (cross-) transmission. 

 



 

 

3.6 Discussion 

Among the first few cycles (~50 cycles), sub-millisecond pulses failed to trigger the crystallization for 

Sb2S3. We increased the voltage of a 500-μs pulse up to the material ablation voltage level without observing 

any crystallization. This suggests that the thermally induced Sb2S3 crystallization process is relatively slow, 

different from laser-induced crystallization105, where a crystallization speed of tens of nano-second was 

demonstrated. We attribute this behavior to the difficulty in the initial nucleation. This hypothesis is 

supported by the fact that after the first few crystallization processes (conditioning), consecutive 

crystallization process could happen only at 100 μs, increasing the speed and lowering the energy by three 

orders of magnitudes. Hence, crystallization becomes easier after the initial nuclei/ defects are formed by 

the first few long thermal processes. We used 200-ms pulses in the reported experiments because the post-

conditioning, shorter 100-𝜇𝑠 pulse condition was identified later in our experiments. 

However, this limited crystallization speed (~ kHz) should not be considered as a problem, because of 

PCMs’ main advantage of nonvolatility, and thus zero-static power consumption. This nonvolatile 

reconfiguration technology is complementary to the traditional high-speed modulators, and will find wide 

applications with infrequent programming needs, where high speed is not essential and the zero-static power 

is more important1,163. Moreover, the slow crystallization speed could prevent unintentional recrystallization 

during long-pulse amorphization, which has a slower cooling rate. Since long-pulse amorphization means 

a smaller vertical thermal gradient, a thicker layer of Sb2S3 can potentially be switched completely. We 

verified that a pulse with 10 µs duration was able to trigger a large degree of amorphization, which is not 

possible in fast PCMs, such as GST. This capability of actuating phase transition in thicker films is useful 

in photonics, since switching a larger volume of PCMs is generally required to provide stronger modulation. 

In addition, such long amorphization pulse reduces the voltage from >10 V (for 500 ns pulses) to ~5 V, 

showing a critical step toward CMOS-compatible operation levels (1~2 V). The voltage could be further 

reduced by designing the heater geometry to improve the thermal delivery efficiency164, such as etching the 



 

 

oxide beneath the silicon165, adding thermal insulation166 layers or using atomically thin 2D material 

heaters107.  

  



 

 

Chapter 4. PCM integration in high-volume silicon photonics167
 

In the previous Chapters, we have shown the usage of low-loss PCMs can significantly improve the 

performance of optical components, such as lower insertion loss, higher extinction ratio, and capability of 

achieving multi-level operation. However, both GST and Sb2S3 components were fabricated through an in-

house fabrication process, which involves e-beam lithography and is not suitable for high-volume 

fabrication. In this Chapter, we demonstrate a scalable platform harnessing the mature and reliable 300-mm 

silicon photonic fab, assisted by an in-house wide-bandgap PCM (Sb2S3) integration process. We show 

various non-volatile programmable devices, including micro-ring resonators, MZIs and asymmetric 

directional couplers, with low loss (~ 0.0044 dB/µm), large phase shift (~ 0.012π/µm) and high endurance 

(> 5,000 switching events). Moreover, we showcase this platform’s capability to handle relatively complex 

structures with multiple PIN diode heaters, each independently controlling an Sb2S3 segment. By reliably 

setting the Sb2S3 segments to fully amorphous or crystalline state, we achieved deterministic multilevel 

operation. An asymmetric directional coupler with two unequal-length Sb2S3 segments showed the 

capability of four-level switching, beyond cross-and-bar binary states. We further showed unbalanced MZIs 

with equal-length and unequal-length Sb2S3 segments, exhibiting reversible switching and a maximum of 

5 (𝑁 + 1,𝑁 = 4) and 8 (2𝑁 , 𝑁 = 3) equally spaced operation levels, respectively. This work lays the 

foundation for future programmable very-large-scale PICs with deterministic programmability.  

4.1 Reproducible zero-change integration of PCMs on silicon photonics 

from a 300-mm fab 

The in-house fabrication process flow for integrating PCMs onto reticles from 300 mm Intel fab is 

shown in Figure 4.1. The initial pure silicon photonic chips were fabricated in a 300-mm semiconductor 

fab in Intel Corp., with silicon etch, doping, top oxide layer growth and metal vias growth included. We 

note that metallization was not available from the commercial manufacturing process at the time of tape-



 

 

out, but it is now possible with a thicker SiO2 top cladding layer. We can include the metallization in the 

future once we develop and test the window opening process for thicker SiO2 cladding. 

 

Figure 4.1: In-house fabrication workflow diagram.  

 

Following is the detailed fabrication process as shown in Figure 4.1. The 300-mm wafer was then diced 

to 2.5 cm × 3.3 cm reticles for in-house Sb2S3 integration. The oxide window for Sb2S3 deposition on the 

intrinsic silicon waveguide was defined by an overlay using a direct write laser lithography tool (DWL, 

Heidelberg DWL66+) with adhesion-promoting primer HMDS (Yeild Engineering Systems, LP 3 A) and 

a positive tone resist AZ-1512 (~1 µm), followed by a partial dry etch of ~ 600 nm SiO2 in a fluorine-based 

inductively coupled plasma etcher (ICP, Oxford PlasmaLab 100 ICP-18). The reticle was then immersed 

in 10:1 buffered oxide etcher (BOE) for 3~4 mins for complete SiO2 removal. To form an ideal Ohmic 

contact, removal of the surface oxide (~100 nm) on the metal vias was done by a second DWL overlay and 

immersing the chips in 10:1 BOE for 75 seconds. The metal contacts were then immediately patterned by 

a third DWL overlay using a negative tone resist NR9G-3000PY. Metallization of Ti/Pt (15 nm/200 nm) 



 

 

was done by electron-beam evaporation (CHA SEC-600) and lift-off. The Sb2S3 window was defined by 

JEOLJBX-6300FS 100 kV electron-beam lithography (EBL) using positive-tone resist double layer 

P(MMA-MAA) Copolymer and PMMA for high-quality liftoff process. A layer of 40-nm-thick Sb2S3 thin 

film was deposited using an Sb2S3 target (Plasmaterial Ltd.) in a magnetron sputtering system (Lesker Lab 

18), followed by a lift-off process. We note the actual Sb2S3 thickness after the liftoff was reduced to ~ 20 

nm due to the trench effect, that the narrow resist trench lowered the deposition rate. We then encapsulated 

the Sb2S3 with 40-nm-thick Al2O3 through thermal ALD (Oxford Plasmalab 80PLUS OpAL ALD) at 150 

°C. To ensure good contact between the electric probe and metal pads while applying electrical pulses, the 

Al2O3 on the metal contacts was removed by defining a window using a fourth DWL overlay with positive 

tone resist AZ1512, then etching in a chlorine-based inductively coupled plasma etcher (ICP-RIE, Oxford 

PlasmaLab 100 ICP-18).  

To highlight the simplicity of this fabrication process, we compare it with the in-house approach in 

Section 2.6. E-beam lithography (EBL) is one of the limiting factors for scalability and costs much money 

and time. The PCM integration process for Intel reticles only takes one EBL to precisely place the PCM 

thin films on the silicon waveguides for asymmetric DCs, which could be eliminated for micro-ring 

resonators and MZIs by using Heidelberg to define a PCM stripe much wider than the alignment error of 

Heidelberg (±0.5 µm). We note the Heidelberg can be replaced with a deep-UV stepper for 

photolithography, which usually has higher alignment accuracy, smaller critical feature size, as well as 

better scalability. In contrast, our previous fully in-house fabrication process used six EBL steps: (1) define 

the photonic structures such as waveguides, micro-rings, grating couplers etc.; (2) and (3) define n++ and 

p++ doped region; (4) define metal pads; (5) define the PCM pattern; (6) define the Al2O3 etch window. In 

addition, additional fabrication steps were required, such as silicon etching, ion implantation, and annealing 

for dopant activation. Therefore, we conclude that the process of integrating PCMs onto Intel reticles takes 

significant less time than the fully in-house fabrication approach.  

 



 

 

Figure 4.2(a) shows a schematic of the PCM-silicon hybrid photonics platform, on which different 

functional optical components are fabricated, including MRRs, MZIs and asymmetric DCs. All the 

components are armed with p++-intrinsic-n++ (PIN) doped silicon heaters for in situ electrical tuning of 

the PCMs. Electrical signals are applied on the platinum pads, which are well separated from the optical 

waveguides and are connected to the doped silicon regions by vertical metal vias. We emphasis that such 

multi-layer fabrication is crucial for low-loss routing of metal wires in VLS programmable PICs and is not 

available in our in-house fabrication. Figure 4.2(b) shows the photograph of a 300-mm wafer fabricated by 

Intel, which is diced into 2.5 cm × 3.3 cm reticles.  

We developed an in-house process to open 4-μm-wide oxide windows followed by Sb2S3 deposition 

and patterning. We performed optical mode simulations to quantify the scattering loss due to the oxide 

window, suggesting less than -0.01 dB scattering loss (99.8% power coupling) per window. Such a small 

scattering loss comes from our 300-nm-thick waveguide, which confines 90.6% of optical mode in the 

silicon core. The in-house fabrication process can potentially be extended to whole-wafer level. An optical 

micrograph of the reticle after Sb2S3 integration is shown in Figure 4.2(c).  

 



 

 

Figure 4.2: Schematic and photograph of the fabricated wafers and reticles.  

(a) Schematic of the silicon-PCM hybrid platform. The 300-mm semiconductor fab promises high-volume 

manufacturing, and many reticles can be fabricated simultaneously for fast prototyping. The PCMs are 

integrated in-house by opening an oxide window on the optical waveguides followed by deposition and 

patterning processes. A schematic showing the cross-sectional view of the PCM-silicon hybrid platform is on 

the right. (b) Photograph of the 300-mm pure silicon photonic wafer, which is diced into multiple 2.5 cm × 3.3 

cm reticles. (c) Optical micrograph of a reticle after integrating the low-loss PCM Sb2S3, showing various 

optical components such as micro-ring resonators (MRRs), Mach-Zehnder interferometers (MZIs), asymmetric 

directional couplers (DCs), quasi-continuously (QC) tunable MZIs and DCs. (Scale bar: 500 µm) 

 

We note that the excess optical loss arising from such in-house fabrication is negligible as optical mode 

perturbation caused by the oxide etching and Sb2S3 deposition is minor. This was verified by similar quality 

factors (Q-factor) of the MRRs, which only slightly reduced from ~1.27×105 to ~1.01×105 as fitted by 

Lorentz model in Figure 3.6. This implies an excess loss of only 0.02 dB for 20-µm-long Sb2S3. We 

highlight that such a highly scalable platform can immediately enable screening and fast testing of different 

PCMs by simply changing the sputtering targets or using other deposition methods such as evaporation. 

We also emphasize that the PCM integration process is at the back end of a 300-mm fab line, and zero 

change is required on the 300-mm fab process, crucial to reliable and massive production.  

 

Figure 4.3: Measured optical transmission spectra for a micro-ring resonator with 20-μm-long oxide 

window.  



 

 

(a) The response of a micro-ring resonator before any in-house process. Fitting to a Lorentz shape gives a Q-

factor of 127k in the zoomed in plot. (b) The response of a micro-ring resonator after opening oxide windows, 

metallization and 20-nm-thick Sb2S3 deposition. Fitting to a Lorentz shape gives a Q-factor of 101k in the 

zoomed in plot, indicating very small extra loss due to the in-house fabrication process of ~0.02 dB, estimated 

using the formula in ref52. 

 

4.2 Characterization results  

4.2.1 Experimental measurement results 

We first demonstrate reversibly switchable MRRs and MZIs on this platform using electrical control. 

We will further show asymmetric DCs with deterministic multi-level behavior in the next section. Although 

the devices were designed and tested at telecom O-band (1,260 ~ 1,360 nm), we do not foresee any 

fundamental problem with extending the operation to the telecom C-band (~1,550 nm) as Sb2S3 has even 

lower absorption loss in the longer wavelength regime (see the optical refractive index in Figure 

3.1)104,113,146.  



 

 

 

Figure 4.4: Reversible reconfiguration of micro-ring resonators and Mach-Zehnder interferometers 

through electrically controlled 20-nm-thick Sb2S3 thin films.  

(a) Optical micrograph image of an MRR. An MRR is coupled with waveguide, and only a small portion of 

the MRR is viewable due to the metal pads.  (Scale bar: 20 µm) (b) Reversibly tuning an MRR for 5 cycles by 

electrically switching 20-nm-thick, 25-μm-long Sb2S3 thin film on the intrinsic waveguide. A resonance red 

shift Δ𝜆𝑟  of 0.40 nm and Q-factor reduction from 5.66×104 to 3.91×104 were observed after Sb2S3 

crystallization. The shaded region indicates the standard deviation. Pulse conditions: 6.9 V, 500 ns for 

amorphization and 2.9 V, 200 ms for crystallization. (c) Optical micrograph image of an MZI. (Scale bar: 50 

µm) (d) Reversibly tuning of an unbalanced MZI for 5 cycles by electrically switching 20-nm-thick, 60-μm-

long Sb2S3 thin film on the intrinsic waveguide. A spectral blue shift of 0.75 nm was observed. Pulse conditions: 

14.7 V, 500 ns for amorphization and 5.5 V, 200 ms for crystallization. (e) Push-and-pull operation of the dual-

arm reconfigurable MZI by independently controlling the upper and lower MZI arms, showcasing a larger 

phase shift and deterministic four-level operation. The structural phases of the Sb2S3 on the upper 𝜎𝑢 and lower 

arm 𝜎𝑙 are denoted in the label as 𝜎𝑢𝜎𝑙.  

 

Figure 4.4(a) shows an optical microscope image of a fabricated non-volatile tunable MRR loaded with 

25-μm-long, 20-nm-thick Sb2S3 thin film. The PIN diode had a resistance of ~ 62.5 Ω and a turn-on voltage 



 

 

of ~0.8 V (see Figure 4.12 in a later section). Figure 4.4(b) shows 5 reversible switching cycles with a 

resonance shift of ~ 0.40 nm by applying three electrical pulses with amplitude of 6.9 V (2.9 V) and duration 

of 500 ns (200 ms) to switch the Sb2S3 into its a- (c-) phase. The free spectral range (FSR) of this MRR was 

measured as ~ 2.7 nm, suggesting a round-trip phase shift of ~0.3π or 0.012π/µm. This matches very well 

with the simulated phase shift ~0.0125π/µm in Figure 4.5(a), indicating a complete phase change of Sb2S3. 

By fitting the ring’s spectrum to a Lorentzian line shape, we extracted a Q-factor of 5.66×104 (3.91×104) 

for a- (c-) Sb2S3. Therefore, the excess loss of c-Sb2S3 was estimated as 0.11 dB according to the Q-factor 

reduction52, i.e., loss per π was around 0.4 dB, which is slightly higher than the simulated loss per 𝜋 of 

~0.26 dB in Figure 4.5(b), and can be attributed to extra scattering at c-Sb2S3 grain boundaries113. We note 

that the Q-factor reported here reduces compared to the initial Q-factor (~1.27×105) due to extra optical 

loss caused by carrier migration from the doping region to the intrinsic region after high voltage pulses 

were applied. This excess loss was estimated as ~ 0.2 dB and can be eliminated in the future by enlarging 

the intrinsic region width. 

 

Figure 4.5: Numerical simulation results for 𝝅  phase shift length 𝑳𝝅  and loss per 𝝅  when varying 

waveguide width and Sb2S3 thickness.  

(a) Simulation results for 𝐿𝜋, where different curves represent different Sb2S3 thickness. A larger waveguide 

width and Sb2S3 thickness lead to a smaller 𝐿𝜋. For our Sb2S3-based phase shifters, 400 nm waveguide width 

and 20 nm Sb2S3 thickness are used, leading to 𝐿𝜋 ≈ 80𝜇𝑚 and unit length phase shift of ~0.125𝜋/𝜇𝑚. (b) 

Simulation results for the loss per 𝜋. A larger waveguide width and PCM thickness lead to a larger loss. The 

loss per 𝜋 is ~0.26 𝑑𝐵 and unit length loss is ~0.00325𝑑𝐵/𝜇𝑚 for our design. 



 

 

 

We emphasize that the smaller resonance shift compared to recently reported results113 is due to 

differences in the waveguide geometry as the waveguides here are thicker (300 nm compared to 220 nm) 

and narrower (400 nm compared to 500 nm), which significantly reduces the optical mode interaction with 

Sb2S3 (Figure 4.6). To further increase the optical phase shift, we can design thinner and wider waveguides 

to enhance the light-matter interaction or use low-loss PCMs with a larger optical refractive index contrast 

such as Sb2Se3
106. Another approach is to increase the PCM thickness, which, however, incurs additional 

optical scattering106 and difficulty in reversible switching168.  

 

 

Figure 4.6: Numerical simulation results for 𝝅 phase shift length 𝑳𝝅 when varying Sb2S3 thickness and 

waveguide height.  

Different curves represents the simulation results for waveguide height of 200, 300, and 400 nm. An increase 

of Sb2S3 thickness or decrease of the waveguide height lead to a decrease of 𝐿𝜋. The impact of waveguide 

height is significant as it determines the confinement of the optical mode, hence the interaction strength 

between light and Sb2S3. 

  

Similar to the MRR, the low-loss phase shifter functionality was also demonstrated in an unbalanced 

MZI in Figure 4.8(c), where 60-μm-long, 20-nm-thick Sb2S3 was deposited on both arms. Figure 4.8(d) 



 

 

presents the reversible switching result of single arm for 5 cycles. Applying the amorphization (14.7 V, 500 

ns) and crystallization (5.5 V, 200 ms) electrical pulses led to a resonance shift of ~ 0.75 nm for an FSR of 

~ 6.3 nm, i.e., a phase shift of ~0.24π. The electrical voltage is higher than MRRs due to impedance 

mismatch with the function generator during measurement. Moreover, the longer metal wire also led to a 

lower unit length conductivity in this MZI (Figure 4.7). We also note the MZI presents a smaller unit length 

phase shift (0.004π/µm) than the MRR (0.012π/µm). We attribute it to the relatively thick oxide residue on 

the MZI waveguides, which physically separated the optical modes from the Sb2S3 thin film and weakened 

the effective index contrast. This issue could be resolved by another wet-etch process to ensure complete 

removal of the oxide on waveguides.  

 

Figure 4.7: Measured I-V curve of the micro-ring resonator and Mach-Zehnder interferometer.  

The current is normalized with respect to the PIN junction length. The unit length conductivity for the MRR 

and the MZI is 0.629 mS/μm and 0.306 mS/μm, respectively. The MRR shows a higher unit length conductivity, 

which leads to the lower phase-transition voltages. This is attributed to the longer metal wire in the MZI and 

could be improved in the future fabrication by making the metal wire wider. 

 

Dual-arm controlled MZIs support push-and-pull type operation, as shown in Figure 4.4(e). With input 

light injected from the upper input port, switching the Sb2S3 to the c-phase on the upper (lower) arm led to 



 

 

a blue (red) spectral shift at the bar output port. We denote the structural phases 𝜎 of Sb2S3 on the upper 

and lower arms as a sequence 𝜎𝑢𝜎𝑙. Figure 4.4(e-iii) shows a large relative phase shift of ~0.7π between ca 

and ac phase and four distinct levels. We attribute the difference in spectral shift on two arms to incomplete 

phase transition or material ablation, which can be improved by further optimizing the pulse conditions. 

Previous simulation suggested an ablation-free voltage range of 0.6 ~ 1.0 V, which highly depends on the 

PIN heater design164. For our PIN design and experiment, we typically observed Sb2S3 ablation when the 

voltage was increased by 1 ~ 1.2 V (0.5 ~ 0.7 V) above the minimum amorphization (crystallization) 

voltage. The tighter crystallization parameter window is attributed to the much longer crystallization pulse 

duration (200 ms). We note that both the incomplete phase transition and the material ablation could be a 

result of non-uniform heating produced by sub-optimal heater design in Figure 4.13. If the center 

temperature is significantly higher than the surrounding temperature, the middle portion of PCMs may be 

amorphized (ablated) while materials at the edge remain unswitched168. In the future, a curved heater 

design169 can be adopted to improve the heating uniformity hence a complete phase transition without 

material ablation. This MZI device maintained good optical contrast after more than 500 switching events 

in Figure 4.8.   

 

Figure 4.8: Endurance test of an unbalanced MZI.  

We sent in amorphization and crystallization pulses for 500 switching events and the contrast still remains. 

There is a DC shift in the transmission level, which could be attributed to partial material damage and could 

be resolved later by optimizing the electrical pulse condition.  

 



 

 

 

4.3 Asymmetric directional couplers with multiple Sb2S3 segments 

On this scalable platform, we show a novel method to achieve deterministic multilevel or quasi-

continuous tuning, which is generally a challenging task due to the inherent stochastic nature of electrically 

controlled PCMs160. 

4.3.1 Improved design procedure 

The design of the Sb2S3-loaded asymmetric DCs follows the procedure described in the previous 

chapter113,144, but here we further improve the design flow and allow a straightforward design without 

calculating the loss manually as in Equations (2.3) and (2.4). Two waveguides of different widths are 

closely positioned to facilitate evanescent coupling, which allows optical power to transfer from one 

waveguide to the other. According to the coupled mode theory33, the power transfer efficiency 𝜂  is 

expressed by 𝜂 =
1

√1+(δ/κc) 
2 

, where δ =
𝜋

𝜆0
(𝑛2 − 𝑛1) is the waveguide detuning, 𝜅𝑐 =

𝜋

𝜆0
(𝑛𝑒 − 𝑛𝑜) is the 

coupling strength, 𝑛1,2 is the effective refractive index for two waveguides assuming no coupling, 𝑛𝑒,𝑜  is 

the effective index of the even (odd) super-modes formed by the coupled waveguide system, 𝜆0 is the 

vacuum wavelength of light. The maximum power transfer efficiency 𝜂𝑚𝑎𝑥 = 1 is reached when 𝑛1 = 𝑛2, 

which is referred to as the phase-matching condition. In such phase-matched system, the coupling length 

𝐿𝑐  to achieve complete power transfer is expressed as 𝐿𝑐 =
𝜋

2𝜅𝑐
=

𝜆0

2(𝑛𝑒−𝑛𝑜)
, where 𝑛𝑒,𝑜  can be obtained 

from numerical simulators. In general, a larger gap between two waveguides leads to a smaller coupling 

strength 𝜅𝑐 and hence a longer 𝐿𝑐. To reconfigure the directional coupler, Sb2S3 is deposited and patterned 

on top of the narrower waveguide. The widths of the waveguides are carefully designed to achieve the 

phase matching condition in one state of Sb2S3, in our case, the a-state. When switching Sb2S3 to the c-state, 

both the coupling strength 𝜅𝑐 and the waveguide detuning 𝛿 are changed due to effective index difference 

of the PCM loaded waveguide, thus a change in the output power. By judicious design of the gap, a complete 

switching from cross-state to bar-state can be achieved144. As such, selectively switching part of the Sb2S3 



 

 

can provide intermediate output states113,114,170. It is important to highlight that while such intermediate 

levels naturally emerges with low-loss PCM Sb2S3, achieving them using absorptive PCMs like GST 

requires additional device design1.  

We describe the steps to design the asymmetric directional coupler based on Sb2S3 in detail here. 

1. Fix the width 𝑤1 of the bare silicon waveguide WG1 and the height of Sb2S3 ℎ𝑆𝑏𝑆 to a reasonable 

number. We used 𝑤1 = 500 nm and ℎ𝑆𝑏𝑆 = 20 𝑛𝑚 in this design. 

2. Design the width 𝑤2 of the Sb2S3-loaded waveguide WG2 to phase match with WG1. We should 

separately simulate WG1 and WG2. We can set the Sb2S3 to a-phase (c-phase also works113), sweep 

𝑤2, and calculate WG2’s effective index difference Δ𝑛 from WG1. The phase matching condition 

is obtained when Δ𝑛 = 0, see Figure 4.9(b). 

3. Design the gap between WG1 and WG2. First, we need to calculate the effective index 𝑛𝑒,𝑜 of the 

super-mode of the coupled waveguide system by a Lumerical FDE simulation. The coupling length 

𝐿𝑐  can be calculated as 𝐿𝑐 =
𝜆0

2(𝑛𝑒−𝑛𝑜)
, where 𝜆0  is the vacuum wavelength, see Figure 4.9(c). 

Second, we can use the optical mode in isolated WG1 as the input field and propagate through the 

coupled waveguide system with a length of 𝐿𝑐 using the propagate command in Lumerical. A high 

cross transmission should always present for a-Sb2S3, see Figure 4.9(d-i). The key requirement for 

gap optimization is to ensure high bar-state transmission for c-Sb2S3, see Figure 4.9(d-ii). We 

should pick the smallest gap to make the device more compact. 

4. All the key parameters are obtained. The last step is to verify the design by running an FDTD 

simulation, see Figure 4.10. 



 

 

 

Figure 4.9: Schematic of the Sb2S3-based asymmetric directional coupler and main results during 

simulation.  

(a) Schematic. Left: top view. Right: cross-sectional view. All the design dimensions are included. (b) 

Difference between the effective index of waveguide WG1 and WG2 when they are fully decoupled, i.e., Δ𝑛 =

𝑛2 − 𝑛1 . The phase matching condition is achieved at 𝑤2 = 473.2 nm , where Δ𝑛 = 0 . (c) Numerical 

simulated coupling length with respect to gap. (d) The calculated optical transmission at cross and bar port for 

both phases of Sb2S3. A consistent cross state for a-Sb2S3 verifies the coupling length is calculated correctly. 

The smallest gap that fulfills this requirement is ~310 nm, which renders the smallest 𝐿𝑐 . The design, 

highlighted by a red star, shows a high extinction ratio between cross and bar output port of ~25 dB (35 dB) 

for a-Sb2S3 (c-Sb2S3).  

 

 

Figure 4.10: Simulated electric field distribution and transmission spectra for quasi-continuously tunable 

beam splitters with 2 Sb2S3 segments.  

(a)-(d) The field distribution and (e)-(h) the transmission spectra. The optical characteristics are programmable 

depending on the phases of two Sb2S3 segments indicated in the titles of the sub-figures. We note here the 



 

 

optical splitting ratios are different from the experimental results and it can be attributed to fabrication 

imperfections such as variations of Sb2S3 deposition thickness, waveguide width and waveguide gap.  

 

Figure 4.11(a) shows our scheme to achieve deterministic multilevel operation, where multiple 

segments of Sb2S3 thin films are individually controlled using interleaved PIN doped silicon heaters. Each 

Sb2S3 segment is switched electrically and independently to fully-a- or fully-c-phase in a repeatable fashion. 

By encoding the state of each Sb2S3 segment, multiple deterministic operation levels were obtained. To 

avoid potential thermal crosstalk, the segments are separated by 1 μm. One potential pitfall of this scheme 

is the relatively complex structure, incurring a higher device failure rate due to fabricated dimension 

inconsistencies. In that regard, mature wafer-scale silicon photonic fab/foundries provide a reliable solution.  

As a proof-of-concept demonstration, we designed a two-segment asymmetric DC, which functions at 

bar(cross)-state when both Sb2S3 segments are in c(a)-phase, and other configurations function as 

intermediate levels as shown in Figure 4.10(e)-(h). Two segments were designed deliberately with different 

lengths to achieve four operation levels. Such multilevel asymmetric DC was fabricated, and Figure 4.11(b) 

shows the microscope image, where the longer, 70-μm-long Sb2S3 segment (labeled 2) is twice as long as 

the shorter, 35-μm-long one (labeled 1). The orientation of the PIN diode heaters was designed to have 

opposite polarity to reduce unwanted crosstalk due to potential doping region mixing. The measured 

transmission spectrum is shown in Figure 4.11(c), where the insets indicate the structural phase of each 

Sb2S3 segment. As shown in Figure 4.11(c-i), initially two Sb2S3 segments were in the cc-state, and the 

asymmetric DC guided the light to the bar-port with an insertion loss of ~ 0.7 dB and extinction ratio of ~ 

12.1 dB at 1,343 nm. When switching Sb2S3 Segment 2 with short electrical pulses with amplitude of 7.8 

V and duration of 500 ns, a complete spectrum flip was observed as in Figure 4.11(c-ii), exhibiting high 

cross-port transmission with an insertion loss of ~1.0 dB and an extinction ratio of ~ 8.6 dB at 1,343 nm. 

We further amorphized Segment 1 using 7-Volt, 500-ns electrical pulses and measured spectra in Figure 

4.11(c-iii) and (c-iv). An intermediate operation level with a splitting ratio of ~32:68 between the bar- and 

cross-port was realized in Figure 4.11(c-iii). The extinction ratio increased from 12.1 dB in the cc-



 

 

configuration to 15.7 dB in the ac-configuration. Segments 1 and 2 were switched back to c-phase with 3-

Volt, 200-ms electrical pulses. We repeated the switching experiment 5 times and plot the standard 

deviation as the shaded region in Figure 4.11(c). The only slight standard deviation showcases the 

deterministic nature of this multi-segment approach. The discrepancy of the experimental performance from 

the simulation can be attributed to the fabrication imperfection and can be overcome by finer device 

parameter sweep during tape-out and better process control. 

Figure 4.11(d) shows the cross-port transmission at 1,330 nm for 52,600 switching events by 

alternatively sending in amorphization and crystallization pulses to Segment 2. We note that one switching 

event is accomplished by applying three pulses, crucial for repeatable Sb2S3 phase transition113. To facilitate 

the cyclability test, we used three faster pulses for crystallization with amplitude 3.3 V and duration 2 ms 

and the amorphization pulses remained the same. The thermal stabilization time between pulses was set to 

~40 ms, allowing us a rate of around  
1

3×(40+2+0.5)𝑚𝑠
≈ 8 𝐻𝑧 and to finish all 52,600 switching events 

within ~ 2 hours. The optical contrast remained high ~ 8.5 ± 1.0  dB after 5,000 switching events, 

showcasing excellent cyclability. After that, we observed a gradual drift in the c-Sb2S3 level and a decrease 

in the contrast, which could be attributed to partial Sb2S3 damage. The latter may also be a result of PIN 

heater degradation indicated by the IV characteristics change after 52,600 events (Figure 4.12). As such, 

we had to increase the amorphization (crystallization) pulse amplitude to 8.3 V (3.3 V) at event 17,000. An 

optical misalignment occurred at ~ 17,000 switching events, indicated by a low optical transmission level 

and is highlighted in Figure 4.11(d) by a gray box region. After we realigned the optical setup and increased 

the pulse amplitude, the optical contrast recovered to ~ 4.5 ± 0.7 dB until ~ event 30,000 and gradually 

decreased again. The Sb2S3 stopped responding to any electrical pulses after 50,000 switching events. 

Scanning electron microscope images in Figure 4.13 show some black areas at the edge of waveguides after 

the cyclability test, which could be a result of Sb2S3 ablation or thermal reflowing. Several strategies can 

be used to further improve the cyclability, such as developing more endurable PCMs171, engineering the 

thickness and material of the encapsulation layer172, patterning the PCMs into subwavelength 



 

 

nanostructures92, and engineering the microheaters and pulse conditions to provide a uniform temperature 

distribution169. 

 

Figure 4.11: Quasi-continuously tunable asymmetric directional coupler with two individually controlled 

Sb2S3 segments.  

(a) Schematic with metal pads omitted for visualization, (b) Optical micrograph of an QC-DC with two Sb2S3 

segments. (Scale bar: 20 µm) (c) Transmission measurement results at both bar and cross port for different 

phase sequences of the material, demonstrating 4 operation levels: (i) cc (ii) ca (iii) aa (iv) ac, where we have 

denoted the phases of segments 1 and 2 as 𝜎1𝜎2. The transmission spectra were normalized to a reference 

waveguide on chip. The device was reconfigured with electrical pulses five times and the shaded region 

indicates the standard deviation, showing excellent deterministic quasi-continuous tuning beyond binary levels. 

The switching condition was three 7.8V, 500ns pulses for amorphization and three 3V, 200ms pulses for 

crystallization. (d) Endurance test recording cross-port optical transmission while switching Sb2S3 Segment 2 

at 1,330 nm for 52,600 times (26,300 cycles). The switching condition was three 7.8V, 500ns pulses for 

amorphization and three 3.3V, 2ms pulses for crystallization. Little performance degradation was observed 



 

 

before 5,000 switching events, exhibiting a large optical contrast of 8.5 ± 1.0 dB. The contrast decreases after 

then and an optical misalignment happens at around event 17,000, highlighted by the gray box labeled “optical 

misalignment”. After realigning the setup, a contrast of 4.5 ± 0.7 dB was preserved. The Sb2S3 became 

inactive after around 50,000 switching events as indicated by the gray box labelled “Sb2S3 inactive”. 

 

 

Figure 4.12: The I-V curve before and after the PIN diode heater was switched for 26,300 cycles.  

Although a small decrease in the current is observed, the PIN diode heater still functions well. 

 



 

 

 

Figure 4.13: The scanning electron microscope images after switching 50,000 switching events.  

(a) The SEM image for the device after 50,000 switching events. (b) The zoom-in image of a. Black areas after 

switching many cycles, indicate Sb2S3 ablation and thermal reflow, pointed by white arrows. (c) The SEM 

image for another device which was not switched. (d) The zoom-in image of (c). Scalebar: 10 µm for (a, c) 

and 200 nm for (b, d). 

 

4.4 Quasi-continuously tunable MZIs with equal and unequal-length multi-

segments schemes 

The idea was further extended to more Sb2S3 segments for a larger number of operation levels in 

unbalanced MZIs under single-arm operation in Figure 4.14. We designed, fabricated, and characterized 

two types of devices, with 𝑁 Sb2S3 segments having the same or different lengths. Since the phase shift Δ𝜑 

induced by each Sb2S3 segment is proportional to its length, the equal length scheme loses some encoding 

capability by the redundant Sb2S3 configurations, leading to (𝑁 + 1) level. In the unequal-length scheme, 

we design the lengths of segments as a geometric sequence with a common ratio of 2. For example, we 

denote the length of Sb2S3 segment 𝑖  as 𝐿𝑖 , then 𝐿1: 𝐿2: 𝐿3 = 1: 2: 4  for a three-segment device. 



 

 

Redundancies in this configuration were avoided, and the system implemented at most 8 (2𝑁 , 𝑁 = 3) 

distinct levels with equal channel spacing.  

 



 

 

Figure 4.14: Quasi-continuously tunable Mach-Zehnder interferometers with four equal and unequal 

Sb2S3 segments.  

(a) Schematic, (b) Optical micrograph of the quasi-continuously tunable MZIs with equal and unequal Sb2S3 

segments. (Scale bar: 100 µm) (c, d) Zoomed-in optical micrograph of (c) equal (d) unequal Sb2S3 segment 

lengths. (Scale bar: 10 µm) (e) Measured optical transmission spectra after gradual amorphization of equal-

length Sb2S3 segments. Five operation levels were achieved with spectral shift of 1.125 ± 0.26 nm per step. 

Reversible switching was demonstrated by bringing the system level back through one-by-one crystallization 

in a different order. The structural phase of Sb2S3 segments is indicated near each curve with the same 

convention in Figure 4.11. (f) Experimental demonstration of 8 (2𝑁 , 𝑁 = 3) operation levels enabled by 

unequal segment lengths in a geometric series with common ratio of 2. The spectrum is continuously shifted 

with uniform step resonance shift of 0.41 ± 0.12 nm or 0.059π. We highlight that measurement for both (e) and 

(f) were repeated five times and the shaded regions indicate the standard deviation. The barely visible shaded 

region indicates a highly deterministic quasi-continuous operation. (Note: metal pads were omitted in the 

schematic for visualization.) 

 

Figure 4.14(b) shows the optical microscope image of the fabricated quasi-continuously tunable MZIs 

with four equal (see zoomed-in picture in Figure 4.14(c)) and unequal (see zoomed-in picture in Figure 

4.14(d)) Sb2S3 segments. In both cases, the total length of four Sb2S3 segments was ~ 80 µm to provide a 

large phase shift. Figure 4.14(e) shows the measured results for the equal length scheme, where we achieved 

5 (𝑁 + 1,𝑁 = 4) operation levels and demonstrated reversible switching between different levels. We first 

tested the pulse conditions of each segment with measured impedance of ~106, 130, 156, 185 Ω. This 

variation could be attributed to the difference in the metal wires’ length and could be resolved by 

engineering the geometry of these wires. However, we note the switching voltage levels were similar after 

matching the load impedance of the function generator, showing less than ±0.5 V variation for 

amorphization. We used three ~8.8 V, 500 ns (~3.6 V, 200 ms) pulses for amorphization (crystallization). 

The waterfall plot of Figure 4.14(e) shows the measured bar-port spectrum labeled with Sb2S3 segments’ 

phases when light was injected from the upper input port and tuning the lower MZI arm. All Sb2S3 segments 

were first set to the c-state, then amorphized one-by-one from Segment 1 to Segment 4. A gradual blue shift 

is observed, agreed with the smaller refractive index of a-Sb2S3 compared to c-Sb2S3. The Sb2S3 segments 

were then crystallized one-by-one in a different order until all were switched to the c-phase. Figure 4.14(e) 



 

 

shows that the final transmission spectrum (the top curve) aligning well with the initial one (the bottom 

curve), indicating a reversible switching behavior. The spectral shift was estimated as 1.125 ± 0.26 nm per 

step. The slight non-uniformity of the spectral shift could be attributed to incomplete phase change in the 

Sb2S3 segments and non-uniformity of Sb2S3 films, which could be resolved by further optimizing the 

electrical pulse condition and the fabrication process. We emphasize that no thermal crosstalk was 

observed, showcasing the thermal-crosstalk-free advantage of PCM tuning.  

Compared to the equal length scheme, which only achieves 𝑁 + 1 distinct optical levels, the unequal 

length scheme can achieve a maximum of 2𝑁 levels with 𝑁 Sb2S3 segments due to the strict non-redundant 

configurations. Figure 4.14(f) shows our measurement results with three unequal Sb2S3 segments with a 

total length of 40 µm. The longest Segment 4 was not used in this proof-of-concept experiment. Like the 

previous experiment for the equal length scheme, we first set all the Sb2S3 segments to the a-phase, and 

then programmed the binary phases of three Sb2S3 segments to demonstrate 3 bits or 8 (2𝑁, 𝑁 = 3) distinct 

levels by different combination of the segments’ phases, labeled in Figure 4.14(f). We report a uniform step 

resonance shift of ~0.41 ± 0.12 nm and a step phase shift of ~0.059π. We note that we observed a red shift 

when amorphizing the upper arm in another MZI and achieved a larger spectral shift after pulse condition 

optimization. Such an exponential increase in the number of levels is critical to reducing the complexity of 

the control circuits for large-scale integration. We emphasize that the measurements for both equal-length 

and unequal-length devices were conducted for five reversible switching cycles. The standard variation was 

shown by a barely visible shaded region in Figure 4.14(e) and (f), implying a highly deterministic quasi-

continuous tuning behavior.  

4.5 Discussion 

We note that this idea of using multiple segments of PCMs has been shown recently in another work 

with a different PCM GSSe and tungsten heaters170. We emphasize that it was limited to pure-amplitude 

modulation in a 1 × 1 waveguide switch, but we show phase-only modulation with the low-loss PCM Sb2S3 

in 2 × 2 asymmetric directional couplers and MZIs, which are crucial building blocks for large-scale PICs. 



 

 

We note although we only demonstrated integrating Sb2S3, any other PCMs can be easily tested on this 

versatile platform by simply changing the sputtering targets or using different deposition methods such as 

evaporation. Such a platform can open new opportunities in fast-prototyping and examination of PCMs. 

Lastly, we compare our scalable platform with other existing PCM-PIC works in Table 6, showing similar 

performance. The energy efficiency for switching can be further improved by reducing the distance from 

the waveguides to metal vias.  

In summary, we demonstrated a scalable programmable PIC platform by combining the mature and 

reliable 300-mm silicon photonic fab with in-house backend-of-line integration of the low-loss PCM Sb2S3. 

Non-volatile electrically programmable MRRs, MZIs, and asymmetric directional couplers were shown 

with low loss, large phase shift, and high endurance. Moreover, thanks to the platform’s capability of 

handling complex structures, we showed a novel scheme to achieve deterministic multi-level operation by 

independently controlling multiple Sb2S3 segments. We experimentally demonstrated such deterministic 

quasi-continuous tuning behavior in both asymmetric directional couplers and MZIs, showcasing at most 

𝑁 bits or 2𝑁 optical levels with 𝑁 Sb2S3 segments by careful geometry engineering. Overall, our work lays 

the foundation for very-large-scale programmable PICs with zero-static power consumption and 

deterministic multi-level operations. 

The table below compares our PCM-based asymmetric directional coupler with other PCM-based 

devices. We would like to point out that the higher energy consumption of our directional coupler can be 

attributed to the relatively large distance between the metal via and the intrinsic region, which leads to 

voltage drop in the doping region and hence large power waste. It could be improved in future tape-outs by 

reducing this distance. The relatively larger footprint of our devices is caused by the thick silicon waveguide 

used and Sb2S3’s smaller refractive index change (~0.6) compared to other PCMs such as GST (~3) and 

Sb2Se3 (~0.8). We can further improve this by utilizing Sb2Se3, trying to switch to a thicker PCM film, as 

well as using wider and thinner waveguides. We emphasize although PCMs are inherently immune to 

thermal crosstalk, the highest temperature different PCMs can tolerate highly depends on the material 

property. Here we quantify thermal immunity of each PCM by the glass transition temperature, which is 



 

 

the lowest temperature the device’s optical characteristics may be changed. Lastly, scalability is measured 

by whether the photonic structures were fabricated by e-beam lithography (low scalability) or 

photolithography (high scalability).  

Table 5: PCM integrated photonic device performance comparison 

Ref Year PCM Structure 
IL 

(dB) 

Energy per 

switch (nJ) 

PCM 

length 

(µm) 

Number 

of cycles 

Crosstalk 

immune 

temperature 

(°C) 

Scalability 

173 2015 GST 
MRR  

1 × 1 
N.R. 

0.533 

(3.1) c 
1.0 50  150 Low 

174 2017 GST 
MRR 

1 × 1 

5.1 / 

4.3 a 

0.19 

(17.1) c 
60.0 1,000  150 Low 

162 2018 GST 
WG 

1 × 1 
N.R. 0.68 2 10 150 Low 

52 2018 GST 
MRR 

1 × 1 
0.54 

0.62 

(0.2) c 
2 > 2 150 Low 

92 2019 GST 
MRR 
1 × 2 

0.75 / 
0.46 a 

0.25 
(11) c 

25.0 > 2 150 Low 

175 2019 GST 
WG 

1 × 1 
1.5 0.7 3.6 N.R. 150 Low 

53 2019 GSST 
MRR 

1 × 1 
< 0.5 N.R. 5 b > 3 200 Low 

176 2021 GST 
MZI  

2 × 2 
0.5 

14 

(9.5 × 105) c 
43.0 600 150 Low 

158,159 2021 GST 
WG 

1 × 1 
0.9 0.38 10.0 1,000 140 Low 

177 2021 GST Crossing 0.4 0.8 2 N.R. 150 Low 

178 2023 Sb2Se3 
SF 

3 × 3 
N.R. N.R. > 500 N.R. 200 High 

179 2017 GST 
WG 

1 × 1 
4.8 

20 

(7.2 × 106) c 
1 ~10  150 Low 

112 2019 GST 
MMI 

1 × 1 
1 

10 

(9) c 
1 > 1500 150 Low 

111 2020 GST 
MRR f 

1 × 1 
0.06 

8 

(78) c 
3 > 500 150 Low 

106 2021 Sb2Se3 
MZI f 

2 × 2 
> 0.3 

176 

(3.8 × 103) c 
> 100 b > 125  200 Low 

180 2022 GST 
MMI 

1 × 1 
3.7 

4.5 

(4.08) c 
1 N.R. 150 Low 

59 2022 GST 
DC f 
2 × 2 

2.0 
380 

(6.8 × 103) c 
50 > 2,800  150 Low 

107 2022  Sb2Se3 
MRR f 

1 × 1 
0.33 

9.25 

(1.4 × 103) c 
6 > 1,100 200 Low 

170 2022 GSSe 
WG 

1 × 1 
0.12 400 80 5 × 105 325 Low 

114 2023 GST 
MZI f 

2 × 2 
1  

81 

(1.9× 103) c 
10 > 3 150 Low 

113 2023 Sb2S3 DC f < 1 

56 

(3.4 × 106 

1.7× 103) c, e 

79 > 800 275 Low 

181 2023 Sb2Se3 DC f < 0.36 N.R. 33 > 1000 200 Low 



 

 

This 

work 
2023 Sb2S3 DC f ~1 

1216.8 

(7.2 × 107) c 105 > 2500 275 High 

(ER: extinction ratio, IL: insertion loss, BW: bandwidth, MMI: Multi-mode interferometer, MZI: Mach-Zehnder 

Interferometer, MR(D)R: Micro-ring(disk) resonator, WG: waveguide, DC: directional coupler, N.R.: not reported) 

Note: here, only devices with reversible switching are compared. aFor drop and through ports, respectively. bEstimated 

from figures. cEnergy per switching event for amorphization (crystallization). dThe study did not perform multiple runs to 

verify the multilevel reliability. eFor ring resonators. A smaller crystallization energy was achieved with shorter 100-µs 

crystallization pulses. fThis work contains multiple devices; only one device is picked for easy comparison. 

 

  



 

 

Chapter 5. Nonvolatile transmissive spatial light modulator with PCMs182 

Different from the previous Chapters mainly focusing on PICs, we switch our interest to metasurface 

applications in this Chapter. Free-space modulation of light is a key enabling technology behind optical 

communications, holography, ranging, and virtual/ augmented reality. Traditional spatial light modulators 

(SLMs) based on liquid crystal (LC) on silicon183 or micro-electro-mechanical systems (MEMS)184 employ 

large pixels (~10𝜇𝑚 × 10𝜇𝑚),  resulting in bulky devices and generally require large driving voltage. To 

address these limitations, recent years have seen tremendous effort to realize free-space light control based 

on subwavelength diffractive optical metasurfaces integrated with active materials.116–118,120–123,185–187 

Metasurfaces can support resonances that enable substantial phase or amplitude modulation with small 

pixel-size. Since a smaller active volume is modulated, lower energy and faster modulation speed can be 

achieved. For example, metasurface resonators (Q~550) based on organic electro-optic (EO) polymers have 

enabled GHz modulation speed;119 LCs combined with a Huygen’s metasurface significantly reduced the 

pixel size (~1 µm) and LC thickness (~1.5 µm) required to attain a full 2π phase shift range;116 metasurfaces 

based on plasmonic resonances coupled to epsilon-near-zero materials have enabled a full 2π modulation 

with independent control of amplitude and phase;117 a large phase-modulation of ~1.3π was reported by 

tuning a plasmonic metasurface using graphene.123 Nevertheless, these approaches are all based on volatile 

changes such as the Pockels effect or free carrier dispersion, necessitating a constant power supply to hold 

the static state. To ensure time-multiplexed pixel control, most of these volatile SLMs require an active 

memory matrix with transistors to hold the written state. In addition, the most prevalent LC-based phase-

only SLMs suffer from detrimental pixel crosstalk188 and temporal phase fluctuations or jitter.124 To 

overcome some of these limitations, a promising solution is to modulate light using non-volatile materials, 

which can drastically improve energy efficiency and avoid phase flickering. Moreover, the control 

complexity can be significantly reduced due to the built-in memory of the non-volatile reconfiguration. 

Chalcogenide-based PCMs are ideal candidates to realize such functionality,126 thanks to their non-volatile 

microstructural phase transition,49 large contrast in complex refractive index (typically Δn ≥ 1),50 and 



 

 

CMOS compatibility.47 Since PCMs hold their state once configured, a truly ‘set-and-forget’ switching 

element can be realized. In fact, PCMs have already attracted considerable attention to create tunable 

metasurfaces for applications such as varifocal lensing,122,127 beam steering,128,129 intensity 

switching,120,121,130,131 and spectral filtering.132 Despite the progress, non-volatile optical phase-only 

modulation in transmission - a highly desirable feature for SLMs - remains elusive. This is because previous 

works on nonvolatile tunable metasurfaces have used lossy PCMs such as GST120,131 or GSST121 which 

concomitantly induce large absorption upon phase transition, prohibiting optical phase-only modulation. 

The use of metallic heaters further leads to ohmic losses and makes transmissive operation difficult. In fact, 

to date, all electrically tunable PCM meta-optics have been demonstrated in reflection mode. Although non-

volatile phase-only control has been demonstrated in mid infrared where GST is transparent,133 the PCM 

was switched optically, which prevents a fully integrated and scalable platform. A similar work has shown 

near 2π phase shift via optical switching of low-loss PCM Sb2S3 in the visible spectrum,189 but phase 

modulation is accompanied by a large amplitude change due to non-negligible loss of crystalline Sb2S3. 

By leveraging a thin layer of low-loss PCM Sb2Se3,190 we demonstrate a non-volatile, electrically 

programmable metasurface for phase-only modulation of free-space NIR light in transmission. The ultra-

low loss of Sb2Se3 enables decoupling of the phase modulation from amplitude modulation in NIR. A phase-

only modulation of ~0.25π in simulation and ~0.2π in experiment is achieved by coupling the PCM to a 

high-Q diatomic metasurface (Q ~ 409). The tunable metasurface also demonstrates large endurance with 

over ~1,000 transitions. We further exploited a guided-mode resonance, albeit with a lower Q-factor, to 

facilitate a larger electric field overlap with the Sb2Se3 layer, which enables a full 2π phase shift. Beyond 

global control of the entire metasurface, we further show independent electrical control of 17 pixels, verified 

by optical microscope images and the spectral response, which shows discrete and reversible resonance 

levels as function of the respective configuration. A deterministic multi-level resonance tuning of the 

metasurface is achieved by switching the meta-molecules one-by-one with a total resonance shift of ~8 nm 

at a center wavelength of ~1230 nm. By imparting different phase profiles through independently 

controlling individual pixels, we demonstrate dynamic beam focusing with three distinct focal distances. 



 

 

This work constitutes a crucial step towards a truly non-volatile “set-and-forget” transmissive phase-only 

SLM. 

5.1 High-Q silicon diatomic metasurfaces 

Sb2Se3 undergoes a refractive index contrast of ~0.7 and exhibits zero loss190 near 1550nm upon phase 

transition, which stipulates a ~1.1µm PCM thickness to obtain π phase shift. Although this is significantly 

thinner than LC used in commercial SLMs, it still poses a severe challenge in reversible switching the 

PCMs as PCMs need to be melt-quenched to be amorphized and such a large thickness prevents the critical 

cooling rate to be reached191. Motivated by the use of micro-ring resonators to increase the modulation in 

integrated photonics192, a high-Q planar resonator can also be used to enhance the phase-modulation of free-

space light by thin film PCMs, while allowing thinner device layer compared to Fabry-Perot cavities. Earlier 

works on high-Q planar resonator focused on high contrast gratings (HCGs)193, while the more recent works 

studied guided mode resonances194,195 and quasi-bound-state-in-continuum (q-BIC) in periodic 

nanostructures with in-plane asymmetry196,197. Here we combine the idea of HCGs and q-BIC by 

introducing an asymmetry into the periodicity of traditional HCGs, realizing a diatomic metasurface (Figure 

5.1a). Such diatomic gratings have also been proposed theoretically in several previous works198–200 but 

have not yet been demonstrated experimentally. Figure 5.1(a) shows that the symmetry of the HCG is 

broken by slightly displacing one of the gratings relative to the other so that the grating spacing becomes 

dissimilar i.e., 𝑎1 ≠ 𝑎2. We denote such displacement or perturbation by 𝛿 =
𝑎2−𝑎1

Λ
 where Λ is the period 

of the diatomic gratings. The duty cycle Γ is defined as 
𝑤

Λ𝐻𝐶𝐺
 for HCGs and 

2𝑤

Λ
 for diatomic gratings, where 

w is the grating width. It can be seen that Γ is essentially identical for the two types of gratings because the 

period of diatomic gratings becomes doubled. The periodicity doubling causes the folding of the first 

Brillouin zone such that the dark modes that were originally at the edge of the Brillouin zone, which is 

outside the free space light cones, get folded into the interior of the light cones, allowing for free-space 

excitation198,199. Figure 5.1(b) shows the simulated spectral response of HCGs (δ=0) and diatomic gratings 



 

 

(δ=0.05) upon excitation of normal incident transverse-magnetic (TM) polarized light near 1550nm 

wavelength (Λ=900nm and Γ=0.7). When δ=0, the dark modes are not accessible to the free-space 

excitation and hence no resonance is observed. While a small perturbation of δ=0.05 is enough to introduce 

a sharp q-BIC resonance with its signature Fano line shape. The pronounced resonance is also clearly visible 

where the q-BIC resonance represents a ~40 times electric field enhancement compared to when δ=0. By 

fitting the experimentally measured resonance we extract a large Q factor of 409 for heavily N-doped 

220nm silicon-on-insulator (SOI) at doping concentration~1020cm-3, as shown in inset of Figure 5.1(b). 

Such a high experimental Q factor is attributed to the ultra-smooth etching side wall as revealed by Figure 

5.1(c), despite the additional loss from the doping. Figure 5.1(d) shows the experimentally measured 

resonances as the period increases from 900nm to 1000nm (silicon thickness=220nm, δ=0.05 and Γ=0.7). 

The resonance shifts almost linearly across the telecommunication wavelength range without significant 

change in the Q factor – a highly-desirable characteristic of diatomic gratings198. In contrast, changing the 

period of HCGs will lead to dramatic change in the Q factor as reported previously193. Now it becomes clear 

why diatomic grating makes an ideal platform for demonstrating nonvolatile phase-only modulation with 

low-loss PCMs. First, the high Q resonance is very sensitive to loss. Cladding the metasurface with lossy 

materials will lead to a drastic reduction in the Q factor and hence obscuring the phase response, which is 

shown in the simulation where 20nm GST is cladded on the diatomic gratings. Secondly, 1D gratings with 

equal width w allow identical microheater resistance, which is essential for uniform heating of PCMs 

cladded on top. Using dissimilar gratings width199 or introducing notches194 in the gratings will also lead to 

high Q resonances, but each grating will have different resistance and distinct Joule heating from current 

injection. Lastly, the resonance can be fine-tuned to arbitrary wavelength by changing the period without 

degrading the Q factor, enabling operation over a large wavelength range limited only by the material 

bandgap. Figure 5.2(a) and (b) show the electric field distribution near the diatomic grating when 𝛿 = 0 

and 𝛿 = 0.05. It is observed clearly that the a resonance mode localized in the air emerges when 𝛿 ≠ 0.  



 

 

 

Figure 5.1: High-Q silicon diatomic metasurfaces.  

(a) Schematics of the high contrast gratings (top) and diatomic gratings (bottom). ΛHCG and Λ are the periods 

of the high contrast gratings and diatomic gratings respectively. The spacing between the gratings are denoted 

as a, a1, and a2. w is the grating width. (b) Simulated transmission spectrum of the high contrast gratings (blue) 

and diatomic gratings (orange) under normal incident TM polarized light. Silicon thickness is 220nm, 

Λ=900nm and Γ=0.7. Inset: Experimental spectrum of a resonance fitted by Fano equation, Λ=950nm and 

Γ=0.7. (c) SEM of fabricated silicon diatomic gratings. (d) Experimental spectrum of diatomic gratings with 

different periods from 900nm to 1000nm, where δ=0.05 and Γ=0.7. 

 

 



 

 

 
Figure 5.2: Electrical field amplitude of the diatomic resonance.  

(a) The electric field profile when δ=0, corresponding to the high contrast gratings. (b) The field profile of 

diatomic metasurfaces on resonance. The unit for the scale bar is micrometer. Λ=900nm and Γ=0.7. 

 

5.2 A nonvolatile electrically reconfigurable metasurface for phase-only 

control 

In order to dynamically control the diatomic metasurface, we heavily dope the SOI into 

microheaters201,202 that is used to switch the Sb2Se3 cladded on top in Figure 5.3(a). The Phosphorus doped 

(doping concentration~1020cm-3) SOI is then etched into diatomic gratings before uniformly cladded with 

20nm Sb2Se3, followed by 40nm atomic-layer-deposited (ALD) Al2O3 encapsulation to prevent oxidation 

and PCM re-flowing during switching. Ohmic contacts are then formed by Ti/Au electrodes. The cross-

section of the metasurface is shown in Figure 5.3(c). Current is injected into the highly doped silicon 

gratings via electrical pulses that causes joule heating, which in turn switches the PCMs. A two-objective 

transmission setup is built to probe the reversible switching of Sb2Se3-loaded diatomic metasurface. The 

fabricated chip is wire bonded to a customized printed circuit board (PCB), labelled PCB1 in Figure 5.3(b), 

connected to a second customized PCB carrying a microcontroller, labelled PCB2 in Figure 5.3(b). PCB2 

is then connected to an arbitrary function generator (AFG) that generates pulses. The microcontroller on 

PCB2 can be programmed by a computer to individually address metasurface pixels fabricated on the same 



 

 

chip. Figure 5.3(d) shows the optical micrograph of a 30-µm aperture metasurface pixel on a chip that has 

been wire bonded to the PCB.  

Below is the detailed fabrication process of this device. The reconfigurable diatomic metasurface is 

fabricated on a 220-nm-thick silicon layer on top of a 3-μm-thick buried oxide layer (SOITECH) with back-

side-polished silicon. The blanket SOI wafer is first implanted by phosphorus ions with a dosage of 2 × 

1015 cm-2 and ion energy of 40 keV at a tilt angle of 7°. Subsequently, the wafer is annealed at 950 °C for 

30 min to activate the dopants. The metasurface pattern is defined by a JEOLJBX-6300FS 100kV electron-

beam lithography (EBL) system using positive tone ZEP-520A resist. 220 nm fully etched gratings are 

made by an inductively coupled plasma reactive ion etching (ICP-RIE) process in Florine-based gases. 

Before metallization, the surface native oxide was removed by immersing the chips in 10:1 buffered oxide 

etchant (BOE) for 15 seconds to ensure Ohmic contact. A second EBL exposure using positive tone 

poly(methyl methacrylate) (PMMA) resist is subsequently carried out to create windows for the Ti/Au 

deposition. After development, 5nm Ti followed by 150 nm Au was electron beam evaporated onto the 

chip. The lift-off of Ti/Au was completed again by immersing the chip in methylene chloride. Note that for 

the individual control of meta-molecules we replaced Au with Pt for electrodes to avoid the melting of the 

traces at high voltages. The third EBL step is used to expose the PMMA resist before depositing Sb2Se3 via 

magnetron sputtering. The Sb2Se3 is sputtered using a magnetron sputtering system at 30 W RF power 

under a deposition pressure of 4 mTorr and Ar flow of 30 sccm. The deposition rate for Sb2Se3 is ~1 nm/min. 

Additionally, the samples are capped with 10 nm of SiO2 sputtered in situ (150 W RF power, 4 mTorr 

pressure, and Ar flow of 30 sccm), to prevent oxidation during sample shipping. The atomic ratio of Sb2Se3 

after deposition is confirmed using XPS to be Sb:Se ≈ 44:56 which is close to the sputtering target 

stoichiometry of Sb:Se ≈40:60. Immediately after lifting off the PCM in methylene chloride, a 40nm ALD 

Al2O3 is grown on the chip to protect the PCM from oxidation and reflowing during switching. To allow 

good adhesion between the wedge bonds and the metal pads, the 5th EBL step is used to open windows in 

PMMA resist at the wire bonding regions for Al2O3 etching. The Al2O3 on top of the contacts is etched 

away using ICP-RIE etching in Chlorine-based gases. Then the PCMs are initialized into the fully 



 

 

crystalline state by rapid thermal annealing (RTA) at 200°C for 10 min under N2 atmosphere before 

measurements. Finally, the chip is wire bonded onto the custom-made PCBs using a wire bonder 

(Westbond) via gold ball-wedge bonds. 

 

Figure 5.3: A nonvolatile electrically reconfigurable metasurface based on Sb2Se3.  

(a) Schematic of the transmissive tunable diatomic metasurface based on Sb2Se3. S (G), signal (ground) 

electrode. (b) The optical setup for probing the metasurface in transmission along with the devices under test 

wire bonded to a customized PCB. AFG is the arbitrary function generator. (c) The schematic of the tunable 

metasurface cross section. (d) Optical micrograph of a single metasurface pixel on the chip under test.  

 



 

 

 

Figure 5.4: Optical measurement setup for in situ electrical control of metasurface.  

(a) Transmissive spectral measurement setup. (b) Optical phase measurement based on a Mach-Zehnder 

Interferometer. (c) Motorized stage setup for far-field beam profile imaging. 

 

The experimental setup we built for reversible switching experiment is shown in Figure 5.4(a) and 

described as following. Each metasurface pixel on the chip is connected to a metal pad that is wire bonded 

to a pin on the carrier PCB or PCB1. A hole (0.8cm in diameter) is opened at the center of PCB1 to allow 

direct light transmission through PCB. To individually control the pixels, a second customized PCB (PCB2) 

carrying a microcontroller (Arduino Nano) is inserted into the predefined pins on PCB1. Most commercial 

microcontrollers normally supply low voltage (≤5 V) and slow speed (tens of megahertz), whereas high 

amplitude (>10 V) and short (nanosecond falling edge) pulses are required to amorphized the PCMs. Hence 

the microcontroller cannot be directly used to switch PCMs, instead an external function generator is 

connected to the PCB2 as a source of excitation. The microcontroller can be programmed to turn on/off 17 



 

 

MOSFETs connected in series with the on chip microheaters. By switching the MOSFET, the voltage drop 

across the microheater can be controlled independently when an electrical pulse is applied. The resistance 

of the microheaters is measured using a source meter (Keithley 2450). A 30 𝜇𝑚 × 30 𝜇𝑚 large metasurface 

pixel typically has resistance of ~115 Ω at 1 mV DC bias, with ~ 45 Ω due to the relatively long Pt wires. 

The resistance of a single channel is ~1.45 kΩ. The SET and RESET pulses were generated from an 

arbitrary function generator (Keysight 81150A). To reconfigure the metasurface, we used a voltage pulse 

of 15 V (~10 mA/channel), 1.25 µs pulse width, and 8 ns rising/trailing edge to induce the amorphization. 

For crystallization, multiple voltage pulses of 6 V (~4.1 mA/channel), 50 µs pulse width, and 30 µs trailing 

edge is used. Our current device size is limited by the maximum voltage (20 V at 50 Ω load impedance) of 

our function generator. A larger aperture size will require function generator that can source higher voltage 

to amorphize the PCMs.  

Reversible tuning of the diatomic resonance is shown in Figure 5.5(a) (simulation) and 4.3(b) 

(experiment). The spectra of three consecutive switching cycles are plotted in Figure 5.5(b) and the shaded 

regions indicate standard deviations between the cycles. The small standard deviation reveals excellent 

cycle to cycle reproducibility. The experimentally extracted spectral shift (Δλ~1.2𝑛𝑚) matches very well 

with the simulated shift (Δλ~1.3𝑛𝑚). Reversible tuning of the diatomic resonance is shown in Figure 5.5(a) 

for simulation and Figure 5.5(b) for experiment, respectively. The spectra of three consecutive switching 

cycles are plotted in Figure 5.5(b) and the shaded regions indicate standard deviations between the cycles. 

The small standard deviation reveals excellent cycle to cycle reproducibility. The experimentally extracted 

spectral shift (Δλ~1.2 𝑛𝑚) matches very well with the simulated shift (Δλ~1.3 𝑛𝑚). Figure 5.5(c) shows 

the phase shift (Δϕ) and transmission contrast (Δ𝑇%) between the two optical states near the 1518 nm 

resonance. A good agreement between the simulation and the experiment can be clearly observed. We 

extract a maximum phase shift of ~0.2π (~0.25π) near the resonance wavelength via digital holography 

experiment (simulation) with less than 10% measured change in transmission. To measure the phase shift 

caused by the switching, we built a free-space MZI in Figure 5.4(b) and switched the pixel in-situ. The 



 

 

interference fringes between the signal beam through the metasurface and the reference beam are taken by 

the IR camera. The images are captured at 11 different wavelengths from 1513 nm to 1523 nm for the two 

optical states, averaged over 20 frames. The phase of each optical state is calculated by first applying a high 

pass filter on the image in the Fourier domain, and then take the argument of the filtered image in the real 

space. The phase shift is the difference between the extracted phase in crystalline state and amorphous state. 

We note that, this phase shift is ~10 times larger than what can be achieved by switching a 20-nm-thick 

blanket film of Sb2Se3 without a metasurface. The extracted Q factor only decreases from 312 to 271 upon 

crystallization, representing only 13% reduction, which shows that minimal loss is introduced by the phase 

transition. In contrast, the high Q resonance is completely suppressed by the same thickness of GST cladded 

on the metasurface due to its high optical loss in Figure 5.6(d). This is done by fitting the spectrum to a 

Fano resonance via 𝑇 =∣ 𝑎 + 𝑗𝑏 +
𝑐

𝐸−𝐸0+𝑗γ
∣2, where a, b, and c are constant real numbers. E is the photon 

energy and E0 is the central resonance energy. 2γ is the line width of the resonance. So, the Q factor is 

calculated to be 
𝐸0

2γ
. 

Finally, we show that the tunable metasurface is robust against switching over 1,000 times without 

degradation in contrast, as shown in Figure 5.5(d). The tunable laser is parked near the resonance to detect 

a measurable change in the transmission. Optical micrographs taken before and after the cyclability test 

indicate no ablation is caused by the switching across the entire 30 × 30µm2 metasurface. Apart from 

switching Sb2Se3, we also show that the doped silicon microheater is a versatile platform for switching 

other PCMs, for example we have also reversibly switched GST using doped silicon on sapphire 

microheaters to realize a broadband tunable notch filter in transmission (Figure 5.6).  



 

 

 

Figure 5.5: Reversible switching of the diatomic metasurface and nonvolatile phase-only modulation.  

(a) Simulated spectral and phase shift caused by the phase transition of 20-nm-thick Sb2Se3. Δλ  is the 

wavelength shift of the resonance dip. The diatomic metasurface is designed to have Λ=900 nm, Γ=0.7, and 

δ=0.05. a(c)Sb2Se3: amorphous(crystalline) Sb2Se3. (b) Measured reversible switching of the diatomic 

resonance. The switching conditions are 3.6 V, 50 µs pulse width, 30 µs trailing edge for SET and 11.6 V, 1 

µs pulse width, 8 ns trailing edge for RESET. Three consecutive cycles are plotted where the shaded area 

indicates the standard deviation between the cycles and the solid line indicates the average. The spectrum is 

normalized with that of bare SiO2 on silicon. (c) Phase shift (Δϕ, red) and transmission contrast (Δ𝑇%, blue) 

between two optical states extracted from simulation (top) and experiment (bottom). The phase is measured at 

11 different wavelengths with 1nm spacing and averaged over three switching cycles. The standard deviation 

over the cycles is shown by the error bars. (d) Cyclability of the tunable metasurface for 1,000 switching events. 

The switching conditions are 3.6 V, 50 µs pulse width, 30 µs trailing edge for SET and 11.6 V, 600 ns pulse 

width, 8 ns trailing edge for RESET. Each pulse is temporally separated for 2 seconds to ensure long thermal 

relaxation. The data is filtered by a two-point moving average to reduce fluctuation caused by thermal and 

mechanical noises. 

. 

 



 

 

 
Figure 5.6: Reversible switching of GST blanket film using doped Si heater. 

(a) A silicon-on-sapphire chip wire bonded to a customized PCB. (b) A switchable GST pixel on the chip. (c) 

Schematic of the device structure. The Si is 500nm thick and the GST is 20nm thick. (d) Reversible switching 

of the transmission spectrum of a 10µm large GST pixel measured by a FTIR. The spectrum is averaged across 

five cycles. 2.5V, 50µs wide, 50µs trailing edge pulses are used for crystallization, 7V, 200ns wide, 8ns trailing 

edge pulses are used for amorphization. 

 

5.3 Individually addressing meta-gratings 

In this section, we demonstrate the manipulation of light in spatial domain by individually programming 

the meta-gratings. To demonstrate the spatial light modulation, we found a TE-polarized guided mode 

resonance with much stronger electric field interaction with Sb2Se3 in Figure 5.7(b), which is utilized in the 

following experiment to achieve a larger resonance shift (~10 nm measured in experiment) and phase shift 

(~ 2π near the resonance extracted from the simulation), albeit at the expense of a lower Q-factor (~100) 

than the q-BIC designs. We note that such phase shift could in principle be measured with the same digital 

holography setup in Figure 5.4(b), but we do not have a laser source with fine enough linewidth near this 

new resonance wavelength in Figure 5.8. The linewidth of the laser source can influence the phase 



 

 

measurement result significantly by smearing out the interference effects at different wavelengths. Figure 

5.4(c) and (d) show that near the resonance, a wavelength dependent phase shift is presented, and at certain 

wavelengths the amplitude contrast is low. Figure 5.4(e) and (f) show the experimentally measured 

spectrum for a(c) Sb2Se3, as well as the transmission contrast in %, showing a similar phase-only behavior 

near 1225 – 1228 nm.  

 

Figure 5.7: Transverse-electric polarized guided-mode resonance mode with strong field overlap with 

Sb2Se3, achieving larger resonance shift and a full 2π phase modulation.  



 

 

(a) The transmission spectra for light with normal incident (green) and with a 3° (blue) and 8° (orange) angle. 

The resonance is only found with a non-zero incident angle. It is shown that the resonance is not clearly visible 

below 3° and a further increase of the angle leads to a decrease in Q-factor. Therefore, we chose 3° because it 

supports a resonance mode with a relatively high Q-factor. The following results are simulated with an incident 

angle of 3°. (b) Electric field distribution for amorphous-Sb2Se3 at resonance wavelengths 1228 nm. The white 

contour represents the metasurface structure. The field on Sb2Se3 is much stronger compared to the quasi-BIC 

mode in Figure 5.2. Using crystalline-Sb2Se3 gives a similar field profile at 1238 nm due to the relatively thin 

Sb2Se3. (c) Calculated transmission spectra at amorphous (denoted "amor", blue dashed line) and crystalline 

(denoted "crys", orange dashed line) state. The transmission change is plotted in black. A large spectral shift 

of ~10nm is observed. The transmission change is highly wavelength dependent and is close to 0 % (25%) 

when 𝜋 (2π) phase shift is achieved. (d) Wavelength-dependent phase difference (denoted "diff", black line) 

between the amorphous (blue dashed line) and crystalline (orange dashed line) state with a maximum phase 

modulation of full 2π. One can achieve an arbitrary phase shift (0~2π) near the resonances (near 1228 or 1238 

nm, denoted by the gray boxes) by precisely controlling the laser wavelength. (e) Measured optical 

transmission spectrum when all Sb2Se3 was set to amorphous (blue line) and crystalline (orange line) phases. 

The data is smoothened by a 72-data moving average filter and the raw data is indicated by light colors. We 

note that the raw data is only normalized for the transmission to a reference background (a box area with silicon 

etched away). The absolute transmission is around 86% (83%) at minimum for amorphous (crystalline) phase. 

This transmission is much higher compared to the simulation, which can be attributed to finite metasurface 

area and the 0-th order transmission. (f) Relative transmission change Δ𝑇 when changing the entire metasurface 

from amorphous to crystalline phase. A maximum change of ~ 15% (2.5%) is observed near the crystalline 

(amorphous) phase resonance, indicating a very small amplitude change. 

 

 

Figure 5.8: Spectrum of the Fianium broadband source combined with a grating.  

 



 

 

To ensure the uniformity of resistance across the channels, we derive an analytical model for the metal 

wires from the metasurface to the metal pads, depicted in Figure 5.9(a). The model was then used to design 

the metal wire geometry to obtain uniform resistance, which is essential for the simultaneous control of 

multiple channels. 

 

 

Figure 5.9: Metal wire engineering to achieve equal resistance across 17 channels.  

(a) Schematic for a single metal wire with parameters denoted. 17 channels have different 𝑤𝑑, hence geometry 

engineering is necessary to achieve uniform resistance. For the first model, only the light gray region is used. 

In the second model, the dark gray region is added to avoid narrow width at the turn, enhancing the wire's 

robustness against high current. (b) Calculated resistance using model 1 (blue) and model 2 (orange), showing 

matching results of a large resistance variation of around 25Ω across the channels if Ln is constant. (c) 

Optimized design for uniform resistances by varying 𝐿𝑛 across the channels (blue). The orange line with circle 

(star) markers is the resistance calculated from the analytical model (COMSOL simulation). (d) Experimentally 

measured resistance across 17 channels.  

 



 

 

We denote the resistance of the rightmost straight wire as 𝑅1 and that of the tilted part from points 

(𝑥1, 𝑦1) to (𝑥2, 𝑦2) as 𝑅2. The pad resistance is usually small and is ignored. Therefore, the total resistance 

is calculated as 𝑅 = 𝑅1 + 𝑅2.  

In our first model, where the dark gray part is not considered, 𝑅1 is given by: 

 
𝑅1 =

𝜌𝑙𝑛
ℎ𝑤𝑛

 (5.1) 

Where 𝜌 = 10.8 × 10−8Ω/𝑚 is the resistivity of Pt, ℎ = 250 𝑛𝑚 is the metal wire thickness, 𝑙𝑛 and 

𝑤𝑛 are denoted in the schematic. The tilted wire can be approximated by a trapezoid, the resistance of which 

could be obtained by segmenting the wire into small rectangular pieces and integrating them over the entire 

length. Although this method ignores the boundary effects of electric currents, it works very well for a 

slowly varying shape. Considering an isosceles trapezoid with a top width of 𝑤𝑡, bottom width of 𝑤𝑏 and 

height of 𝑙𝑡, the resistance is calculated as 𝑅 =
𝜌

ℎ𝑎0
ln (

𝑎0𝑙𝑡

𝑤𝑡
+ 1), where 𝑎0 = (𝑤𝑏 −𝑤𝑡)/𝑙𝑡. Here, 𝑤𝑡 and 

𝑤𝑏 could be approximated as: 

 
{
𝑤𝑡 = 𝑤𝑛 ⋅ cos(𝜃)

𝑤𝑏 = 𝑤𝑝 ⋅ cos(𝜃)
 (5.2) 

Where 𝑤𝑛 and 𝑤𝑝 are constants as denoted in the schematic, and 𝜃 is the tilting angle (0~π/2) of the 

wire, given by  

 𝜃 = |arctan (
𝑦2 − 𝑦1
𝑥2 − 𝑥1

)| 

 

(5.3) 

 Where |𝑎| means taking the absolute value of 𝑎. The length of the tilted wire can be obtained by 𝑙𝑡 =

√(𝑥1 − 𝑥2)
2 + (𝑦1 − 𝑦2)

2. Therefore, the total resistance is given by the following equations: 



 

 

 

{
 
 
 
 

 
 
 
 𝑅 =

𝜌𝑙𝑛
ℎ𝑤𝑛

+
𝜌

ℎ𝑎0
ln (

𝑎0𝑙𝑡
𝑤𝑡

+ 1)

𝑎0 = (𝑤𝑏 −𝑤𝑡)/𝑙𝑡
𝑤𝑡 = 𝑤𝑛 ⋅ cos(𝜃)

𝑤𝑏 = 𝑤𝑝 ⋅ cos(𝜃)

𝜃 = |arctan (
𝑦2 − 𝑦1
𝑥2 − 𝑥1

)|

𝑙𝑡 = √(𝑥1 − 𝑥2)
2 + (𝑦1 − 𝑦2)

2

 (5.4) 

In the experiment, we observed that the corners were the most prone to damage. This is attributed to a 

narrow wire at the corner, especially for edge channels with large tilting angles. We add extra metal wires 

drawn in dark gray to improve the robustness. The second model for this revised structure is given by: 

 

{
 
 
 
 

 
 
 
 𝑅 =

𝜌(𝑙𝑛 − 𝑙𝑒𝑥/2)

ℎ𝑤𝑛
+

𝜌

ℎ𝑎0
ln (

𝑎0𝑙𝑡
𝑤𝑡

+ 1)

𝑎0 = (𝑤𝑏 −𝑤𝑡)/𝑙𝑡
𝑤𝑡 = 𝑤𝑛 ⋅ 𝑐𝑜𝑠(𝜃) + 𝑙𝑒𝑥 ⋅ sin (𝜃)

𝑤𝑏 = 𝑤𝑝 ⋅ cos(𝜃)

𝜃 = |arctan (
𝑦2 − 𝑦1
𝑥2 − 𝑥1

)|

𝑙𝑡 = √(𝑥1 − 𝑥2)
2 + (𝑦1 − 𝑦2)

2

 (5.5) 

 As a sanity check, Eq. (5.5) reduces to Eq. (5.4) if 𝑙𝑒𝑥 = 0. The accuracy of our model is also verified 

by Figure 5.9(b), where two models match well. It can also be seen that there is a large resistance variation 

of ~25Ω across the 17 channels if we use the same Ln. By varying the Ln spatially, it is possible to obtain 

uniform resistance (variation ~ ±0.5Ω, < 1%) across the channels as shown in Figure 5.9(c). The analytic 

model also agrees with the COMSOL simulation with less than 2Ω difference. The optimized metal wire 

shape is shown in the inset of Figure 5.9(d). In experiments, the doped-silicon gratings contribute to a much 

higher resistance than the simulation with an average resistance of ~1460 Ω across 17 channels, see Figure 

5.9(d). However, the variation in the resistance is only around ±10 Ω (~1% of the resistance) which will 

not lead to significant variation in Joule heating.  

Besides optical resonance engineering, we also designed electrical signal routing and improved the 

uniformity of the electrical response. Each pixel of the metasurface must be individually addressed to fully 

control the optical wavefront. As shown in Figure 5.10(a)-(c), such controllability of individual pixel is 



 

 

realized using a source metal wire and 17 separate ground wires for each channel (two periods). The wires 

are electrically isolated with each other by over-etching silicon and wire-bonded to the PCB 1 in Figure 

5.3(b). Each ground wire is switched ON/OFF by PCB2 through a CMOS transistor. The source wire 

delivers the voltage pulses to all the ON channels with equal current.  

 

Figure 5.10: Electrically addressing individual meta-atoms.  

(a) Micrograph image of the metasurface with individual meta-atom controllability, where each channel is 

wire-bonded to a PCB. (b) Zoomed-in micrograph image that shows 17 electrical wire fan-outs. (c) Schematic 

plot. (d) IR images showing four different metasurface configurations. The brighter (dark) color shows an 

amorphous (crystalline) Sb2Se3 due to the slight absorption difference. (e) Simulated spectra corresponding to 

each configuration in d. (f) Measured spectra. The spectral shifts agree well with the simulation. (g) Waterfall 

plot for the normalized transmission spectra as switching meta-atoms one-by-one. The result is averaged over 

9 repeated experiments, and the barely visible shaded region indicates the standard deviation. The resonance 

dip shifts gradually from ~1235 nm to ~1228 nm with little variation. (h) Fitted Fano-resonance spectral shift 

with respect to the number of channels switched. The resonance wavelength gradually blue shifts as more 

gratings are switched. The minimal error bar indicates a near-deterministic multilevel operation.   

 



 

 

We implement arbitrary phase masks via the following two-step sequence: (1) first, all meta-molecules 

are electrically switched to the crystalline phase; (2) a single amorphization pulse is then applied to switch 

the selected meta-molecules. Amorphization, instead of crystallization, is used for independent control to 

avoid channel crosstalk. We observed large crosstalk if crystallization is used to switch the individual 

channel, as the longer crystallization pulses lead to a more severe thermal dispersion (Figure 5.11). This 

can be easily verified in COMSOL simulator Figure 5.12. Additionally, the lower temperature threshold 

for crystallization compared to amorphization (200 °C for crystallization vs. 620 °C for amorphization) can 

also contribute to crosstalk. Using this two-step switching method, four different configurations of the 

metasurface were obtained, shown in the IR camera images in Figure 5.10(e) and (f): (i) all amorphous 

metasurface; (ii), (iii) hybrid metasurface with an amorphous period two and three times the original 

channel pitch respectively, and (iv) all crystalline metasurface. Movie S1 shows the reversible switch 

between these configurations. The insets at the bottom left corner of each subplot show a schematic for the 

respective metasurface configurations for clarity. The bright (dark) lines correspond to amorphous- 

(crystalline-) Sb2Se3 from the slight absorption difference. The patterns are implemented by the same phase 

transition condition (15 V, 1.25 µs for amorphization and 6 V, 50 µs for crystallization), showcasing a 

universal method to set arbitrary patterns. Besides the IR camera images, we also measured the spectra of 

the configurations in, which qualitatively matches with simulations. 

The individually addressable metasurface was first globally switched to crystalline phase; then 

arbitrarily phase pattern was imparted by selecting the channels via Arduino and applying the same voltage 

pulse for amorphization. Starting from crystalline phase is preferred because unintentional switching of the 

neighbor pixels could be avoided by the more stringent amorphization transition condition. This is also 

verified by heat transfer simulation. Figure 5.10(d) shows four example images taken by an IR camera, 

where the brighter (darker) parts are amorphous- (crystalline-) Sb2Se3 due to the slight absorption 

difference. The patterns are implemented by the same phase transition condition (15 V, 1.25 µs for 

amorphization and 6 V, 50 µs for crystallization).  



 

 

 

Figure 5.11: Normalized transmission spectra showing evidence of thermal crosstalk-free performance 

during amorphization process.  

Normalized transmission spectra for hybrid gratings with an amorphous period two and three times the original 

channel pitch. (a, b) We start from an all-crystalline state and reversibly switch the selected channels via 

amorphization. Different colors show multiple switching cycles. After applying the amorphization pulse, we 

can return the spectrum to the initial state every time using a crystallization pulse, as indicated by the 

overlapping between the dashed and black lines. (c, d) We start from an all-amorphous state and reversibly 

switch the selected channels via crystallization. The crystallization pulse leads to unintentional crystallization 

of the adjacent meta-atoms due to thermal crosstalk. As a result, the subsequent amorphization pulses cannot 

return the systems to the initial all-amorphous state and the switching is irreversible. In contrast, the 

amorphization pulses we used in a, b will not switch the adjacent channels due to a much smaller thermal mass 

and higher temperature threshold. We note a global amorphization pulse could bring the metasurface back to 

the initial all-amorphous state, ruling out the possibility of device damage.   



 

 

 

Figure 5.12: Joule heating simulation in COMSOL for understanding of the thermal crosstalk-free 

performance during amorphization.  

(a) Simulated temperature distribution for amorphization using a 10.8V, 1µs pulse. (b) Simulated temperature 

distribution for crystallization using a 4.5V, 50µs pulse. The grating structure is overlaid on the plot. The 

electric current is only injected in the center diatomic grating (light orange), while the adjacent gratings (gray) 

are heated up due to thermal crosstalk. The temperature drops quickly away from the center. The amorphization 

temperature drop by ~280 K on the nearest gratings due to the shorter pulse width. Since the adjacent gratings 

do not reach the melting point, the thermal crosstalk will not lead to inadvertent switching of the nearby cells. 

In contrast, the temperature only drops by ~60K for the nearest gratings after the crystallization pulse, which 

is not enough to suppress accidental crystallization. 

 

Apart from collectively switching multiple channels, we further showed that the pixels can be 

independently addressed one by one. Increasing the number of amorphized channels causes a gradual 

spectral shift of the resonance from 1235 nm to 1228 nm, as shown in Figure 5.10(g). We fitted the 

resonance to a Fano line shape and extracted the resonance shift versus the number of meta-molecules 

switched in Figure 5.10(h). The small error bars averaging over nine repeated cycles suggest a highly 

deterministic multilevel operation. Compared to the partial amorphization of a single pixel via pulse 

amplitude or width modulation, individual tuning of the meta-molecules is much more reliable since each 

meta-molecule undergoes a full amorphization process. The resonance shift is not observed when more 

than ten meta-molecules were switched because the focused probe beam with a natural Gaussian shape has 

low intensity at the edge.  

 



 

 

 

Additionally, this decrease in the resonance shift from the center to the edge could be attributed to the 

non-local nature of the diatomic grating metasurface. Since the meta-atoms at the center have the most 

neighbors, the resonance effect is the strongest. The meta-atom at the edge has the least neighbor and 

coupling, hence weaker resonance shift. This behavior is verified through another rigorous coupled wave 

analysis (RCWA) simulation, where a super cell with 17 meta-molecules is simulated with only one 

crystalline meta-molecule while keeping all other meta-molecules amorphous. Figure 5.13 shows the 

resonance shift when placing the crystalline meta-molecule at different positions from 1 (edge) to 9 (center). 

Due to the symmetry, here we only show half of the total meta-molecules. Switching a center meta-molecule 

into the crystalline phase clearly causes a larger resonance red shift compared to an edge one. 

 

Figure 5.13: Simulated resonance shift of the metasurface when different meta-atoms are switched.  

(a) Simulated transmission spectrum. The resonance is first simulated (indexed 0 in the x-axis) when Sb2Se3 

on all the meta-atoms are set to amorphous phase. Then we separately simulated when only one meta-atom is 

switched to the crystalline phase from edge (indexed 1) to center (indexed 9). A larger resonance shift is 

observed when the meta-atom closer to the center is crystallized. (b) Plot of resonance shift versus the index 

of meta-atoms switched. A much larger shift (> 12 times) in the center (0.968 nm) is observed compared to the 

edge (0.078 nm). This simulation validates our observation in experiments that outer meta-atoms could not 

produce an observable resonance shift.  

 



 

 

The fitted Q factor (Figure 5.14) first decreases from ~120 to ~90 until 4 meta-molecules are switched 

and then increases to ~120. We attribute the lower Q-factor to inhomogeneous broadening, resulting from 

the coexistence of the amorphous and crystalline resonance mode. As more meta-molecules are switched, 

the amorphous resonance gradually dominates, leading to an increase in the Q-factor. 

 

Figure 5.14: Fitted Q-factor of the metasurface versus the number of switched meta-molecules.  

The Q-factor first decreases because of the inhomogeneous broadening of the amorphous-crystalline hybrid 

resonance mode and then increases as the amorphous resonance mode dominates. The Q reduction at the fully 

amorphous state near the end is not physical and could be attributed to a Fano resonance fitting error. 

 

Resonances are in general angle dependent since the effective index of systems changes as the angle 

changes. This could also be observed in our TM-polarized quasi-BIC mode and TE-polarized guided-mode 

resonance, as shown in the energy-momentum spectrum (E-k spectrum) in Figure 5.15(a) and (b) below. 

Due to the angle dependence, the input FOV is highly limited to < 0.5 degrees. Fortunately, it is possible 

to overcome this angle-dependence by flatband resonances, which have been demonstrated203,204 by 

breaking the vertical symmetry of the metasurface and therefore hybridizing multiple resonant modes. One 

example design for the TE-polarized guided mode resonance in Figure 5.15(c) and the simulated E-k 

spectrum is presented in Figure 5.15(d), where a resonance shift below 1 nm is observed across a half angle 

~ 10°. Such exotic resonance modes can create new opportunities for metasurface-based SLMs with a large 

input FOV.  



 

 

 

 

Figure 5.15: Simulated energy-momentum spectrum for both resonance modes and a flatband 

metasurface design for wide input FOV.  

(a) TE-polarized guided-mode resonance. The resonance red shifts and Q-factor reduce as the angle increases. 

(b) TM-polarized quasi-BIC mode. The resonance blue shifts as the angle increases. (c) Geometry of the 

designed flatband metasurface. Parameters are 𝑎 = 0.62, 𝜂 = 0.37, 𝛿 = 0.1. The thickness of the intrinsic 

silicon slab (doped silicon gratings) is 140 (380) nm. (d) A TE-polarized guided-mode resonance with an angle-

independent resonance within 10° half angle (resonance variation smaller than 1 nm).  

 

Lastly, as a proof-of-concept demonstration, we show that such individually addressable metasurface 

could enable tunable wavefront control. Switching the meta-atom induces a local, arbitrary optical phase 

shift near the resonance wavelength (0 ~ 2π, controlled by the laser wavelength). Using the angular 

spectrum propagation method and assuming a phase shift of π/2, we simulated the simple situations, where 



 

 

only the center 4, 8, or 12 meta-atoms were amorphized while all the other parts were crystalline. The 

simulation results in Figure 5.16(a) suggest weak focusing phenomena with different focal lengths (10, 30, 

75 µm). We experimentally imparted the phase profile by collectively switching center pixels and measure 

the 3D intensity distribution using a moving stage setup. Figure 5.16(b) shows the measurement results in 

the yz plane and the focal distance changes for different phase profiles, showing good agreement with the 

simulated results. The measured beam focusing is weak and non-ideal, primarily due to the relatively wide 

linewidth of our laser source (~ 1 nm) in Figure 5.8, which smeared out the optical phase contrast. This 

experiment could be improved by optimizing the resonance to a wavelength range that is supported by 

commercial tunable continuous-wave lasers, such as 1260 ~ 1360 nm or 1465 ~ 1565 nm. The focus 

efficiency could be improved by engineering phase profile and by using more unit cells.  

 

Figure 5.16: Tunable free-space beam focusing.  

(a) Angular spectrum propagation simulation for the situation where only the center 4, 8, or 12 gratings are 

amorphized and other gratings are kept crystalline. A phase shift of π/2 is assumed. Weak focusing could be 

observed and the focal lengths are clearly different (~10, 30, 75 µm) for three configurations. (b) Measured 

beam profile. The shift in focal lengths agrees well with the simulation. The plotted data is normalized to global 

crystalline configuration to avoid objective focusing and then averaged in the x-direction over the center 10 

pixels of the metasurface. 

 

Table 6 compares our Sb2Se3-based SLM with other works, including some commercial products based 

on liquid crystals (LCs) and digital mirror devices (DMDs). First, most of the commercial products work 

in the reflection mode instead of transmission, despite the need for transmissive SLMs in many applications 

such as AR/VR. A typical transmissive SLM is in the second row of Table 6, which has a large pixel size 



 

 

(36 µm compared to 1.8 µm in our work) and a large optical loss (transmittance ~ 28% compared to ~ 90% 

in our work). The limited pixel size further leads to a small output field of view (FOV) of ~ 1 degree88, 

constrained by the first-order diffraction. Thanks to the smaller pixel size, our output FOV is ~ 56 degrees. 

We note the resolution of our work could be easily extended to 0.9 µm with a PCB supporting more 

channels to support a large enough area metasurface. Our work also shows excellent power consumption 

compared to other reported values. We note the number reported in the table assumes a switching frequency 

of 1Hz for our device. In practice, once one programmed the PCM-based SLM, it can self-hold the state for 

much longer than 1 second with no energy consumption. For example, if the state is held for one hour, the 

power consumption per pixel becomes 0.06 (0.42) nJ/s for RESET (SET), showcasing orders of magnitude 

higher performance than the state-of-the-art commercial products. The pixel speed is around 10 kHz, mainly 

limited by the long SET process. This could be potentially resolved in the future by exploring faster PCMs. 

Table 6: Comparison of our Sb2Se3-based transmissive SLMs with other SLMs 



 

 

(LC: Liquid crystal; N.R.: not reported; DMD: digital mirror devices; EO: electro-optical; FC: free-

carrier dispersion effect; PCM: phase-change materials; *estimated or calculated values from figures or 

related parameters; † Assume a programming frequency of 1 second. This number corresponds to the 

switching energy for each phase transition. ) 

  

Ref 
Model 

name 

Tuning 

Method 

Transmissive 

or reflective 

Phase-

only 

Pixel 

size 

(µm) 

Optical 

loss (dB) 

* 

Pixel power 

consumption 

(nJ/s) 

Pixel Speed 

(MHz) 
Nonvolatile 

205 
TELCO-

033 
LC Reflective Yes 3.74 -1.4 N.R. 0.24 * No 

206 LC 2012 LC Transmissive Yes 36 -5.5 N.R. 0.05 * No 

207 
DLP650

0 
DMD Reflective No 7.56 ~0 ~ 3100 * 11.88 * No 

88 
Lab 

work 
LC Transmissive Yes ~1 -4.4 N.R. N.R. No 

208 
Lab 

work 

EO 

polymer 
Reflective No 390 -1 < 10000 * 50 No 

209 
Lab 

work 
LC Reflective Yes ~1 -3 N.R. 0.0005 No 

22 
Lab 

work 
FC Reflective Yes 5 -15.2 1.9 × 106 * 5.4 No 

210 
Lab 

work 
FC Reflective Yes 5 -0.2 675 * 135 No 

This 

work 

Lab 

work 
PCM Transmissive Yes 1.8 (0.9) -0.65 

200 

(RESET) 

1500 (SET)† 

0.01 Yes 



 

 

Chapter 6. Challenges and opportunities 

Despite impressive progress, several limitations and challenges exist in PCM-integrated photonics, 

including high optical loss of traditional PCMs, unreliable multi-level operation with electrical control, low 

endurance, and difficulty identifying accurate phase transition conditions. This section discusses these 

challenges and a few potential opportunities to mitigate them. 

6.1 High loss of traditional PCMs  

The high material loss of traditional PCMs, e.g., GST52,63,95, GeTe45, GSST53, and GSSe97, can introduce 

extra device insertion loss. Deposited on top of Si3N4 or silicon waveguides, PCMs provide strong 

reconfigurability. However, close to the optical mode, they also induce significant optical absorption, 

especially in crystalline phases. Moreover, non-uniform PCM deposition, and mode mismatch at PCM 

interfaces can cause additional scattering loss55. Physically separating the PCM further from the waveguide 

can reduce the loss at the price of switching contrast. Such optical loss limits applications to direct intensity 

modulation, although generally phase-only modulation is desired. For example, a well-known way to switch 

between different optical paths is to use an MZI21,39,211. Using GST in one MZI arm to modulate the phase, 

its excess loss in the crystalline state drastically attenuates the signal in that arm, making the interference 

imperfect and reducing the extinction ratio.   

To mitigate the high loss in the crystalline GST/GSST for low-loss applications, a three-waveguide 

device geometry (Figure 6.1) was conceptualized57–59. In this scheme, only the middle waveguide is 

deposited with PCMs, and is purposefully phase matched with other two bare silicon waveguides when the 

PCM is in the amorphous state. Therefore, light can couple to the cross output port for a-PCM. Once the 

PCM is switched to the crystalline phase, a large effective index mismatch significantly reduces the 

coupling, so the light directly transmits through the bar port and barely interacts with the lossy c-PCM, 

hence bypassing the high loss. However, we note that this geometry only works for PCMs transparent in 

the amorphous state and requires a significant refractive index contrast. A more promising solution will be 

wide bandgap PCMs, which are transparent in both phases and introduce negligible excess losses after 



 

 

deposited on waveguides (Figure 6.1). Specifically, two classes of PCMs, Sb2Se3 and Sb2S3
104,105, have very 

low material absorption loss (𝜅𝑎 ≈ 0 , 𝜅𝑐 < 0.005) in the telecommunication wavelength range (near 

1310nm and 1550nm) due to their wide bandgap. The extra scattering loss due to mode mismatch could be 

relaxed by patterning the PCM into a taper shape55.  

 

6.2 Reliable multi-level operation with electrical control 

It is difficult to achieve reliable multi-level operation with electrical control, due to PCMs’ stochastic 

nature160, where only two states (complete amorphous or complete crystalline phase) can be reliably 

produced. While using a fs-pulsed laser, reliable multi-level operations52,62,65,66,69 have been reported, 

electrical control of intermediate states remains unreliable59. Specifically, each time an amorphization event 

is triggered, the stochastic melt-quench process results in a slightly different amorphous state. It then leads 

to significant distinctions in the subsequent long-range nucleation and growth process, hence the PCM ends 

up with a stochastic partial crystalline phase160.  

Electrically controlled PCM waveguide switches with a reliable multi-level operation were 

demonstrated recently using segmented heaters for both phase-only55 and amplitude modulation97. Here we 

propose a similar scheme97 using multiple PIN silicon heaters, as shown in Fig 3a and 3c. Each heater 

separately controls a short PCM patch in a binary but reliable manner. The device then performs reliable 

multi-level operation by programming the states of all PCM patches individually.  

Furthermore, using this similar idea, beam splitters with quasi-continuously tunable splitting ratios, an 

essential component in many PICs, can be fabricated, as shown in Figure 6.1(b) and (d). Traditionally, such 

tunable beam splitters are realized in an MZI configuration, which may require active lengths exceeding 

100 µm21,39,143,211,212. The length of the phase shifter in an MZI is given by ~𝜆/2Δ𝑛𝑒𝑓𝑓, with 𝜆  being the 

wavelength and Δ𝑛𝑒𝑓𝑓being the effective index change. As Δ𝑛𝑒𝑓𝑓 is generally small using traditional 

modulation methods, the phase shifter is long. Additionally, two fixed beam splitters add considerable 

footprint (e.g., ~36μm in Ref. [211]), but do not contribute to the active tuning. In contrast, the PCM device 



 

 

footprint is significantly reduced to ~30 μm58, further improving the PIC’s integration density. As 

conceptualized in Figure 6.1, quasi-continuously programmable beam splitters can be realized by placing 

low-loss PCM directly on one of two waveguides to form an asymmetric directional coupler144, and 

controlling with segmented heaters. With lossy PCMs, however, additional design considerations are 

necessary to mitigate material loss, such as adapting a three-waveguide geometry in Figure 6.1(d), as 

explained earlier.  

In all these devices, the central design principle is to change the coupling between waveguides using 

PCM phase transition. The typical design procedure57–59,144,213 starts with picking a PCM thickness and 

performing optical mode simulations of the directional couplers using an electromagnetic field solver. One 

can then obtain the coupling length as 𝐿𝑐𝑜𝑢𝑝𝑙𝑒 = 𝜆/2(𝑛𝑒 − 𝑛𝑜), with 𝜆 being wavelength and 𝑛𝑒 and 𝑛𝑜 

being the effective index of the even and odd supermode of coupled waveguides, respectively. Then, by 

optimizing the waveguides gap and PCM thickness, a tunable beam splitter is designed144.  

 

Figure 6.1: Quasi-continuously programmable phase shifters, photonic switches, and beam splitters with 

PCMs and PIN diode-based doped silicon micro-heaters on a silicon-on-insulator (SOI) chip.  

(a, b) A phase shifter and a tunable beam splitter using transparent PCMs; (c, d) A photonic amplitude switch 

and a tunable beam splitter using PCMs absorptive in the crystalline phase but low loss in the amorphous phase. 

(a) The phase of light is changed due to the PCM refractive index difference between two states. (b) Upon 

PCM phase transition, the coupling strength between two waveguides is changed, and the splitting ratio 

changes. (c) Light is switched on and off, harnessing the PCM loss difference. (d) The middle waveguide is 

phase matched with the other two waveguides when the PCM is in the amorphous state. In the crystalline state, 



 

 

the light will not be coupled to the lossy c-PCM due to the large refractive index change, and the high crystalline 

material loss is effectively mitigated. 

 

6.3 Limited PCM endurance 

Another important unsolved issue is the limited PCM endurance214, i.e., how many times the PCM can 

be cycled through. Although recent years have witnessed orders of magnitude improvement in optical PCM 

devices’ cyclability (10 cycles96 to half 106 cycles97 with different PCMs and heater designs, see Table 7), 

it is still lower than what has been achieved in electronic memories (1012 cycles215, and potentially 1015 

cycles216 for GST). In electronic memories, PCMs are typically shaped into tiny units of volume 

(10×10×10nm3)217, whereas the size is much larger (10μm×1μm×10nm)54,59,95 for photonics. Generally, a 

smaller PCM volume suffers less degradation from thermal reflowing issues and is less prone to elemental 

segregation, which causes material failure92. Additionally, oxidization can occur if the material is not 

properly encapsulated, as phase transition involves high temperature. Assuming a linear scaling with 

volume, we expect to achieve around 1010 cycles in photonics. We note that even 1010 cycles are insufficient 

for high-speed modulation. For example, a GHz modulator with an endurance of 1010 would only last 10 

seconds!! While this is a severe limitation, 1010 cycles are sufficiently high for reconfigurable applications. 

Assuming a PIC/meta-optic is changed, kept for 1 second, and changed again for another application, it 

takes more than 300 years to complete 1010 cycles! Besides, the kept time could be much longer for many 

applications, e.g., optical computing or optical programmable gate arrays (OPGAs). Therefore, temporal 

response or the switching times of PCM-based programmable units are of little significance163.  

Different capping materials214 have been used to protect the PCM from thermal reflowing and 

oxidation, including ITO62, SiO2
218, Si3N4

218, ZnS/SiO2
64, Al2O3

59,67,95
, and others218. In particular, recent 

developments show that atomic-layer deposited Al2O3 can provide excellent protection59,67,95, primarily due 

to its highly conformal nature. Further optimizing the material property may also help improve device 

endurance. For example, to avoid element segregation, single-element PCMs219 may be explored. Finally, 



 

 

since smaller volumes of PCM are more durable92, patterning the PCM on a subwavelength scale67,92 could 

improve device endurance. We refer readers to a recent review214 on this topic. 

 

 

Table 7: PCM integrated photonic device performance comparison 

Ref Year PCM Structure 
ER 

(dB) 

IL 

(dB) 

Energy per 

switch (nJ) 

Footprint 

(µm) 

Optical BW 

(nm) 

Tuning 

method, 

cycles 

62 2015 GST 
MRR  

1 × 1 
10 N.R. 

0.533 

(3.1) c 
1 < 1 

Optical, 

50 cycles 

174 2017 GST 
MRR 

1 × 1 
> 5.0 

5.1 / 

4.3 a 

0.19 

(17.1) c 
60 < 1 

Optical, 

1,000 cycles 

92 2019 GST 
MRR 

1 × 2 
14 

0.75 / 

0.46 a 

0.25 

(11) c 
25 < 1 

Optical, 

N. R. 

93 2021 GST 
MZI  

2 × 2 
8.0 0.5 

14 

(9.5 × 105) c 
43 N. R. 

Optical, 

600 

96 2017 GST 
WG 

1 × 1 
1.2 4.8 

20 

(7.2 × 106) c 
1 >100 

Electrical, 

~10 cycles 

112 2019 GST 
MMI 

1 × 1 
6.5 7.5 N. R. 1 >100 

Electrical, 

>1500 

95 2020 GST 
WG  

1 × 1 
5 1.6 

13 

(715) c 
5 >100 

Electrical, 

>500 

55 2021 Sb2Se3 
MZI 

2 × 2 

6.5/ 

15.0 a 
> 0.3 

176 

(3.8 × 103) c 
> 100 b > 15 b 

Electrical, 

>125 cycles 

59 2022 GST 
DC 

2 × 2 
10.0 2.0 

380 

(6.8 × 103) c 
50 > 30 

Electrical, 

>2,800 

cycles 

97 2022 GSSe 
WG 

1 × 1 
0.12 12 N. R. 80 N. R. 

Electrical, 

5 × 105 

(ER: extinction ratio, IL: insertion loss, BW: bandwidth, MMI: Multi-mode interferometer, MZI: 

Mach-Zehnder Interferometer, MR(D)R: Micro-ring(disk) resonator, WG: waveguide, DC: directional 

coupler, N.R.: not reported) 



 

 

Note: here, only devices with reversible switching are compared. Footprint refers to the total device 

length. aFor drop and through ports, respectively. bEstimated from figures. cEnergy per switching event for 

amorphization (crystallization) 

 

6.4 Difficulty identifying phase change conditions  

Identifying the phase transition condition is non-trivial because the amorphization condition can be 

close to the device damaging threshold. We consider the silicon PIN heater and GST as an example in the 

following. In comparison to silicon, GST has lower melting temperature Tm (Tm,Si = 1450°C, Tm,GST = 

650°C), thermal conductivity kth (kth,Si = 148W/m∙K, kth,GST <1 W/m∙K) and specific heat capacity Cp
220 

(Cp,Si = 0.72J/g∙K, Cp,GST = 0.2J/g∙K). A smaller thermal conductivity means that heat diffuses away slower 

and hence becomes more localized within the GST. A smaller specific heat capacity indicates GST requires 

less energy to heat up. Therefore, the low melting point Tm, thermal conductivity kth, and heat capacity Cp 

all together imply that GST is much easier to melt than silicon with the same input energy. However, in our 

experiments, the silicon waveguide could be damaged when applying a slightly higher voltage (~0.3V) than 

the amorphization voltage. Simulations164,221 suggest that such surprising device damage stems from local 

hotspots at the edge of the waveguide, due to current crowding, as the silicon thickness abruptly changes 

between the waveguide and partially etched slab. Another challenge is that the phase change condition is 

not universal. As the phase transition is thermally induced, any change in the substrate or the cladding 

requires re-optimizing the heating conditions. 

Although identifying the accurate phase change conditions requires trials and errors, the conditions are 

reliable, once found. Therefore, the conditions must be identified on several separate testing devices before 

being applied on large-scale PICs. We can also further optimize the material systems to avoid device 

damages. Besides designing new PCMs with lower Tm, exploring other material systems for the photonic 

backbone with higher Tm than silicon, such as Si3N4, can facilitate a larger amorphization window. Another 

important research direction is to optimize the heaters. For doped silicon heaters, the geometry of the doping 



 

 

area and the silicon waveguides can be optimized to eliminate the local hotspots. Furthermore, new heaters, 

such as graphene heaters107,221, should be explored to maximize heat delivery to the PCMs, easing the 

stringent phase change conditions and significantly improving their energy efficiency. 

 

6.5 opportunities toward large-scale systems 

The main advantage of using PCMs in large-scale systems comes from their zero-static power 

consumption. Here, we perform a simple calculation on energy consumption to stress the utility of small 

(or zero) static energy consumption for semi-static active photonics. Let us consider a thermo-optically 

tunable component, which usually consumes 10mW power21,211. The energy consumption to hold the 

system configuration is thus estimated as 10𝑚𝑊 × 3600𝑠 = 36𝐽/ℎ𝑜𝑢𝑟 or 864𝐽/𝑑𝑎𝑦. Compared with the 

typical energy of 1µJ (or less) for switching a PCM device, the static energy consumption is seven orders 

of magnitude larger per hour per switch!! 

In addition to the energy saving benefit, the weak thermo-optic effect usually leads to a very high 

temperature at the heater locations, resulting in thermal crosstalk. Such thermal crosstalk limits the optical 

component density and may cause unreliable operations. PCM devices need a relatively high thermal 

threshold to switch and are unaffected by such thermal crosstalk. Although researchers recently developed 

many algorithms212,222 to compensate for the thermal crosstalk, they require additional software and 

complex control circuits. 

Overcoming the challenges outlined in the previous section will create opportunities to build large-

scale electrically controlled PCM systems for various applications, including quasi-arbitrary unitary 

transformation223, non-blocking switching fabrics224, adding redundancy to optical systems, multi-purpose 

PICs39,41,143,212, optical neural networks67,68,159,175,211, and quantum simulations21. As an overview, Table 8 

compares several state-of-the-art PICs at the device (insertion loss, extinction ratio, active region length, 

and power consumption) and system levels (number of devices, insertion loss, and footprint). We also refer 

the readers to a few review articles222,225,226 on programmable PICs. 



 

 

 

Table 8: State-of-the-art PIC systems comparison 

Ref 
Applic

ation 
Year 

Mechanis

m, 

platform 

Device 

IL 

(dB) 

Device 

ER 

(dB) 

Active 

length 

(µm) 

Device power 

(mW) or 

(mJ/switch) 

# of 

devices 

System 

IL 

(dB) 

System 

footprint 

(µm × µm) 

224 SF 2016 
EO, 

SOI 
0.44 

30 

(18) a 
380 26 56 

6.7 

(14) c 

10700 × 

4400 

143 OPGA 2015 
TO, 

Si3N4 
N.R. N.R. 2100 250 9 1.0 8500 × 3500 

39,212 OPGA 
2017 

(2020) 

TO, 

SOI 
0.6 N.R. 800 b 110 60 0~30 c 1500 × 2000 

41 OPGA 2020 
TO, 

SOI 

0.8 

(9.0) a 

5 

(20) a 
6 1.6 128 

2.8(9.8) 

4(18) c 

300 × 

300b 

67,159 ONN 
2021 

(2022) 

PCM, 

SOI 

0.9 

(6.5) a 

16 

(3.5) a 
10.0 3.81×10-7 4 N.R. N.R. 

21,211 
ONN 

QS 
2017 

TO, 

SOI 
-3×10-3 6.3 33 b 10 48(176) 8.0 1830 × 4100b 

68 ONN 2021 
PCM, 

Si3N4 
0.4 N.R. 2 8×10-7 16/64 12 1140 × 1416 

175 ONN 2019 
PCM, 

Si3N4 
1.5 9 3 7×10-7 4 N.R. 2461 × 270b 

Envisioned PCM systems 
PCM, 

SOI/Si3N4 
0.1 30 <50 10-7 >100 <10 500 × 500 

(ER: extinction ratio, IL: insertion loss, N.R.: not reported, SF: switching fabric, OPGA: optical programmable gate array, ONN: optical neural 

networks, QS: quantum simulation) 

Note: The device power for volatile (non-volatile) devices is in mW (mJ/switching event). aFor drop and through ports, respectively. bEstimated 

from figures. cPerformance for different configurations.  

 

6.5.1 Quasi-arbitrary unitary transformation 

Any arbitrary unitary transformation can be implemented using an array of programmable units 

consisting of a tunable beam splitter and a phase shifter223. Combining these unitary transformations, we 

can implement arbitrary matrix-vector multiplication at the heart of optical signal processing212,227,228, neural 



 

 

networks211, and even quantum simulation21. Although these systems have been demonstrated using the 

thermo-optic effect in the past21,39,211,212, the number of programmable units remains limited due to thermal 

crosstalk, massive energy consumption, and complex control circuit. While a non-volatile programmable 

unit, as shown in Figure 6.2(a), can circumvent these problems. However, no demonstrations using PCMs 

currently exist, due to the challenges in the previous section. While the loss in PCMs can be acceptable for 

a single unit, these losses add up for massive arrays of devices and incur a considerable signal reduction. 

Additionally, the loss can be non-uniform along different paths and further limit the system performance57. 

Therefore, we must ensure an excess insertion loss of <0.1dB and an extinction ratio of > 30dB per 

programmable unit for a large system, with at least ~10 units in the critical path. This is beyond the 

capabilities of traditional PCMs such as GST because 20nm thick c-GST on a standard silicon waveguide 

introduces an insertion loss of around 7dB/μm52,95. Wide bandgap PCMs, e.g., Sb2S3 and Sb2Se3
104,105, 

exhibit a low insertion loss of less than 0.02dB/μm in the same setting, reducing the loss by more than 300 

fold.  The Sb2Se3-based PCM phase shifters55 demonstrate a promising approach to this application, but 

more work is required to further reduce the loss from 0.33dB55 to less than 0.1dB and to achieve more 

operation levels reliably.   

 

6.5.2 Adding optical redundancy with non-blocking switching fabric 

A large PIC may contain some critical components that are more prone to failure, and such single-

device failure can render the whole circuit unusable. A straightforward approach to overcome this is to add 

several backup devices and a switching fabric, bypassing the broken device and rewiring another functional 

device. Figure 6.2(b) shows a proposed schematic of a fully integrated transceiver with multiple on-chip 

lasers and a PCM non-volatile switching fabric. This type of switching fabric demands zero-static energy 

but has no switching speed requirement, rendering PCMs ideal candidates. Other switching fabrics may 

also be established with PCMs, such as the 16 × 16 Benes type non-blocking switch fabric57 in Figure 

6.2(c), which can route all input signals to corresponding output ports simultaneously. We again stress that 



 

 

the successful creation of such a fabric requires optimized insertion loss (< 0.1 𝑑𝐵) and extinction ratio 

(> 30𝑑𝐵). Although this is possible in theoretical designs57,213, currently reported PCM programmable 

units (listed in Table 8) do not yet meet such stringent demands, primarily due to fabrication imperfections 

and residual material losses. Therefore, further research and engineering on loss reduction techniques are 

needed. 

 

Figure 6.2: Schematic of non-volatile arbitrary unitary transformation and switching fabrics.  

(a) Three schematics of programmable units for arbitrary unitary transformation based on PCMs (orange). The 

tunable beam splitter is boxed in purple dotted lines. (b) Proposed switching fabrics for adding optical 

redundancy to all-integrated systems on a chip. Because on-chip lasers are prone to failure, one can fabricate 

several backup lasers together. The PCM-based switching fabric in the dotted light blue box can be placed 

between the laser and the large-scale PICs (such as an array of ring modulators in optical transceivers). The 

switching fabric routes a functioning laser to the PIC using zero-static energy, akin to a free-space 

optomechanical setup. (c) Proposed 16 ×16 Benes type non-blocking switching fabrics for non-volatile signal 

routing. We envision all the tunable beam splitters will be programmed electrically. 

 

6.5.3 Optical neural networks 

Deep neural networks have been an immensely promising platform for many problems, such as image 

recognition and classification229–231, natural language processing232–234, image-to-image translation235,236, 



 

 

artwork generation237, and mathematical problem solving238–240. With the enlarging problem space, 

increasingly larger neural networks are implemented, which, unfortunately, requires prohibitive amount of 

computational resources.  

Photonics has the potential advantage of being lower loss and faster than electronics, providing power-

efficient and high-speed computation. Therefore, implementing part of the artificial neural networks, 

especially the linear operations, using optics is a promising alternative. Recently, rudimentary deep neural 

networks211, convolutional neural networks67,68, reservoir networks241, and generative adversarial 

networks71 have all been implemented using photonics. 

 

Figure 6.3: Schematic of PCM-based non-volatile platform to perform linear operations in an optical 

neural network.  

(a) A forward matrix-vector multiplier for deep neural network based on PCM tunable beam splitters; (b) A 

convolutional block using a tunable attenuator.  

 

In optical neural networks (ONNs), the input vector is encoded in the phase and amplitude of the input 

light. The weights (matrices) are implemented by PICs using tunable devices, such as MZI or crossing 

meshes, as shown in Figure 6.3(a) and (b), respectively. Therefore, a matrix-vector multiplication is 

performed by propagating the light through the PIC.  Notably, once the neural networks are trained, the 

weights are fixed, hence a fixed PIC configuration. In other words, holding the PIC for a long time is 

necessary, making PCM photonics an encouraging solution. Indeed, spiking neurosynaptic networks with 

self-learning capability175 and convolutional neural networks67,68 have already been demonstrated with 

GST. Nevertheless, all the ONNs mentioned above relied on optical control, complicating the laser 



 

 

alignment, and potentially hindering the scalability. On the contrary, electrically controlled PCM devices 

and electronic integrated circuits (EIC) could be packaged together242,243, even monolithically integrated on 

the same chip244,245, offering a higher degree of integration and flexibility. Therefore, electrically controlled 

PCM photonics is a promising path toward energy efficient and large-scale optical neural networks. 

 

6.5.4 Non-volatile electro-optical programmable gate array 

PICs primarily work in application-specific ways, where they are carefully designed and optimized for 

a particular task. However, there is an increasing interest in creating a general-purpose PIC, i.e., developing 

optical programmable gate arrays (OPGAs)39,41,143,212, an idea inspired by the electronic field-programmable 

gate array (FPGA). These OPGAs can be reconfigured to implement different functionalities. Each gate 

consists of the programmable unit described earlier (Fig. 4a) and is connected in meshes39,41,143,212, such as 

the rectangular meshes shown in Figure 6.4(a).   

 

Figure 6.4: Non-volatile programmable gate arrays based on PCMs.  

(a) Schematic of a non-volatile programmable gate array in a rectangular mesh. (b) Schematic of two meshes 

shown in the dotted purple box. Each edge of the mesh denotes a programmable unit. (c) Working principles 

of OPGAs. The light traveling path, denoted by orange, can be changed by reprogramming the OPGA. In this 

way, the OPGA can implement various functions on the same chip, such as delay lines, ring filters, MZI filters, 

and double ring filters.  

 



 

 

 

The resulting chip can perform various functions by appropriately programming the array, as shown in 

Figure 6.4(c). The functionalities of the OPGA are optimal when each unit is non-resonant, making them 

applicable over a broad wavelength range. Obviously, this type of PIC does not require frequent 

reconfiguration. So “set-and-forget” type devices are demanding. To date, most of the demonstrated 

OPGAs are, however, based on thermo-optic effects39,41,143,212, consuming an enormous amount of static 

energy, see Table 8. Besides, the compactness of PCM devices offers more available meshes and 

functionalities39. We see many exciting opportunities in PCM-based zero static-energy OPGAs using the 

wide-bandgap PCMs and the electrically controlled quasi-continuous tuning devices. 

 

6.6 Visible phase-change material photonics  

While integrated photonics research was pioneered in the infrared telecommunication bands, in recent 

years, visible integrated photonics has generated strong interest for many applications, including 

optogenetics246, spectroscopy247,248, augmented reality visors249, and quantum optics250,251. Barring some 

rudimentary demonstrations of PCM-integrated visible photonics54,  this field remains largely unexplored 

due to the narrow bandgap of traditional PCMs and ensuing strong absorption in the shorter wavelengths. 

For example, traditional PCMs, such as GST52, GSST53, GeTe45, and GSSe56, are very lossy at visible 

wavelengths in both amorphous and crystalline phases, see Table 1. Sb2S3, explored in the late twentieth 

century for memory applications252, has a wide bandgap and zero loss in the amorphous state down to a 

wavelength of ~ 600nm104,105. Table 1 compares different PCMs at 640nm, and Sb2S3 shows the best 

performances in all three FOM. Thus, Sb2S3 is particularly suitable for visible or near-IR (>600nm) 

applications, specifically for routing of single photons emitted from semiconductor color centers, such as 

silicon-vacancy (SiV)250 and nitrogen-vacancy (NV)251 centers in diamond.  



 

 

The finite c-Sb2S3 loss in visible wavelength could be mitigated using ring resonators54 or three-

waveguide directional couplers58 on a Si3N4 platform. In Figure 6.5(a), we show the schematic of a tunable 

beam splitter made of Sb2S3 on Si3N4, where the state of Sb2S3 can be switched by external heaters such as 

ITO, graphene, or tungsten. Another possible approach can be to integrate Sb2S3 on wide bandgap 

semiconductors, such as gallium phosphide253, which can be directly doped to create a heater, similar to 

what has been demonstrated in silicon photonics55,59,95.  

One promising application of this wide bandgap PCM could be to create a visible switching fabric to 

connect functional quantum devices and route single photons on-chip. Quantum devices, such as single-

photon sources, on-chip superconducting detectors, or quantum frequency converters, have a moderate to 

low success probability, which diminishes as many such devices are hardwired. One way to circumvent this 

issue is the “pick-and-place” technique, where pre-characterized functional devices are picked up and 

placed on a previously defined PIC254,255. However, the scalability of such approaches may be limited. In 

contrast, by using a non-volatile switching fabric, we can connect selected functional quantum devices on-

chip.  

The emission wavelengths of most solid-state single-photon sources and the atomic transitions of alkali 

and alkaline earth metals used in quantum memory are in the visible or near-infrared wavelength range. By 

using non-volatile and wide bandgap Sb2S3, we can realize low-loss reconfigurable visible PICs. 

Additionally, unlike thermo-optic or free-carrier dispersion effects, PCMs can be efficiently tuned even at 

cryogenic temperatures. The relatively low energy density for switching PCMs55,95 will not provide a 

considerable energy burden to tune them inside a cryostat. Alternatively, these devices can even be 

programmed at room temperature before putting in a cryogenic environment. We note that, for quantum 

photonics, the loss per programmable unit needs to be even lower than what is required for classical 

applications; specifically, we estimate the required loss < 0.01𝑑𝐵 per unit21. 

Sb2S3 can also enable biological applications. For example, Figure 6.5(c) shows the concept of a 

switching fabric that can be used in light delivery for optogenetics246, where by simultaneously stimulating 

selective regions in the brain, collective neural functions can be studied. PCMs render a much safer solution 



 

 

than the previously reported thermal tuning246. Moreover, the programming process could even happen 

before placing the probe inside the brain, mitigating potential heat-related harm. 

 

Figure 6.5: Non-volatile active photonics in the visible employing emerging wide bandgap PCM Sb2S3.  

(a) Schematic of an electrically tunable 2 × 2 tunable beam splitter in the visible spectrum. (b) Schematics for 

switching single photons for quantum applications in a cryogenic temperature, achieving on-demand splitting 

ratios. (c) Concept of a non-volatile switch for optogenetic applications246. 

 

6.7 Laser rewriteable phase-change material integrated photonics 

Lithography-free laser writing256–261 is becoming increasingly popular in integrated photonics as it 

provides significant flexibility for rapid prototyping and on-demand designs. However, traditional laser 

writing techniques in glass require costly fs-pulse lasers. Additionally, they only induce miniscule material 

refractive index change (Δn≈0.02) and are limited to large waveguides (~10μm wide). Moreover, the 

writing process is generally irreversible256.  

As the dominant optical data storage materials, PCMs have been reversibly switched with ns-lasers in 

commercial CD writers60. Compared to fs-lasers, ns-lasers are significantly cheaper. In CDs, data is 

recorded by amorphizing PCM strips and erased with crystallization pulses. The minimal feature size of a 

commercial Blue-Ray CD is as tiny as ~150𝑛𝑚60. For integrated photonics, one major obstacle was, again, 



 

 

the high loss of traditional PCMs. However, the loss can be overcome by exploiting emerging wide bandgap 

PCMs and designing low-loss waveguides.  

 

Figure 6.6: Lithography-free rewritable PICs using ns-lasers and PCMs.  

(a) Proposed laser written PIC in wide bandgap PCM (Sb2Se3 or Sb2S3). Optical waveguides are defined by 

two Sb2S3 strips (in white, pointed by purple arrows) amorphized by a focused ns blue pulse laser and leaving 

the center portion crystalline. Prefabricated grating couplers can provide the input/ output interfaces. The 

photonic circuitry is completely reprogrammable and only limited by material loss and laser writing resolution. 

(b) An example material stack consists of a silicon substrate, a few microns of silicon dioxide, tens of 

nanometers Sb2S3, and a few hundred nanometers Si3N4 cladding. When a strip of Sb2S3 is switched from 

crystalline (dark orange) to amorphous (light orange) phase, a waveguide mode can be confined, located mainly 

in the Si3N4 layer to reduce loss. (c) Concept of photonic devices beyond waveguides that can be implemented 

in this platform, such as directional couplers, multi-mode interferometers, and ring resonators. Here orange 

indicates the unwritten c-Sb2S3. 

 

Laser-written PCM photonic devices have already been demonstrated for meta-optics64 and tunable 

beam splitters93, which were limited to a single device. In Figure 6.6(a), we envision expanding this 

technology to create large-scale rewritable PICs, and in fact a few recent works have been published on the 

perspective178,262,263. A carefully designed material stack builds canvases for rewritable photonic 

architecture. An example material stack in Figure 6.6(b) includes tens of nanometer thick c-Sb2S3 on a few 

microns thick layer SiO2, cladded with a few hundred nanometers of Si3N4. In this configuration, the optical 

guided mode is mainly confined within the Si3N4 cladding layer to ensure low absorptive loss from c-Sb2S3.  



 

 

The relatively thin layer of Sb2S3 and Si3N4 also allows Sb2S3 to be switched entirely and reversibly with 

high spatial resolution from the top. A ns-pulsed laser at a shorter wavelength (e.g., 450nm), similar to a 

CD-writer, selectively “patterns” (i.e., amorphizes) the PCM. The probe laser at IR wavelength is guided 

by the total internal reflection at the interface of patterned a-PCM and unpatterned c-PCM. As illustrated 

in Figure 6.6(c), many other functionalities can potentially be implemented beyond simply routing through 

an optical waveguide, such as ring resonators, directional couplers, and multi-mode interferometers. The 

same chip/wafer can be switched back to its crystalline state by rapid thermal annealing63 and rewritten 

again.  

Compared with previously reported multi-purpose PICs39,41,143,212, the achievable functions in rewritable 

PICs are only limited by the laser writing resolution and the material losses, and not by the discrete meshes. 

Thus, such laser-writable PCM can be considered as programmable PICs at the most extreme level, where 

each pixel can be programmed. This process can be much faster, and more cost-effective for simple designs 

and rapid prototyping than fabricating a PIC in a nanofabrication facility. However, we note that the 

proposed laser writing approach is unlikely to entirely replace traditional nanofabrication. The limited 

spatial resolution of laser writing and repeatability of PCM switching will limit possible applications. 

Instead, it should be considered as a new path - a fast, cheap, and rewritable platform for devices with large 

feature sizes. 

 

 

  



 

 

References 

1. Chen, R. et al. Opportunities and Challenges for Large-Scale Phase-Change Material 

Integrated Electro-Photonics. ACS Photonics 9, 3181–3195 (2022). 

2. Karabchevsky, A., Katiyi, A., Ang, A. S. & Hazan, A. On-chip nanophotonics and future 

challenges. Nanophotonics 9, 3733–3753 (2020). 

3. Chen, H.-T., Taylor, A. J. & Yu, N. A review of metasurfaces: physics and applications. Rep. 

Prog. Phys. 79, 076401 (2016). 

4. Molesky, S. et al. Inverse design in nanophotonics. Nature Photonics 12, 659–670 (2018). 

5. Khorasaninejad, M. et al. Metalenses at visible wavelengths: Diffraction-limited focusing 

and subwavelength resolution imaging. Science 352, 1190–1194 (2016). 

6. Ossiander, M. et al. Extreme ultraviolet metalens by vacuum guiding. Science 380, 59–63 

(2023). 

7. Tseng, E. et al. Neural nano-optics for high-quality thin lens imaging. Nat Commun 12, 6493 

(2021). 

8. Zhan, A. et al. Low-Contrast Dielectric Metasurface Optics. ACS Photonics 3, 209–214 

(2016). 

9. Huang, K., Fang, J., Yan, M., Wu, E. & Zeng, H. Wide-field mid-infrared single-photon 

upconversion imaging. Nature Communications 13, (2022). 

10. Abrams, N. C. et al. Silicon Photonic 2.5D Multi-Chip Module Transceiver for High-

Performance Data Centers. Journal of Lightwave Technology 38, 3346–3357 (2020). 

11. Atabaki, A. H. et al. Integrating photonics with silicon nanoelectronics for the next generation 

of systems on a chip. Nature 556, 349–354 (2018). 

12. Wang, X. et al. Integrated photonic encoder for low power and high-speed image processing. 

Nat Commun 15, 4510 (2024). 



 

 

13. Choi, M. et al. Transferable polychromatic optical encoder for neural networks. Preprint at 

https://doi.org/10.48550/ARXIV.2411.02697 (2024). 

14. Xu, X. et al. 11 TOPS photonic convolutional accelerator for optical neural networks. Nature 

589, 44–51 (2021). 

15. Zhang, H. et al. An optical neural chip for implementing complex-valued neural network. 

Nat Commun 12, 457 (2021). 

16. Liu, C. et al. A programmable diffractive deep neural network based on a digital-coding 

metasurface array. Nature Electronics 5, 113–122 (2022). 

17. Carolan, J. et al. On the experimental verification of quantum complexity in linear optics. 

Nature Photonics 8, 621–626 (2014). 

18. Arrazola, J. M. et al. Quantum circuits with many photons on a programmable nanophotonic 

chip. 591, (2021). 

19. Bao, J. et al. Very-large-scale integrated quantum graph photonics. Nat. Photon. 1–9 (2023) 

doi:10.1038/s41566-023-01187-z. 

20. Carolan, J. et al. Variational quantum unsampling on a quantum photonic processor. Nature 

Physics 16, 322–327 (2020). 

21. Harris, N. C. et al. Quantum transport simulations in a programmable nanophotonic processor. 

Nature Photonics 11, 447–452 (2017). 

22. Park, J. et al. All-solid-state spatial light modulator with independent phase and amplitude 

control for three-dimensional LiDAR applications. Nature Nanotechnology 16, 69–76 (2021). 

23. Li, B., Lin, Q. & Li, M. Frequency–angular resolving LiDAR using chip-scale acousto-optic 

beam steering. Nature 620, 316–322 (2023). 



 

 

24. Zhang, X., Kwon, K., Henriksson, J., Luo, J. & Wu, M. C. A large-scale 

microelectromechanical-systems-based silicon photonics LiDAR. Nature 603, 253–258 

(2022). 

25. Yang, K. Y. et al. Inverse-designed non-reciprocal pulse router for chip-based LiDAR. 

Nature Photonics 14, 369–374 (2020). 

26. Bian, L. et al. A broadband hyperspectral image sensor with high spatio-temporal resolution. 

Nature 635, 73–81 (2024). 

27. Zhou, Z. et al. Electrically tunable planar liquid-crystal singlets for simultaneous 

spectrometry and imaging. Light Sci Appl 13, 242 (2024). 

28. Dolia, V. et al. Very-large-scale-integrated high quality factor nanoantenna pixels. Nat. 

Nanotechnol. 19, 1290–1298 (2024). 

29. Yang, Z., Albrow-Owen, T., Cai, W. & Hasan, T. Miniaturization of optical spectrometers. 

Science 371, (2021). 

30. Zhang, Z. et al. Folded Digital Meta-Lenses for on-Chip Spectrometer. Nano Lett. (2023) 

doi:10.1021/acs.nanolett.3c00515. 

31. Li, A. et al. Advances in cost-effective integrated spectrometers. Light Sci Appl 11, 174 

(2022). 

32. Qiao, Q. et al. MEMS-Enabled On-Chip Computational Mid-Infrared Spectrometer Using 

Silicon Photonics. ACS Photonics (2022) doi:10.1021/acsphotonics.2c00381. 

33. Osgood, R. & Meng, X. Principles of Photonic Integrated Circuits: Materials, Device 

Physics, Guided Wave Design. (Springer International Publishing, Cham, 2021). 

doi:10.1007/978-3-030-65193-0. 



 

 

34. Chen, X. et al. The Emergence of Silicon Photonics as a Flexible Technology Platform. 

Proceedings of the IEEE 106, 2101–2116 (2018). 

35. Bogaerts, W. et al. Programmable photonic circuits. Nature 586, 207–216 (2020). 

36. Shen, Y. et al. Deep learning with coherent nanophotonic circuits. Nature Photon 11, 441–

446 (2017). 

37. Arrazola, J. M. et al. Quantum circuits with many photons on a programmable nanophotonic 

chip. Nature 591, 54–60 (2021). 

38. Rogers, C. et al. A universal 3D imaging sensor on a silicon photonics platform. Nature 590, 

256–261 (2021). 

39. Pérez, D. et al. Multipurpose silicon photonics signal processor core. Nat Commun 8, 636 

(2017). 

40. Liao, L. et al. High speed silicon Mach-Zehnder modulator. Opt. Express, OE 13, 3129–3135 

(2005). 

41. Zhang, W. & Yao, J. Photonic integrated field-programmable disk array signal processor. 

Nat Commun 11, 406 (2020). 

42. Errando-Herranz, C. et al. MEMS for Photonic Integrated Circuits. IEEE Journal of Selected 

Topics in Quantum Electronics 26, 1–16 (2020). 

43. Koeber, S. et al. Femtojoule electro-optic modulation using a silicon–organic hybrid device. 

Light: Science & Applications 4, e255–e255 (2015). 

44. He, M. et al. High-performance hybrid silicon and lithium niobate Mach–Zehnder 

modulators for 100 Gbit s −1 and beyond. Nature Photonics 13, 359–364 (2019). 

45. Xu, P., Zheng, J., Doylend, J. K. & Majumdar, A. Low-Loss and Broadband Nonvolatile 

Phase-Change Directional Coupler Switches. ACS Photonics 6, 553–557 (2019). 



 

 

46. Wuttig, M., Bhaskaran, H. & Taubner, T. Phase-change materials for non-volatile photonic 

applications. Nature Photon 11, 465–476 (2017). 

47. Fang, Z., Chen, R., Zheng, J. & Majumdar, A. Non-volatile reconfigurable silicon photonics 

based on phase-change materials. IEEE Journal of Selected Topics in Quantum Electronics 

1–1 (2021) doi:10.1109/JSTQE.2021.3120713. 

48. Abdollahramezani, S. et al. Tunable nanophotonics enabled by chalcogenide phase-change 

materials. Nanophotonics 9, 1189–1241 (2020). 

49. Raoux, S., Xiong, F., Wuttig, M. & Pop, E. Phase change materials and phase change memory. 

MRS Bull. 39, 703–710 (2014). 

50. Shportko, K. et al. Resonant bonding in crystalline phase-change materials. Nature Materials 

7, 653–658 (2008). 

51. Wuttig, M., Bhaskaran, H. & Taubner, T. Phase-change materials for non-volatile photonic 

applications. Nature Photonics 11, 465–476 (2017). 

52. Zheng, J. et al. GST-on-silicon hybrid nanophotonic integrated circuits: a non-volatile quasi-

continuously reprogrammable platform. Optical Materials Express 8, 1551 (2018). 

53. Zhang, Y. et al. Broadband transparent optical phase change materials for high-performance 

nonvolatile photonics. Nature Communications 10, 4279 (2019). 

54. Fang, Z. et al. Non-Volatile Reconfigurable Integrated Photonics Enabled by Broadband 

Low-Loss Phase Change Material. Advanced Optical Materials 9, 2002049 (2021). 

55. Ríos, C. et al. Ultra-compact nonvolatile photonics based on electrically reprogrammable 

transparent phase change materials. arXiv (2021). 

56. Miscuglio, M. & Sorger, V. J. Photonic tensor cores for machine learning. Applied Physics 

Reviews 7, 031404 (2020). 



 

 

57. Zhang, Q. et al. Broadband nonvolatile photonic switching based on optical phase change 

materials: beyond the classical figure-of-merit. Optics Letters 43, 94 (2018). 

58. Xu, P., Zheng, J., Doylend, J. K. & Majumdar, A. Low-Loss and Broadband Nonvolatile 

Phase-Change Directional Coupler Switches. ACS Photonics 6, 553–557 (2019). 

59. Chen, R. et al. Broadband Nonvolatile Electrically Controlled Programmable Units in Silicon 

Photonics. ACS Photonics 9, 2142–2150 (2022). 

60. Wuttig, M. & Yamada, N. Phase-change materials for rewriteable data storage. Nature Mater 

6, 824–832 (2007). 

61. Pernice, W. H. P. & Bhaskaran, H. Photonic non-volatile memories using phase change 

materials. Appl. Phys. Lett. 101, 171101 (2012). 

62. Ríos, C. et al. Integrated all-photonic non-volatile multi-level memory. Nature Photon 9, 

725–732 (2015). 

63. Rios, C., Hosseini, P., Wright, C. D., Bhaskaran, H. & Pernice, W. H. P. On-Chip Photonic 

Memory Elements Employing Phase-Change Materials. Advanced Materials 26, 1372–1377 

(2014). 

64. Wang, Q. et al. Optically reconfigurable metasurfaces and photonic devices based on phase 

change materials. Nature Photonics 10, 60–65 (2016). 

65. Cheng, Z. et al. Device-Level Photonic Memories and Logic Applications Using Phase-

Change Materials. Advanced Materials 30, 1802435 (2018). 

66. Li, X. et al. Fast and reliable storage using a 5 bit, nonvolatile photonic memory cell. Optica, 

OPTICA 6, 1–6 (2019). 

67. Wu, C. et al. Programmable phase-change metasurfaces on waveguides for multimode 

photonic convolutional neural network. Nat Commun 12, 96 (2021). 



 

 

68. Feldmann, J. et al. Parallel convolutional processing using an integrated photonic tensor core. 

Nature 589, 52–58 (2021). 

69. Cheng, Z., Ríos, C., Pernice, W. H. P., Wright, C. D. & Bhaskaran, H. On-chip photonic 

synapse. Science Advances 3, e1700160. 

70. Ríos, C. et al. In-memory computing on a photonic platform. Science Advances 5, eaau5759. 

71. Wu, C. et al. Harnessing optoelectronic noises in a photonic generative network. Science 

Advances 8, eabm2956. 

72. Feldmann, J. et al. Calculating with light using a chip-scale all-optical abacus. Nat Commun 

8, 1256 (2017). 

73. Michel, A. K. U. et al. Using low-loss phase-change materials for mid-infrared antenna 

resonance tuning. Nano Letters 13, 3470–3475 (2013). 

74. Abdollahramezani, S. et al. Electrically driven reprogrammable phase-change metasurface 

reaching 80% efficiency. Nat Commun 13, 1696 (2022). 

75. Zhang, Y. et al. Electrically reconfigurable non-volatile metasurface using low-loss optical 

phase-change material. Nat. Nanotechnol. 16, 661–666 (2021). 

76. Wang, Y. et al. Electrical tuning of phase-change antennas and metasurfaces. Nature 

Nanotechnology 16, 667–672 (2021). 

77. Julian, M., Williams, C., Borg, S., Bartram, S. & Kim, H. J. Reversible optical tuning of 

GeSbTe phase-change metasurface spectral filters for mid-wave infrared imaging. Optica 7, 

28–30 (2020). 

78. Shalaginov, M. Y. et al. Reconfigurable all-dielectric metalens with diffraction-limited 

performance. Nature Communications 12, 1225 (2021). 



 

 

79. Dong, W. et al. Tunable Mid-Infrared Phase-Change Metasurface. Advanced Optical 

Materials 6, 1–6 (2018). 

80. Badloe, T., Lee, J., Seong, J. & Rho, J. Tunable Metasurfaces: The Path to Fully Active 

Nanophotonics. Advanced Photonics Research 2, 2000205 (2021). 

81. Mikheeva, E. et al. Space and Time Modulations of Light with Metasurfaces: Recent 

Progress and Future Prospects. ACS Photonics (2022) doi:10.1021/acsphotonics.1c01833. 

82. Afridi, A. et al. Electrically Driven Varifocal Silicon Metalens. ACS Photonics 5, 4497–4503 

(2018). 

83. Colburn, S., Zhan, A. & Majumdar, A. Varifocal zoom imaging with large area focal length 

adjustable metalenses. Optica, OPTICA 5, 825–831 (2018). 

84. Wang, Y. et al. Helicity-dependent continuous varifocal metalens based on bilayer dielectric 

metasurfaces. Optics Express 29, 39461 (2021). 

85. Wei, S. et al. A Varifocal Graphene Metalens for Broadband Zoom Imaging Covering the 

Entire Visible Region. ACS Nano 15, 4769–4776 (2021). 

86. Miller, S. A. et al. Large-scale optical phased array using a low-power multi-pass silicon 

photonic platform. Optica, OPTICA 7, 3–6 (2020). 

87. Klopfer, E., Dagli, S., Barton III, D., Lawrence, M. & Dionne, J. A. High-Quality-Factor 

Silicon-on-Lithium Niobate Metasurfaces for Electro-optically Reconfigurable Wavefront 

Shaping. Nano Letters 22, 1703–1709 (2022). 

88. Li, S. Q. et al. Phase-only transmissive spatial light modulator based on tunable dielectric 

metasurface. Science 364, 1087–1090 (2019). 

89. Mansha, S. et al. High resolution multispectral spatial light modulators based on tunable 

Fabry-Perot nanocavities. Light Sci Appl 11, 141 (2022). 



 

 

90. Li, L. et al. Electromagnetic reprogrammable coding-metasurface holograms. Nat Commun 

8, 197 (2017). 

91. Li, J. et al. Addressable metasurfaces for dynamic holography and optical information 

encryption. Science Advances 4, eaar6768 (2018). 

92. Wu, C. et al. Low-Loss Integrated Photonic Switch Using Subwavelength Patterned Phase 

Change Material. ACS Photonics 6, 87–92 (2019). 

93. Delaney, M. et al. Nonvolatile programmable silicon photonics using an ultralow-loss 

Sb2Se3 phase change material. Science Advances 7, eabg3500. 

94. Farmakidis, N. et al. Plasmonic nanogap enhanced phase-change devices with dual electrical-

optical functionality. Science Advances 5, eaaw2687 (2019). 

95. Zheng, J. et al. Nonvolatile Electrically Reconfigurable Integrated Photonic Switch Enabled 

by a Silicon PIN Diode Heater. Advanced Materials 32, 2001218 (2020). 

96. Kato, K., Kuwahara, M., Kawashima, H., Tsuruoka, T. & Tsuda, H. Current-driven phase-

change optical gate switch using indium–tin-oxide heater. Appl. Phys. Express 10, 072201 

(2017). 

97. Meng, J. et al. Electrical Programmable Low-loss high cyclable Nonvolatile Photonic 

Random-Access Memory. (2022) doi:10.48550/arXiv.2203.13337. 

98. Taghinejad, H. et al. ITO-based microheaters for reversible multi-stage switching of phase-

change materials: towards miniaturized beyond-binary reconfigurable integrated photonics. 

Opt. Express, OE 29, 20449–20462 (2021). 

99. Zheng, J. et al. GST-on-silicon hybrid nanophotonic integrated circuits: a non-volatile quasi-

continuously reprogrammable platform. Opt. Mater. Express, OME 8, 1551–1561 (2018). 



 

 

100. Zhang, H. et al. Nonvolatile waveguide transmission tuning with electrically-driven ultra-

small GST phase-change material. Science Bulletin 64, 782–789 (2019). 

101. Zheng, J. et al. Nonvolatile Electrically Reconfigurable Integrated Photonic Switch Enabled 

by a Silicon PIN Diode Heater. Advanced Materials 32, 2001218 (2020). 

102. Wu, C. et al. Low-Loss Integrated Photonic Switch Using Subwavelength Patterned Phase 

Change Material. ACS Photonics 6, 87–92 (2019). 

103. Seok, T. J., Kwon, K., Henriksson, J., Luo, J. & Wu, M. C. Wafer-scale silicon photonic 

switches beyond die size limit. Optica, OPTICA 6, 490–494 (2019). 

104. Delaney, M., Zeimpekis, I., Lawson, D., Hewak, D. W. & Muskens, O. L. A New Family of 

Ultralow Loss Reversible Phase-Change Materials for Photonic Integrated Circuits: Sb2S3 

and Sb2Se3. Advanced Functional Materials 30, 2002447 (2020). 

105. Dong, W. et al. Wide Bandgap Phase Change Material Tuned Visible Photonics. Advanced 

Functional Materials 29, 1806181 (2019). 

106. Ríos, C. et al. Ultra-compact nonvolatile phase shifter based on electrically reprogrammable 

transparent phase change materials. PhotoniX 3, 26 (2022). 

107. Fang, Z. et al. Ultra-low-energy programmable non-volatile silicon photonics based on 

phase-change materials with graphene heaters. Nat. Nanotechnol. 17, 842–848 (2022). 

108. Lu, L. et al. Reversible Tuning of Mie Resonances in the Visible Spectrum. ACS Nano 15, 

19722–19732 (2021). 

109. Moitra, P. et al. Programmable Wavefront Control in the Visible Spectrum Using Low-Loss 

Chalcogenide Phase-Change Metasurfaces. Advanced Materials 35, 2205367 (2023). 

110. Spallholz, J. E. On the nature of selenium toxicity and carcinostatic activity. Free Radical 

Biology and Medicine 17, 45–64 (1994). 



 

 

111. Zheng, J. et al. Nonvolatile Electrically Reconfigurable Integrated Photonic Switch Enabled 

by a Silicon PIN Diode Heater. Advanced Materials 32, 2001218 (2020). 

112. Zhang, H. et al. Nonvolatile waveguide transmission tuning with electrically-driven ultra-

small GST phase-change material. Science Bulletin 64, 782–789 (2019). 

113. Chen, R. et al. Non-volatile electrically programmable integrated photonics with a 5-bit 

operation. Nat Commun 14, 3465 (2023). 

114. Zhang, C. et al. Nonvolatile Multilevel Switching of Silicon Photonic Devices with 

In2O3/GST Segmented Structures. Advanced Optical Materials 11, 2202748 (2023). 

115. Zhou, W. et al. In-memory photonic dot-product engine with electrically programmable 

weight banks. Nat Commun 14, 2887 (2023). 

116. Li, S.-Q. et al. Phase-only transmissive spatial light modulator based on tunable dielectric 

metasurface. Science 364, 1087–1090 (2019). 

117. Park, J. et al. All-solid-state spatial light modulator with independent phase and amplitude 

control for three-dimensional LiDAR applications. Nat. Nanotechnol. 16, 69–76 (2021). 

118. Benea-Chelmus, I.-C. et al. Electro-optic spatial light modulator from an engineered organic 

layer. Nat Commun 12, 5928 (2021). 

119. Benea-Chelmus, I.-C. et al. Gigahertz free-space electro-optic modulators based on Mie 

resonances. Nat Commun 13, 3170 (2022). 

120. Wang, Y. et al. Electrical tuning of phase-change antennas and metasurfaces. Nat. 

Nanotechnol. 16, 667–672 (2021). 

121. Zhang, Y. et al. Electrically reconfigurable non-volatile metasurface using low-loss optical 

phase-change material. Nat. Nanotechnol. 16, 661–666 (2021). 



 

 

122. Wang, Q. et al. Optically reconfigurable metasurfaces and photonic devices based on phase 

change materials. Nature Photon 10, 60–65 (2016). 

123. Sherrott, M. C. et al. Experimental Demonstration of >230° Phase Modulation in Gate-

Tunable Graphene–Gold Reconfigurable Mid-Infrared Metasurfaces. Nano Lett. 17, 3027–

3034 (2017). 

124. García-Márquez, J., López, V., González-Vega, A. & Noé, E. Flicker minimization in an 

LCoS spatial light modulator. Opt. Express, OE 20, 8431–8441 (2012). 

125. Nobukawa, T., Katano, Y., Muroi, T., Kinoshita, N. & Ishii, N. Reduction of spatio-temporal 

phase fluctuation in a spatial light modulator using linear phase superimposition. OSA 

Continuum, OSAC 4, 1846–1858 (2021). 

126. Abdollahramezani, S. et al. Tunable nanophotonics enabled by chalcogenide phase-change 

materials. Nanophotonics 9, 1189–1241 (2020). 

127. Shalaginov, M. Y. et al. Reconfigurable all-dielectric metalens with diffraction-limited 

performance. Nature Communications 12, 1225 (2021). 

128. Galarreta, C. R. de et al. Nonvolatile Reconfigurable Phase-Change Metadevices for Beam 

Steering in the Near Infrared. Advanced Functional Materials 28, 1704993 (2018). 

129. Chu, C. H. et al. Active dielectric metasurface based on phase-change medium. Laser & 

Photonics Reviews 10, 986–994 (2016). 

130. Galarreta, C. R. de et al. Reconfigurable multilevel control of hybrid all-dielectric phase-

change metasurfaces. Optica, OPTICA 7, 476–484 (2020). 

131. Abdollahramezani, S. et al. Electrically driven reprogrammable phase-change metasurface 

reaching 80% efficiency. Nat Commun 13, 1696 (2022). 



 

 

132. Julian, M. N. et al. Reversible optical tuning of GeSbTe phase-change metasurface spectral 

filters for mid-wave infrared imaging. Optica, OPTICA 7, 746–754 (2020). 

133. Leitis, A. et al. All-Dielectric Programmable Huygens’ Metasurfaces. Advanced Functional 

Materials 30, 1910259 (2020). 

134. Chen, R. et al. Low-loss multilevel operation using lossy phase-change material-integrated 

silicon photonics. JOM 4, 031202 (2024). 

135. Yariv, A. Coupled-mode theory for guided-wave optics. IEEE Journal of Quantum 

Electronics 9, 919–933 (1973). 

136. Zheng, J., Zhu, S., Xu, P., Dunham, S. & Majumdar, A. Modeling Electrical Switching of 

Nonvolatile Phase-Change Integrated Nanophotonic Structures with Graphene Heaters. ACS 

Appl. Mater. Interfaces 12, 21827–21836 (2020). 

137. Zhang, Q. et al. Broadband nonvolatile photonic switching based on optical phase change 

materials: beyond the classical figure-of-merit. Opt. Lett., OL 43, 94–97 (2018). 

138. De Leonardis, F., Soref, R., Passaro, V. M. N., Zhang, Y. & Hu, J. Broadband Electro-Optical 

Crossbar Switches Using Low-Loss Ge 2 Sb 2 Se 4 Te 1 Phase Change Material. J. Lightwave 

Technol. 37, 3183–3191 (2019). 

139. Zhang, Y. et al. Transient Tap Couplers for Wafer-Level Photonic Testing Based on Optical 

Phase Change Materials. ACS Photonics 8, 1903–1908 (2021). 

140. Kato, K., Kuwahara, M., Kawashima, H., Tsuruoka, T. & Tsuda, H. Current-driven phase-

change optical gate switch using indium–tin-oxide heater. Appl. Phys. Express 10, 072201 

(2017). 



 

 

141. Chen, Y., Ho, S.-T. & Krishnamurthy, V. All-optical switching in a symmetric three-

waveguide coupler with phase-mismatched absorptive central waveguide. Appl. Opt., AO 52, 

8845–8853 (2013). 

142. Chen, Y., Ho, S.-T. & Krishnamurthy, V. All-optical switching in a symmetric three-

waveguide coupler with phase-mismatched absorptive central waveguide. Applied Optics 52, 

8845 (2013). 

143. Zhuang, L., Roeloffzen, C. G. H., Hoekman, M., Boller, K.-J. & Lowery, A. J. Programmable 

photonic signal processor chip for radiofrequency applications. Optica, OPTICA 2, 854–859 

(2015). 

144. Teo, T. Y. et al. Comparison and analysis of phase change materials-based reconfigurable 

silicon photonic directional couplers. Opt. Mater. Express, OME 12, 606–621 (2022). 

145. Meng, J. et al. Electrical Programmable Low-loss high cyclable Nonvolatile Photonic 

Random-Access Memory. (2022) doi:10.48550/arXiv.2203.13337. 

146. Fang, Z. et al. Non-Volatile Reconfigurable Integrated Photonics Enabled by Broadband 

Low-Loss Phase Change Material. Advanced Optical Materials 9, 2002049 (2021). 

147. Parize, R. et al. In situ analysis of the crystallization process of Sb2S3 thin films by Raman 

scattering and X-ray diffraction. Materials & Design 121, 1–10 (2017). 

148. Constantin-Popescu, C. et al. New phase change materials for active photonics. in Active 

Photonic Platforms 2022 vol. 12196 26–37 (SPIE, 2022). 

149. Geler-Kremer, J. et al. A ferroelectric multilevel non-volatile photonic phase shifter. Nat. 

Photon. 16, 491–497 (2022). 

150. Song, L., Li, H., Dai, D. & Dai, D. Mach–Zehnder silicon-photonic switch with low random 

phase errors. Opt. Lett., OL 46, 78–81 (2021). 



 

 

151. Xu, P., Zheng, J., Doylend, J. K. & Majumdar, A. Low-Loss and Broadband Nonvolatile 

Phase-Change Directional Coupler Switches. ACS Photonics 6, 553–557 (2019). 

152. Li, J. et al. Performance Limits of Phase Change Integrated Photonics. IEEE J. Select. Topics 

Quantum Electron. 30, 1–9 (2024). 

153. Johnson, G. K., Papatheodorou, G. N. & Johnson, C. E. The enthalpies of formation of 

SbF5(l) and Sb2S3(c) and the high-temperature thermodynamic functions of Sb2S3(c) and 

Sb2S3(l). The Journal of Chemical Thermodynamics 13, 745–754 (1981). 

154. Senkader, S. & Wright, C. D. Models for phase-change of Ge2Sb2Te5 in optical and 

electrical memory devices. Journal of Applied Physics 95, 504–511 (2004). 

155. Njoroge, W. K., Wöltgens, H.-W. & Wuttig, M. Density changes upon crystallization of 

Ge2Sb2.04Te4.74 films. Journal of Vacuum Science & Technology A 20, 230–233 (2002). 

156. Rayleigh, Lord. On the Stability, or Instability, of certain Fluid Motions. Proceedings of the 

London Mathematical Society s1-11, 57–72 (1879). 

157. Lian, J., Wang, L., Sun, X., Yu, Q. & Ewing, R. C. Patterning Metallic Nanostructures by 

Ion-Beam-Induced Dewetting and Rayleigh Instability. Nano Lett. 6, 1047–1052 (2006). 

158. Wu, C. et al. Programmable phase-change metasurfaces on waveguides for multimode 

photonic convolutional neural network. Nature Communications 12, 96 (2021). 

159. Wu, C. et al. Harnessing optoelectronic noises in a photonic generative network. Science 

Advances 8, eabm2956 (2022). 

160. Tuma, T., Pantazi, A., Le Gallo, M., Sebastian, A. & Eleftheriou, E. Stochastic phase-change 

neurons. Nature Nanotechnology 11, 693–699 (2016). 



 

 

161. Gutierrez, Y. et al. Interlaboratory Study on Sb2S3 Interplay between Structure, Dielectric 

Function and Amorphous-to-Crystalline Phase Change for Photonics. iScience 25, 104377 

(2022). 

162. Li, X. et al. Fast and reliable storage using a 5 bit, nonvolatile photonic memory cell. Optica 

6, 1–6 (2019). 

163. Zhang, Y. et al. Myths and truths about optical phase change materials: A perspective. 

Applied Physics Letters 118, 210501 (2021). 

164. Erickson, J. R., Shah, V., Wan, Q., Youngblood, N. & Xiong, F. Designing fast and efficient 

electrically driven phase change photonics using foundry compatible waveguide-integrated 

microheaters. Optics Express 30, 13673 (2022). 

165. Masood, A. et al. Comparison of heater architectures for thermal control of silicon photonic 

circuits. in 10th International Conference on Group IV Photonics 83–84 (2013). 

doi:10.1109/Group4.2013.6644437. 

166. Vaziri, S. et al. Ultrahigh thermal isolation across heterogeneously layered two-dimensional 

materials. Science Advances 5, eaax1325 (2019). 

167. Chen, R. et al. Deterministic quasi-continuous tuning of phase-change material integrated on 

a high-volume 300-mm silicon photonics platform. npj Nanophoton. 1, 7 (2024). 

168. Prabhathan, P. et al. Roadmap for phase change materials in photonics and beyond. iScience 

26, 107946 (2023). 

169. Zhang, Y. et al. Electrically reconfigurable non-volatile metasurface using low-loss optical 

phase-change material. Nature Nanotechnology 16, 661–666 (2021). 

170. Meng, J. et al. Electrical programmable multilevel nonvolatile photonic random-access 

memory. Light Sci Appl 12, 189 (2023). 



 

 

171. Xia, J. et al. Ultrahigh Endurance and Extinction Ratio in Programmable Silicon Photonics 

Based on a Phase Change Material with ITO Heater. Laser & Photonics Reviews n/a, 

2300722. 

172. Martin-Monier, L. et al. Endurance of Chalcogenide Optical Phase Change Materials: a 

Review. Optical Materials Express (2022) doi:10.1364/ome.456428. 

173. Rios, C. et al. Integrated all-photonic non-volatile multi-level memory. Nature Photonics 9, 

725–732 (2015). 

174. Stegmaier, M., Ríos, C., Bhaskaran, H., Wright, C. D. & Pernice, W. H. P. Nonvolatile All-

Optical 1 × 2 Switch for Chipscale Photonic Networks. Advanced Optical Materials 5, 

1600346 (2017). 

175. Feldmann, J., Youngblood, N., Wright, C. D., Bhaskaran, H. & Pernice, W. H. P. All-optical 

spiking neurosynaptic networks with self-learning capabilities. Nature 569, 208–214 (2019). 

176. Delaney, M. et al. Nonvolatile programmable silicon photonics using an ultralow-loss Sb 2 

Se 3 phase change material. Sci. Adv. 7, eabg3500 (2021). 

177. Feldmann, J. et al. Parallel convolution processing using an integrated photonic tensor core. 

Nature 589, 52–58 (2021). 

178. Wu, C. et al. Freeform direct-write and rewritable photonic integrated circuits in phase-

change thin films. Science Advances 10, eadk1361 (2024). 

179. Kato, K., Kuwahara, M., Kawashima, H., Tsuruoka, T. & Tsuda, H. Current-driven phase-

change optical gate switch using indium-tin-oxide heater. Applied Physics Express 10, (2017). 

180. Wu, D. et al. Resonant multilevel optical switching with phase change material GST. 

Nanophotonics 11, 3437–3446 (2022). 



 

 

181. Fang, Z. et al. Arbitrary Programming of Racetrack Resonators Using Low-Loss Phase-

Change Material Sb2Se3. Nano Lett. 24, 97–103 (2023). 

182. Fang, Z. et al. Nonvolatile Phase-Only Transmissive Spatial Light Modulator with Electrical 

Addressability of Individual Pixels. ACS Nano 18, 11245–11256 (2024). 

183. Johnson, K. M., McKnight, D. J. & Underwood, I. Smart spatial light modulators using liquid 

crystals on silicon. IEEE Journal of Quantum Electronics 29, 699–714 (1993). 

184. Hornbeck, L. J. Deformable-Mirror Spatial Light Modulators. in Spatial Light Modulators 

and Applications III vol. 1150 86–103 (SPIE, 1990). 

185. van de Groep, J. et al. Exciton resonance tuning of an atomically thin lens. Nat. Photonics 

14, 426–430 (2020). 

186. Wu, P. C. et al. Dynamic beam steering with all-dielectric electro-optic III–V multiple-

quantum-well metasurfaces. Nat Commun 10, 3654 (2019). 

187. Karst, J. et al. Electrically switchable metallic polymer nanoantennas. Science 374, 612–616 

(2021). 

188. Persson, M., Engström, D. & Goksör, M. Reducing the effect of pixel crosstalk in phase only 

spatial light modulators. Opt. Express, OE 20, 22334–22343 (2012). 

189. Moitra, P. et al. Programmable Wavefront Control in the Visible Spectrum Using Low-Loss 

Chalcogenide Phase-Change Metasurfaces. Advanced Materials n/a, 2205367. 

190. Delaney, M., Zeimpekis, I., Lawson, D., Hewak, D. W. & Muskens, O. L. A New Family of 

Ultralow Loss Reversible Phase-Change Materials for Photonic Integrated Circuits: Sb2S3 

and Sb2Se3. Advanced Functional Materials 30, 2002447 (2020). 

191. Wuttig, M. & Yamada, N. Phase-change materials for rewriteable data storage. Nature 

Materials 6, 824–832 (2007). 



 

 

192. Fang, Z. et al. Non-Volatile Reconfigurable Integrated Photonics Enabled by Broadband 

Low-Loss Phase Change Material. Advanced Optical Materials 9, 2002049 (2021). 

193. Chang-Hasnain, C. J. & Yang, W. High-contrast gratings for integrated optoelectronics. Adv. 

Opt. Photon., AOP 4, 379–440 (2012). 

194. Lawrence, M. et al. High quality factor phase gradient metasurfaces. Nature Nanotechnology 

15, 956–961 (2020). 

195. Song, J.-H., van de Groep, J., Kim, S. J. & Brongersma, M. L. Non-local metasurfaces for 

spectrally decoupled wavefront manipulation and eye tracking. Nat. Nanotechnol. 16, 1224–

1230 (2021). 

196. Koshelev, K., Lepeshov, S., Liu, M., Bogdanov, A. & Kivshar, Y. Asymmetric Metasurfaces 

with High-$Q$ Resonances Governed by Bound States in the Continuum. Phys. Rev. Lett. 

121, 193903 (2018). 

197. Liu, Z. et al. High-$Q$ Quasibound States in the Continuum for Nonlinear Metasurfaces. 

Phys. Rev. Lett. 123, 253901 (2019). 

198. Zeng, B., Majumdar, A. & Wang, F. Tunable dark modes in one-dimensional “diatomic” 

dielectric gratings. Opt. Express, OE 23, 12478–12487 (2015). 

199. Overvig, A. C., Shrestha, S. & Yu, N. Dimerized high contrast gratings. Nanophotonics 7, 

1157–1168 (2018). 

200. Kim, J. Y. et al. Full 2π tunable phase modulation using avoided crossing of resonances. Nat 

Commun 13, 2103 (2022). 

201. Zheng, J. et al. Nonvolatile Electrically Reconfigurable Integrated Photonic Switch Enabled 

by a Silicon PIN Diode Heater. Advanced Materials 32, 2001218 (2020). 



 

 

202. Zhang, H. et al. Miniature Multilevel Optical Memristive Switch Using Phase Change 

Material. ACS Photonics 6, 2205–2212 (2019). 

203. Munley, C. et al. Visible Wavelength Flatband in a Gallium Phosphide Metasurface. ACS 

Photonics 10, 2456–2460 (2023). 

204. Nguyen, H. S. et al. Symmetry Breaking in Photonic Crystals: On-Demand Dispersion from 

Flatband to Dirac Cones. Phys. Rev. Lett. 120, 066102 (2018). 

205. Holoeye Spatial Light Modulator - GAEA 2 TELCO 033 | MEETOPTICS. 

https://www.meetoptics.com/adaptive-optics/spatial-light-modulator/s/holoeye/p/GAEA-2-

TELCO-033. 

206. Holoeye Spatial Light Modulator - LC 2012 | MEETOPTICS. 

https://www.meetoptics.com/adaptive-optics/spatial-light-

modulator/s/holoeye/p/LC%202012. 

207. Search results - TI.com. https://www.ti.com/sitesearch/en-

us/docs/universalsearch.tsp?langPref=en-

US&searchTerm=DLP6500&nr=413#q=DLP6500&sort=relevancy&numberOfResults=25. 

208. Benea-Chelmus, I. C. et al. Electro-optic spatial light modulator from an engineered organic 

layer. Nature Communications 12, (2021). 

209. Moitra, P. et al. Electrically Tunable Reflective Metasurfaces with Continuous and Full-

Phase Modulation for High-Efficiency Wavefront Control at Visible Frequencies. ACS Nano 

17, 16952–16959 (2023). 

210. Panuski, C. L. et al. A full degree-of-freedom spatiotemporal light modulator. Nat. Photon. 

16, 834–842 (2022). 



 

 

211. Shen, Y. et al. Deep learning with coherent nanophotonic circuits. Nature Photonics 11, 441–

446 (2017). 

212. Pérez-López, D., López, A., DasMahapatra, P. & Capmany, J. Multipurpose self-

configuration of programmable photonic circuits. Nature Communications 11, (2020). 

213. De Leonardis, F., Soref, R., Passaro, V. M. N., Zhang, Y. & Hu, J. Broadband Electro-Optical 

Crossbar Switches Using Low-Loss Ge2Sb2Se4Te1 Phase Change Material. Journal of 

Lightwave Technology 37, 3183–3191 (2019). 

214. Martin-Monier, L. et al. Endurance of chalcogenide optical phase change materials: a review. 

Opt. Mater. Express, OME 12, 2145–2167 (2022). 

215. Raoux, S., Xiong, F., Wuttig, M. & Pop, E. Phase change materials and phase change memory. 

MRS Bulletin 39, 703–710 (2014). 

216. Kim, S., Burr, G. W., Kim, W. & Nam, S.-W. Phase-change memory cycling endurance. 

MRS Bulletin 44, 710–714 (2019). 

217. Atwood, G. Phase-Change Materials for Electronic Memories. Science 321, 210–211 (2008). 

218. Ohshima, N. Crystallization of germanium–antimony–tellurium amorphous thin film 

sandwiched between various dielectric protective films. Journal of Applied Physics 79, 8357–

8363 (1996). 

219. Aggarwal, S. et al. Antimony as a Programmable Element in Integrated Nanophotonics. Nano 

Letters 22, 3532–3538 (2022). 

220. Gosciniak, J. Ultra-compact nonvolatile plasmonic phase change modulators and switches 

with dual electrical–optical functionality. AIP Advances 12, 035321 (2022). 



 

 

221. Zheng, J., Zhu, S., Xu, P., Dunham, S. & Majumdar, A. Modeling Electrical Switching of 

Nonvolatile Phase-Change Integrated Nanophotonic Structures with Graphene Heaters. ACS 

Applied Materials and Interfaces 12, 21827–21836 (2020). 

222. Xu, X. et al. Self-calibrating programmable photonic integrated circuits. Nat. Photon. 1–8 

(2022) doi:10.1038/s41566-022-01020-z. 

223. Reck, M., Zeilinger, A., Bernstein, H. J. & Bertani, P. Experimental realization of any 

discrete unitary operator. Phys. Rev. Lett. 73, 58–61 (1994). 

224. Lu, L. et al. 16 × 16 non-blocking silicon optical switch based on electro-optic Mach-Zehnder 

interferometers. Optics Express 24, 9295 (2016). 

225. Pérez, D., Gasulla, I. & Capmany, J. Programmable multifunctional integrated nanophotonics. 

Nanophotonics 7, 1351–1371 (2018). 

226. Harris, N. C. et al. Linear programmable nanophotonic processors. Optica 5, 1623 (2018). 

227. Macho-Ortiz, A., Pérez-López, D. & Capmany, J. Optical Implementation of 2 × 2 Universal 

Unitary Matrix Transformations. Laser & Photonics Reviews 15, 2000473 (2021). 

228. Bogaerts, W. et al. Programmable photonic circuits. Nature 586, 207–216 (2020). 

229. LeCun, Y. et al. Handwritten Digit Recognition with a Back-Propagation Network. in 

Advances in Neural Information Processing Systems (ed. Touretzky, D.) vol. 2 (Morgan-

Kaufmann, 1989). 

230. Munder, S. & Gavrila, D. M. An Experimental Study on Pedestrian Classification. IEEE 

Transactions on Pattern Analysis and Machine Intelligence 28, 1863–1868 (2006). 

231. Krizhevsky, A., Sutskever, I. & Hinton, G. E. ImageNet Classification with Deep 

Convolutional Neural Networks. in Advances in Neural Information Processing Systems (eds. 



 

 

Pereira, F., Burges, C. J., Bottou, L. & Weinberger, K. Q.) vol. 25 (Curran Associates, Inc., 

2012). 

232. Bengio, Y., Ducharme, R. & Vincent, P. A Neural Probabilistic Language Model. in 

Advances in Neural Information Processing Systems vol. 13 (MIT Press, 2000). 

233. Collobert, R. & Weston, J. A unified architecture for natural language processing: deep neural 

networks with multitask learning. in Proceedings of the 25th international conference on 

Machine learning 160–167 (Association for Computing Machinery, New York, NY, USA, 

2008). doi:10.1145/1390156.1390177. 

234. Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S. & Dean, J. Distributed Representations 

of Words and Phrases and their Compositionality. in Advances in Neural Information 

Processing Systems (eds. Burges, C. J., Bottou, L., Welling, M., Ghahramani, Z. & 

Weinberger, K. Q.) vol. 26 (Curran Associates, Inc., 2013). 

235. Isola, P., Zhu, J.-Y., Zhou, T. & Efros, A. A. Image-To-Image Translation With Conditional 

Adversarial Networks. in 1125–1134 (2017). 

236. Zhu, J.-Y., Park, T., Isola, P. & Efros, A. A. Unpaired Image-To-Image Translation Using 

Cycle-Consistent Adversarial Networks. in 2223–2232 (2017). 

237. Karras, T., Laine, S. & Aila, T. A Style-Based Generator Architecture for Generative 

Adversarial Networks. in 4401–4410 (2019). 

238. Lagaris, I. E., Likas, A. & Fotiadis, D. I. Artificial neural networks for solving ordinary and 

partial differential equations. IEEE Transactions on Neural Networks 9, 987–1000 (1998). 

239. Raissi, M., Perdikaris, P. & Karniadakis, G. E. Physics-informed neural networks: A deep 

learning framework for solving forward and inverse problems involving nonlinear partial 

differential equations. Journal of Computational Physics 378, 686–707 (2019). 



 

 

240. Sirignano, J. & Spiliopoulos, K. DGM: A deep learning algorithm for solving partial 

differential equations. Journal of Computational Physics 375, 1339–1364 (2018). 

241. Vandoorne, K. et al. Experimental demonstration of reservoir computing on a silicon 

photonics chip. Nature Communications 5, 1–6 (2014). 

242. Li, H. et al. A 3-D-Integrated Silicon Photonic Microring-Based 112-Gb/s PAM-4 

Transmitter With Nonlinear Equalization and Thermal Control. IEEE Journal of Solid-State 

Circuits 56, 19–29 (2021). 

243. Li, H. et al. A 112 Gb/s PAM4 Silicon Photonics Transmitter with Microring Modulator and 

CMOS Driver. Journal of Lightwave Technology 38, 131–138 (2020). 

244. Atabaki, A. H. et al. Integrating photonics with silicon nanoelectronics for the next generation 

of systems on a chip. Nature 556, 349–353 (2018). 

245. Moazeni, S. et al. A 40-Gb/s PAM-4 Transmitter Based on a Ring-Resonator Optical DAC 

in 45-nm SOI CMOS. IEEE Journal of Solid-State Circuits 52, 3503–3516 (2017). 

246. Mohanty, A. et al. Reconfigurable nanophotonic silicon probes for sub-millisecond deep-

brain optical stimulation. Nat Biomed Eng 4, 223–231 (2020). 

247. Cheng, R. et al. Broadband on-chip single-photon spectrometer. Nat Commun 10, 4104 

(2019). 

248. Momeni, B., Hosseini, E. S. & Adibi, A. Planar photonic crystal microspectrometers in 

silicon-nitride for the visible range. Opt. Express, OE 17, 17060–17069 (2009). 

249. Notaros, J., Raval, M., Notaros, M. & Watts, M. R. Integrated-Phased-Array-Based Visible-

Light Near-Eye Holographic Projector. in 2019 Conference on Lasers and Electro-Optics 

(CLEO) 1–2 (2019). doi:10.1364/CLEO_SI.2019.STu3O.4. 



 

 

250. Hepp, C. et al. Electronic Structure of the Silicon Vacancy Color Center in Diamond. Phys. 

Rev. Lett. 112, 036405 (2014). 

251. Schirhagl, R., Chang, K., Loretz, M. & Degen, C. L. Nitrogen-Vacancy Centers in Diamond: 

Nanoscale Sensors for Physics and Biology. Annual Review of Physical Chemistry 65, 83–

105 (2014). 

252. Arun, P., Vedeshwar, A. G. & Mehra, N. C. LASER-INDUCED CRYSTALLIZATION IN 

Sb2S3 FILMS. Materials Research Bulletin 32, 907–913 (1997). 

253. Wilson, D. J. et al. Integrated gallium phosphide nonlinear photonics. Nature Photonics 14, 

57–62 (2020). 

254. Wan, N. H. et al. Large-scale integration of artificial atoms in hybrid photonic circuits. 

Nature 583, 226–231 (2020). 

255. Kim, J.-H., Aghaeimeibodi, S., Carolan, J., Englund, D. & Waks, E. Hybrid integration 

methods for on-chip quantum photonics. Optica, OPTICA 7, 291–308 (2020). 

256. Davis, K. M., Miura, K., Sugimoto, N. & Hirao, K. Writing waveguides in glass with a 

femtosecond laser. Opt. Lett., OL 21, 1729–1731 (1996). 

257. Sotillo, B. et al. Diamond photonics platform enabled by femtosecond laser writing. Sci Rep 

6, 35566 (2016). 

258. Meany, T. et al. Laser written circuits for quantum photonics. Laser & Photonics Reviews 9, 

363–384 (2015). 

259. Dias, A. et al. Femtosecond laser writing of photonic devices in borate glasses 

compositionally designed to be laser writable. Opt. Lett., OL 43, 2523–2526 (2018). 

260. Valle, G. D., Osellame, R. & Laporta, P. Micromachining of photonic devices by 

femtosecond laser pulses. J. Opt. A: Pure Appl. Opt. 11, 013001 (2008). 



 

 

261. Juodkazis, S., Mizeikis, V. & Misawa, H. Three-dimensional microfabrication of materials 

by femtosecond lasers for photonics applications. Journal of Applied Physics 106, 051101 

(2009). 

262. Miller, F. et al. Rewritable photonic integrated circuits using dielectric-assisted phase-change 

material waveguides. Opt. Lett., OL 48, 2385–2388 (2023). 

263. Miller, F. et al. Rewritable Photonic Integrated Circuit Canvas Based on Low-Loss Phase 

Change Material and Nanosecond Pulsed Lasers. Nano Lett. 24, 6844–6849 (2024).