Hillhouse, Hugh WHeins, William Vincent2026-02-052026-02-052026-02-052025Heins_washington_0250O_29181.pdfhttps://hdl.handle.net/1773/55162Thesis (Master's)--University of Washington, 2025Halide perovskite (HP) photovoltaics (PV) are a high-performance, low-cost alternative to traditional crystalline silicon (c-Si) PV, but their short operational lifetimes prevent them from reaching commercial scale. Alongside other processes, HPs chemically degrade under elevated temperature, oxygen, moisture, illumination, and electrical bias, and this chemical decomposition is the primary driver of HP PV performance decline. In previous studies, we proposed mechanisms and developed kinetic rate law models for the chemical decomposition of three relevant HP materials, and we developed predictive machine learning (ML) models for the optoelectronic properties of these HP films and for device operational lifetimes. Although, no mechanistic models of device performance decline exist. Additionally, our initial predictive models of operational lifetimes require our kinetic rate law models to be accurate, and these models are both composition-specific and time-intensive to develop. In this work, we fill these essential gaps by mechanistically modeling degrading perovskite solar cells (PSCs) and improving our predictive ML models for their operational lifetimes. Specifically, we globally fit the non-ideal diode model and a custom, one-dimensional (1D) drift-diffusion model to light and dark current-voltage (J-V) scans over time for devices degrading under varying temperatures, oxygen concentrations, humidities, and illumination intensities. Moreover, we quantify effective fractional active areas and thicknesses of HP films over degradation from in situ dark-field (DF) microscopy measurements, constituting an effective model of degradation profile over degradation. The extracted diode and drift-diffusion fitting parameters and their corresponding derived parameters are mechanistic properties of the device, and coupled with these effective degradation profile parameters, the evolutions of and correlations among these parameters over degradation illuminate the mechanisms of device performance decline. Furthermore, in a unique cumulative sensitivity analysis (CSA), we calculate the exact influence of each fitting parameter on each solar cell parameter over time, quantifying the exact influence of each degradation mechanism on device performance decline. Last, we analyze the relationships between parameter evolutions and degradation conditions including Arrhenius modeling to inform the development of both accelerated aging models and long-lived device design in future work. Overall, these analyses constitute the most advanced mechanistic model of PSC performance decline to date. Then, beyond mechanistic modeling, we utilize these various parameter sets as features in predictive machine learning (ML) models of operational lifetimes (T80), achieving a champion model applicable to all three of our PSC architectures with a mean-normalized root-mean-squared (RMS) error (NRMSE) of 26.5% using features derived only from temperature and the first 20 J-V measurements, without requiring the other degradation conditions or the composition-specific kinetic rate law models necessary for our previous predictive ML model. The predictive ML models developed in this study are the strongest we have produced in both accuracy and applicability, and we thus demonstrate the ability to use empirical and modeling parameters as features to construct predictive ML models for the operational lifetimes of multiple PSC architectures from small device degradation datasets using common, low-cost electronic measurements (e.g., J-V scans). This establishes a strong foundation for and represents a powerful step toward high-throughput device testing and thus long-lived device development.application/pdfen-USnoneMachine LearningMechanisticModelingPerovskitePredictiveSolar CellsEnergyChemical engineeringThesis