Automated Transport of Micro-Objects With Complex Trap Configurations in Holographic Optical Tweezers

Abstract

Optical tweezers use a focused laser beam to trap micro and nano-scale objects through ascattering and gradient force exhibited due to the properties of light. A specialized holographic optical tweezer (HOT) which is centered around the use of a spatial light modulator can generate multiple optical traps in the workspace through holograms. HOTs have found many use cases in micro-robotics and the medical field, where they have been used for cell characterization, cell-to-cell interactions, and micro-robotic actuators. We will begin by presenting a HOT specific framework on which to build control systems intended to augment user control of HOTs. We then integrate the framework with our HOT system and develop a model predictive control (MPC) based adaptive path planning control system capable of transporting polystyrene micro-beads to create formations in the workspace. We develop the MPC based system to consider multiple trap configurations including point, annular, and line traps. The latter two of which can be used to trap multiple micro-beads in a ring or line configuration respectively. The control system leverages GPU compute and Python parallelization to allow for simultaneous path planning and control for multiple optical traps. We then demonstrate the effectiveness of the controller by allowing it to created various micro-object formations automatically while avoiding collisions with obstacles in a crowded workspace. Then, we will focus on the phase retrieval process, another area of research in the HOT field and other applications involving the use of holograms. Phase retrieval is the process of calculating a phase mask for addressing onto an SLM to form a desired hologram, or in the case of HOTs, an intensity distribution in the image plane. We investigate the capability for deep learning models, specifically the convolutional neural network architecture U-Net, to act as the mechanism for phase calculation during operation of the HOT system. By using Red Tweezers, an open-source hologram engine built for calculating phase masks on our HOT system, for creating synthetic training data. Then by carefully choosing hyperparameters and training data distributions we train the model on 2D intensity distributions of the far-field and the phase masks generated by Red Tweezers. We then conducted trials on our HOT system and demonstrate the capability of the U-Net to act as the phase retriever in a single pass. The U-Net allows for constant phase calculation times, whereas the Red Tweezers algorithm would be slowed to update rates insufficient for real-time control using large numbers of optical traps. We aim for the methodologies developed for both of these topics to be useful for further research into autonomy integrated HOT systems.