Learning wavefront shaping through reinforcement learning in a simulation environment

We propose a reinforcement learning based architecture for focusing light from a coherent light source through a thick scattering sample using a spatial light modulator. The light source, modulator, sample and the image sensor are modeled in COMSOL environment. The diffraction pattern recorded at the image sensor is then transformed into MATLAB where an RL agent is used to predict the pattern that will focus the light through the sample. The SLM pattern is then updated accordingly. The SLM is modeled as a discrete set of block designed to introduce a phase shift between 0 and 2 depending on the output of the RL agent. The state space of the RL agent in defined as an array of scalar values in [0,1]. The effect of resolution, pixel pitch and sample thickness are discussed. The samples are modeled as blocks of randomly distributed refractive indices. The model is trained on a set of 40 samples. Then training and test performances are reported. Further, applications on dynamic samples are discussed and results are presented for a number of configurations. The focused and unfocused diffraction patterns are presented for various sample thicknesses and at different epochs.