In this work, we systematically explore the effectiveness of optical aberration correction over a wide range of optical depths using the Segmented and Zernike Polynomial bases. We manufacture a set of 16 optical phantom samples consisting of a clear polyester resin substrate embedded with varying concentrations of aluminum oxide (Al₂O₃) scattering particles. We measured the optical depth of the sample set, recording a range τ = 0−8.5. Using these samples to distort a collimated laser wavefront, we conducted an experiment to determine the range of optical depths over which each optimization basis was effective. The results showed that the Zernike basis was more effective at low optical depths (τ < 2), and the Segmented bases was more effective at higher optical depths (τ > 3).

In a separate experiment, we trained a ResNet Convolutional Neural Network to correct for optical aberrations using single shot correction. First, we generated a dataset of 8 million simulated optical aberrations using random phase modulation patterns, generated from a set of 34 Zernike Polynomials, loaded on to a liquid crystal spatial light modulator. We then applied a transfer learning technique to train an 18 layer ResNet neural network, pretrained on ImageNet, on the simulated dataset. Once trained, we tested the ability of the network to correct for new, randomly generated aberrations. Preliminary data showcases the trained network’s ability to correct for new optical aberrations with high efficacy on a millisecond timescale.