Dual-plane wavefront sensing using a vision transformer

Image-based wavefront sensing using deep-learning allows Zernike coefficients to be estimated directly from intensity measurements. To date, the majority of experiments have focused on using convolutional neural networks to estimate coefficients. Here, we demonstrate a dual-plane wavefront sensor trained using a vision transformer model and compare its performance to that of the widely used convolutional neural network (CNN) architecture. Both results of experiment and simulation indicate that the vision transform can outperform the CNN where image data is significantly downsampled, due to the former's ability to more accurately predict high-order Zernike coefficients.