Machine Learning Wavefront Sensing for the James Webb Space Telescope

The fine alignment of the James Webb Space Telescope (JWST)'s 18 primary mirror segments relies on image-based wavefront sensing. Previously a convolutional neural network (CNN) produced phase estimates with around 0.37 waves RMS error, which were adequate as starting guesses for phase retrieval to accurately predict global Zernike coefficients. In contrast, our study uses a CNN to sense segment piston errors between the 18 segments. Our trained CNN model could retrieve segment piston phase errors of +/- 0.5 waves from a single defocused point-spread function to within 0.02 waves RMS, without the need of additional phase retrieval algorithms.