Testing of machine learning wavefront sensing algorithms on the Tiny Observatory for Telescope Optimization (TOTO) testbed
TL;DR
This work considers a Machine Learning model trained originally on simulated data, and then augmented with real focus diversity data from the Tiny Observatory for Telescope Optimization (TOTO) testbed at the University of Arizona, and compares the wavefront sensing performance of the Machine Learning model with known truth values of the generated dataset.
Abstract
Phase retrieval techniques are utilized to correct low order wavefront aberrations originating from misalignments of the optical system in space based telescope concepts. Traditional phase retrieval involves observation of the Point Spread Function (PSF) and a diversity measurement, usually focus diversity although other measures are possible, to reconstruct the incident wavefront at the science detector. We consider a Machine Learning model trained originally on simulated data, and then augmented with real focus diversity data from the Tiny Observatory for Telescope Optimization (TOTO) testbed at the University of Arizona. We then compare the wavefront sensing performance of the Machine Learning model with known truth values of the generated dataset. The model predictions for low order Zernikes on TOTO data after training and validation show a reasonable agreement with the true Zernike coefficients.