Early results of new data driven machine learning (ML) methods for wavefront estimation have shown some potentially useful improvements in simulation, however rigorous analysis of these simulated results in the astronomical domain have been difficult to evaluate due to the data driven nature of the neural networks applied. Due to their opaque operational nature, these methods require robust performance analysis before they can be considered for practical use.

We present a careful error analysis from the projection of the wavefront estimates on a set of Karhunen-Loeve modes, and compare with similar statistics generated for simulated benchmarks. We then use this technique to examine previously published estimation techniques that estimate wavefronts from Shack-Hartmann wavefront sensor images, outlining their relative strengths and weaknesses for pseudo open-loop control and point spread function reconstruction (PSF-R).

Finally, we conduct a thorough analysis of the effects of noise on conditional Generative Adversarial Network (cGAN) and UNet wavefront estimation methods and provide an assessment of suitability to control and PSF-R applications from simulated results. We find that the UNet Assisted control performance is particularly robust to photon and read-out noise when applied in simulation for low photon regimes.

Jeffrey Smith, Australian National University