Wavefront Estimation with Image-to-Image Translation for Astronomical Instrumentation : AI spectacles for large telescopes

Abstract

The current generation and in-construction next generation of large optical telescopes require accurate estimation of the wavefront error to compensate for image distortions in observations. As there is no direct method of measuring the wavefront directly, it is estimated through transformation to an intensity plot using sophisticated wavefront sensors such as the ShackHartmann Wavefront Sensor (SH-WFS). In particular, a coarse map of the wavefront is indicated by average slope measurements taken from a physical array of lenslets in the SH-WFS. The SH-WFS is incorporated in an Adaptive Optics (AO) system that uses these wavefront estimates to drive a deformable mirror to compensate for aberrations in real-time. We introduce a new method of wavefront estimation that can interpret high spatial frequency information available in the existing SH-WFS of an AO system through the adaptation and development of conditional Generative Adversarial Networks. We investigate these network-based wavefront estimation methods for Point Spread Function Reconstruction (PSF-R) and AO real-time control by designing experiments and training artificial neural networks to demonstrate the benefits of using Machine Learning to enhance wavefront estimation in simulation. We have undertaken a deep simulation study of both applications, demonstrating that these are significantly improved in experimentation with simulated data over the range of expected atmospheric conditions. We find that our data driven methods can significantly exceed performance of incumbent processes for PSF-R (three orders of magnitude) and real-time control (up to 10% Strehl Ratio) of AO systems. We apply statistical methods to examine the wavefront estimates of our networks to explain our assumptions and motivate design choices for each application. These networks are then subject to analysis with simulated noise effects where we demonstrate the strengths and limitations of our network-based methods at the limits of photon flux for our simulated instruments. Our results have the potential to increase the sky coverage of existing and new telescopes, and have exciting prospects for eXtreme Adaptive Optics (XAO) applications such as exoplanet discovery Show more