Machine learning for image-based wavefront sensing

High-contrast imaging systems in ground-based astronomy rely on a precise control
of the wavefront. On one hand, atmospheric turbulence distorts the wavefront which
is corrected by a dedicated adaptive optics system. On the other hand, non-common
path aberrations between wavefront sensor and scientific paths can also overwhelm
a putative scientific signal and need to be corrected. Hence, measuring precisely
the wavefront at the scientific focal plane is of prime interest for applications
such as direct imaging of exoplanets. While early attempts using neural networks
showed some successful results, the latest advances in machine learning have yet
to be fully exploited for the problem of focal plane wavefront sensing. In this
paper, we explore the use of convolution neural networks to perform image-based
wavefront sensing. Based on simulated data, we evaluate neural architectures on
two different data sets, one with only low order aberrations (20 Zernike modes)
and one including higher orders modes (100 Zernike modes). We discuss the
accuracy reached in both cases, and we show that direct phase map reconstruction
outperforms classical modal approaches. The precision achieved ranges typically
between 1% and 10% of the injected wavefront. Finally, we explore the impact of
phase diversity, and we compare our optimized CNN model to a standard iterative
phase retrieval algorithm.