What the speck is that? Improving exoplanet imaging sensitivity by combining machine learning and physical models
TL;DR
By combining a physical optics model with a flexible machine learning model, this work can avoid introducing unphysical features in the resulting PSF estimate and explores the temporal coherence of structures in telemetry and the limitations of learning this relationship from data sampled on such different timescales.
Abstract
Extreme adaptive optics observations of exoplanetary systems are limited at the smallest separations by speckle noise introduced by diffractive effects in the telescope and instrument system. To probe for lower-mass planets on close orbits, we need to improve starlight suppression with both optics and post-processing techniques. The promise of neural networks as universal function approximators led us to investigate their ability to predict a time-evolving point-spread function (PSF) from instrument telemetry (e.g. wavefront sensor data). By combining a physical optics model with a flexible machine learning model, we can avoid introducing unphysical features in the resulting PSF estimate. To make effective use of 2–3.7 kHz wavefront sensor data in analyzing science images taken at 5–10 Hz, we explore the temporal coherence of structures in our telemetry and the limitations of learning this relationship from data sampled on such different timescales.