The future looks dark: improving high contrast imaging with hyper-parameter optimization for data-driven predictive wavefront control

Abstract

The direct imaging and characterization of exoplanets requires extreme adaptive optics (XAO), achieving exquisite wavefront correction (upwards of 90$\%$ Strehl) over a narrow field of view (a few arcseconds). For these XAO systems the temporal error is often a leading term in the error budget, wherein the wavefront evolves faster than the lag between wavefront sensing and control. For atmospheres with high-velocity wind layers, this can result in a wind-driven halo in the coronagraphic dark-zone, limiting sensitivity to faint, close-in companions. The AO system's lag-time is often limited by the wavefront sensor exposure time, especially in the case of fainter guidestars. Predictive control mitigates the temporal error by predicting the shape of the wavefront by time the system correction is applied. One such method of prediction is empirical orthogonal functions (EOF), wherein previous states in the wavefront sensor history are used to learn linear correlations with a minimization problem. This method has been demonstrated on-sky at Subaru/SCExAO and Keck/NIRC2, but has yet to be optimized. With this work as a starting point, we explore the optimal filter hyper-parameter space for implementing EOF on-sky, study its stability under varying atmospheric parameters, and discuss future paths for facilitization of predictive control. This work not only offers a pathway to optimize Keck and Subaru observing, but also acts as a pathfinder for predictive control methods with extremely large telescopes.