SAXO+, the second-stage adaptive optics for SPHERE: a review of six selected control algorithms and the results achieved to date

Abstract

SAXO+ is the second-stage adaptive optics (AO) system that will be added in cascade after SPHERE’s current extreme AO system at the VLT. This upgrade of SPHERE aims to improve coronagraphic contrast by a factor of 10 or more, close to the star, and to extend the performance to fainter, redder targets. To achieve these goals, SAXO+ uses a modulated pyramid wavefront sensor (PyWFS) and an AO loop running more than twice as fast as the current system. A classical optimal modal gain integrator is the current baseline control algorithm. We investigate five alternative algorithms that can provide useful properties, such as prediction to mitigate delay, improved reconstruction for faint objects, and increased tolerance to potential PyWFS nonlinearities. Three of these algorithms are based on linear estimation: LQGSPH, a Linear Quadratic Gaussian estimator, DD4AO, a data-driven method, and TAO-SP, an iterative minimum-variance predictive reconstructor. The other two rely on neural networks: Policy Optimization for AO (PO4AO) and Dual-stage Supervised & Reinforcement Learning (DSRL). All these algorithms use machine learning and are pushed to their limits for SAXO+. Their properties of robustness and unsupervised operation are highly desirable and will be evaluated on-sky. For now, the algorithms are compared using the same COMPASS simulator. Here, we present the current status of this comparison and the lessons learned so far. Final conclusions will come from on-sky comparisons, enabled by the real-time computer (RTC), which has the challenging ability to quickly switch between algorithms.