Deploying AI for active optics control of the Vera C. Rubin Observatory
TL;DR
Two AI-based methods for accelerating wavefront inference are described: AIDonut, a neural-network replacement for the wavefront-estimation step, and TARTS, a larger framework that incorporates donut detection and sensor-level aggregation.
Abstract
The Vera C. Rubin Observatory uses an active optics system (AOS) to deliver consistently high image quality across its 9.6 deg2 field of view. The baseline forward-modeling approach provides physically interpretable wavefront fits, but challenges the cadence required for the Legacy Survey of Space and Time (LSST). We describe two AI-based methods for accelerating wavefront inference: AIDonut, a neural-network replacement for the wavefront-estimation step, and TARTS, a larger framework that incorporates donut detection and sensor-level aggregation. Both systems are trained on Rubin simulations and adapted to Rubin commissioning data through transfer learning, using supervised and unsupervised approaches. Early validation in on-sky tests shows both models converge to image quality and alignment consistent with the forward-modeling baseline. These demonstrations are an important step towards fast, AI-assisted active optics control for Rubin, with detailed physical models remaining central to training, validation, and operational monitoring.