2026 workshop program
AIDonut : First Real-Time Neural Network Control of Telescope Active Optics the Vera Rubin Observatory
The work
Abstract
We present the first successful deployment of a deep learning model for real-time wavefront estimation in the Rubin Observatory Active Optics System (AOS). Building on simulation-trained convolutional neural networks, we applied transfer learning techniques to adapt our model to real telescope data and integrated it into the AOS pipeline for live operations. After testing multiple variations of transfer learning, we find that training directly on real data outperforms all transfer learning strategies, and therefore extend the training set to include data through November. Our approach achieves 30-40 times faster than Danish solver, enabling parallel processing of multiple wavefront sensor images, which is critical for Rubin’s rapid 36-second cadence. Under ideal conditions, the model reduces the median wavefront estimation error by 20.5\% compared to Danish. Performance are consistent in challenging conditions such as blending, high airmass, low signal to noise ratios. Closed-loop testing confirmed the model meets the AOS requirement and successfully converges to required optical quality (PSF FWHM $<$ 0.19”, PSSN $>$ 0.91). This marks the first time a neural network has been used to control telescope optics in real-time operations.