Focal-plane wavefront control using deep learning for high-contrast imaging

On ground-based telescopes, high contrast imaging (HCI) systems suffer the common limitations of atmospheric turbulence that produce phase aberrations on the wavefront. While Adaptive Optics (AO) systems are adept at correcting these aberrations, non-common path aberrations (NCPAs) require additional intervention. NCPAs are caused by the wavefront sensor (WFS) measuring and correcting for a wavefront that is different from the wavefront affecting the science images. These aberrations introduce biases to observations that can be misinterpreted as exoplanets. In the past years we have developed focal-plane wavefront sensing (FPWFS) for vortex coronagraphs, exploring various techniques to lift the sign-ambiguity on even Zernike modes and to estimate NCPAs. Using Deep Learning (DL) algorithms, we have trained models on large laboratory datasets of Zernike coefficients with their associated images, and we have used those models to identify and correct aberrations (Quesnel 2024). Here, we build upon our previous approaches by replacing the DL algorithms with Reinforcement Learning (RL) that allow the real-time training and correction of NCPAs such as water vapor seeing. We have created a simulation that mimics the Mid-infrared ELT Imager and Spectrograph (METIS) and uses RL algorithms to correct the simulated aberrations caused by water vapor seeing. We are currently fine-tuning RL algorithms with the aim to eventually conduct on-sky demonstrations. The development and on-sky demonstration of this framework would be a major milestone for the deployment of FPWFS on METIS and could prove highly valuable for future generations of HCI instruments.