Wavefront prediction using artificial neural networks with CANARY telemetry

In this paper we use artificial neural networks (ANNs) as a nonlinear wavefront predictor for CANARY open-loop telemetry. CANARY is a single channel multi-object adaptive optics demonstrator hosted by the 4.2 m William Herschel Telescope on La Palma island. These datasets were taken by the on-axis 7×7 NGS (natural guide star) Shack-Hartmann wavefront sensor between 28 September and 2 October, 2017. The ANN predictor is trained in simulations, assuming frozen flow turbulence. We show that the ANN predictor did not improve the system performance with a two-frame latency in terms of residual wavefront errors. Analyses with auto-covariance maps show that a stationary layer was observed by CANARY during those nights, indicative of strong dome seeing. This implies the need of a more representative turbulence model for training ANN predictors with non frozen flow observations.