Wavefront Prediction Using Artificial Neural Networks for Adaptive Optics

Latency in the control loop of Adaptive Optics (AO) systems can severely limit its performance. Theories describing the temporal evolution of the atmospheric turbulence, such as the frozen flow hypothesis, justify the feasibility of predicting the turbulence (or equivalently its measurements) to compensate for the resultant temporal error in the system. This will mostly benefit AO assisted High Contrast Imaging (HCI) instruments for enhanced contrast, or wide-field AO systems for improved sky coverage. In this thesis, we explore the potential of an Artificial Neural Network (ANN) as a nonlinear tool for open-loop wavefront prediction. The ANN predictor composes mainly Long Short-Term Memory (LSTM) cells, an ANN type specialised in sequence modelling and prediction. We demonstrate the efficiency and robustness of an ANN predictor both with simulated and on-sky 7 × 7 Shack-Hartmann Wavefront Sensor (SHWFS) CANARY data measured at 150 Hz, an AO demonstrator on the 4.2 m William Herschel Telescope (WHT), La Palma. We provide evidence that in addition to accurately predicting the wavefronts, an ANN predictor is also filtering high temporal frequencies such as Wavefront Sensor (WFS) noise. We show that an ANN predictor is adaptive to time-variant turbulence on sub-second level without user tuning. Specifically, we show that an ANN predictor is capable of predicting both frozen flow and non-frozen flow such as dome seeing, and that the ANN prediction can be based on a per-subaperture basis. As a pioneer, this thesis examines in great detail the characteristics of an ANN wavefront predictor and provides implications towards an on-sky implementation