Abstract

The rise of extreme Large Telescopes (ELTs) poses challenges for high-resolution phase map reconstruction. Despite the pyramid wavefront sensor (PyWFS) promise, its inherent non-linearity is limiting. This study proposes techniques to enhance the non-modulated PyWFS linearity through deep learning, comparing convolutional Neural Networks (CNNs) models (Xception, WFNet, ConvNext) with the transformer model Global Context Vision Transformers (GCViT). Results favor transformers, highlighting CNN limitations near pupil borders. Experimental validation on the PULPOS optical bench underscores the GCViT robustness. Trained solely on simulated data under varied SNR and D/r0 conditions, our approach enables to accurately close the AO loop in a real system and leave behind the reconstruction paradigm based on the interaction matrix. We demonstrate the high performance of the GCViT in closed loop obtaining a Strehl ratio over 0.6 for strong turbulence and nearly 0.95 for weak turbulence on the PULPOS optical bench.