Title: Transformer neural networks for closed-loop adaptive optics using non-modulated pyramid wavefront sensors

Abstract: The Pyramid Wavefront Sensor (PyWFS) is highly nonlinear and requires the use of beam modulation to successfully close an AO loop under varying atmospheric turbulence conditions, at the expense of a loss in sensitivity. In this work we train, analyse, and compare the use of deep neural networks (NNs) as non-linear estimators for the non-modulated PyWFS, identifying the most suitable NN architecture for reliable closed-loop AO. We develop a novel training strategy for NNs that seeks to accommodate for changes in residual statistics between open and closed-loop, plus the addition of noise for robustness purposes. Through simulations, we test and compare several deep NNs, from classical to new convolutional neural networks (CNNs), plus a state-of-the-art transformer neural network (TNN, Global Context Visual Transformer, GCViT), first in open-loop and then in closed-loop. Using open-loop simulated data, we observe that a TNN (GCViT) largely surpasses any CNN in estimation accuracy in a wide range of turbulence conditions. Also, the TNN performs better in simulated closed-loop than CNNs, avoiding estimation issues at the pupil borders. When closing the loop at strong turbulence and low noise, the TNN using non-modulated PyWFS data is able to close the loop similar to a PyWFS with $12\lambda/D$ of modulation. When raising the noise only the TNN is able to close the loop, while the standard linear reconstructor fails, even with modulation. Using the GCViT, we close a real AO loop in the optical bench achieving a Strehl ratio between 0.28 and 0.77 for turbulence conditions ranging from 6cm to 20cm, respectively. In conclusion, we demonstrate that a TNN is the most suitable architecture to extend the dynamic range without sacrificing sensitivity for a non-modulated PyWFS. It opens the path for using non-modulated Pyramid WFSs under an unprecedented range of atmospheric and noise conditions.