A CNN encoder for modal phase reconstruction in Adaptive Optics systems

Pyramid wavefront sensing (pWFS) offers some of the best sensitivity for adaptive optics, but suffers from strong non-linearity. Modulation can extend linear range at the expense of sensitivity. Operating the pyramid without modulation is therefore attractive, but remains challenging. We develop a non-linear CNN reconstructor for pyramid wavefront sensing that maps pWFS images directly to a modal phase representation, and we assess simple hybrid methods combining it with a standard linear least-squares (LS) reconstructor. Using the end-to-end COMPASS simulator, we generate a large open-loop dataset spanning wide ranges of RMS and power spectra to train a compact CNN encoder. We then compare the CNN against an LS baseline and evaluate hybrid schemes in closed-loop simulations over a grid of guide-star magnitudes and Fried parameters, reporting long-exposure Strehl at lambda=1.6 um with controller gain re-optimized per method and bin. We also report preliminary offline bench tests on SCExAO. The CNN reduces open-loop reconstruction error and, in closed loop, outperforms LS across most (magnitude, r_0) conditions, with the largest gains for faint stars where it often closes the loop while LS does not. In very strong turbulence, LS can exceed the CNN; in these cases the hybrid methods are necessary to surpass LS, with a second-stage NN performing best. Inference is real-time capable and the hybrid overhead is negligible. Bench snapshots on SCExAO show successful correction of small static/slow perturbations, with instability for stronger/faster cases. A compact CNN can be trained to perform modal reconstruction for a non-modulated pWFS and improves performance over a classical linear reconstructor in most regimes. When the CNN alone is not optimal, simple hybrid methods achieve the best performance, suggesting a practical way to exploit the pyramid’s sensitivity without modulation.