Modal CNN reconstruction for an unmodulated pyramid wavefront sensor in SAXO+

Abstract

The unmodulated pyramid wavefront sensor offers high sensitivity but requires accurate reconstruction of its nonlinear response. We present a convolutional neural network that directly estimates deformable-mirror modal coefficients from the four pyramid pupil images. Unlike image-to-grid approaches, the modal output can be adapted to different deformable-mirror geometries by changing the modal basis and output dimension. The method was evaluated using end-to-end COMPASS simulations of the SAXO+ two-stage adaptive-optics system. Random residual phase screens were used to generate training data, and the learned reconstructor was assessed in open and closed loop under three representative observing conditions. In open loop, the modal CNN reduced the reconstruction error substantially relative to the calibrated linear reconstructor. In closed loop, the modal CNN alone kept the loop stable in the two bright cases and improved the mean coronagraphic contrast between 4 and 6λ/D by approximately 20–30% relative to the modulated linear reference. In the photon-limited case, the standalone CNN failed to close the loop; a lightweight dense head fusing the CNN and linear estimates restored stability and still improved contrast by 33%, while leaving performance almost unchanged in the bright cases. The modal CNN performed at least as well as the previous U-Net-based approach, while providing greater flexibility for future extreme-adaptive-optics systems and ELT instruments.