OptoMPGD: joint encoder-decoder optimization for coded diffraction phase retrieval with state space models

Recovering complex optical fields from intensity-only measurements is a fundamental, ill-posed inverse problem in computational imaging. Classical iterative methods converge slowly, while learned approaches rely on either local convolutional kernels that miss diffraction’s global energy redistribution or transformers with quadratic computational cost. We present OptoMPGD, a hybrid framework that jointly optimizes a programmable optical encoder and a physics-based deep unfolding decoder for coded diffraction phase retrieval. The decoder unfolds proximal gradient descent with a state space model-based learned prior whose efficient long-range aggregation capability is well matched to the globally distributed dependencies induced by diffraction, providing a global receptive field at linear computational cost. Furthermore, we propose a complex-domain transformer to calibrate residual hardware aberrations during physical deployment. Experiments on a custom 4 f bench show that OptoMPGD surpasses prior methods while requiring fewer coded measurements.