Can the success of digital super-resolution networks be transferred to passive all-optical systems?

Abstract

The deep learning revolution has increased the demand for computational resources, driving interest in efficient alternatives like all-optical diffractive neural networks (AODNNs). These systems operate at the speed of light without consuming external energy, making them an attractive platform for energy-efficient computation. One task that could greatly benefit from an all-optical implementation is spatial super-resolution. This would allow overcoming the fundamental resolution limitation of conventional optical systems, dictated by their numerical aperture. Here, we examine whether the success of digital super-resolution networks can be replicated with AODNNs considering networks with phase-only nonlinearities. We find that while promising, super-resolution AODNNs face two key physical challenges: (i) a tradeoff between reconstruction fidelity and energy preservation along the optical path and (ii) a limited dynamic range of input intensities that can be effectively processed. These findings offer a first step toward understanding and addressing the design constraints of passive, all-optical super-resolution systems.