Mesh effect in diffractive optical processor

A diffractive optical processor (DOP), built from multiple cascaded, programmable free-space diffractive optical components, is capable of performing a broad range of machine learning (ML) tasks in high-throughput and energy-efficient manners. However, transferring simulation-trained designs to hardware with high fidelity remains challenging because differentiable surrogate models often simplify underlying physics for efficient computing, leading to systematic performance degradation upon deployment. Here, we identify a previously ignored factor contributing to this simulation-hardware mismatch, namely the mesh effect arising from how physical hardware features are discretized in differentiable models. This mesh effect becomes pronounced when the unit size of the programmable elements is not commensurate with the operation wavelength. We develop an analytic description of the mesh effect and validate it through numerical studies and experiments on two representative ML tasks, including handwritten-digit classification and solving the two-dimensional Darcy-flow partial differential equation. We show that the common practice of directly using the physical pixel size as the simulation discretization degrades experimental accuracy. To address this, we develop a fully differentiable fine-meshing strategy that refines the computational grids in simulation models, improving deployment fidelity and reducing discrepancy between simulations and experiments. These results highlight discretization-aware training as a practical route to robust, scalable optical learning hardware.