Abstract
ABSTRACT
All‐optical diffractive deep neural networks (D
2
NNs) are developed as the most important architecture for all‐optical intelligent computing, owing to their low power consumption and parallel processing at the speed of light. Currently, a significant challenge in D
2
NNs is exploring new degrees of freedom to expand channel capacity between multiple diffractive layers. Here, we propose a sequenceable scheme by reconfiguring the stacking sequence and interlayer distance to enable high‐capacity processing capability. Experimentally, we have simultaneously achieved 6 classifiers and 6 trigonometric calculators by exchanging the spatial sequence and interlayer distance among the same three‐layer metasurfaces. Moreover, we demonstrate a high‐security information transmission framework using image transformation and a shifting operation. To improve the robustness, we design the polarization‐insensitive metasurfaces to break the barrier of polarization‐dependent multi‐task diffractive processing. The demonstrated phenomena showcase polarization‐insensitive properties, excellent flexibility, and robustness, which provide a paradigm for next‐generation multi‐functional, high‐capacity all‐optical computing and versatile applications.