All‐Optical Meta‐Convolutional Neural Networks for Reconfigurable Image Identification

ABSTRACT Convolutional neural networks (CNNs) dominate visual computing, but their electronic implementations face fundamental bottlenecks in speed and energy efficiency. Although photonic alternatives offer a promising route, all‐optical image recognition using CNNs remains challenging in experiments due to the difficulty of implementing effective weak‐light nonlinearities. Here, we demonstrate a streamlined all‐optical CNN architecture with only physically guided, fixed convolutional and trainable fully‐connected layers (FCLs), which circumvents explicit, energy‐intensive nonlinear layers. Our design leverages a fixed multi‐kernel metasurface operator to perform 24 directional‐differentiation convolutions, thus enhancing feature orthogonality and suppressing crosstalk. This preprocessing enables the system to exploit the weak nonlinearity produced by square‐law detection, thereby achieving high experimental classification accuracy: 86.8% on MNIST and 94.8% on ten‐class gesture recognition. The system exhibits high reconfigurability and robustness. Its all‐optical inference core achieves an equivalent computational throughput of 1.13 × 10 5 TOPS and an energy efficiency of 1.88 × 10 3 TOPS/W—surpassing current CNN implementations. At the system level, it achieves an end‐to‐end throughput of 871 TOPS and an energy efficiency of 4.97 TOPS/W, still exceeding advanced all‐electronic devices and paving the way for ultralow‐latency intelligent hardware.