All‐Optical Nonlinear Activation Functions Based on Photonic Crystal Microcavity

ABSTRACT The growing computational demand of artificial intelligence (AI) calls for energy‐efficient computing architectures. Diffractive deep neural networks () offer a promising route for optical computing, but their lack of nonlinear activation functions limits their representation capability. Here, we demonstrate a passive all‐optical Leak‐ReLU nonlinear activation function based on a 1D photonic crystal microcavity. By exploiting resonantly enhanced thermo‐optic nonlinearity, the cavity exhibits a temperature‐induced resonance blueshift that produces intensity‐dependent nonlinear transmission with tunable activation thresholds through wavelength selection. Integrating the experimentally characterized nonlinear activation function into a improves the classification accuracy by 1.79% on Fashion‐MNIST compared with a two‐layer linear and accelerates convergence by 2.6 times on MNIST. This work provides a compact and passive approach for implementing nonlinear optical neural networks and advances low‐power neuromorphic optical computing.