Tunable Nonlinear Activation Functions Enabled by WO₃ Films for Adaptive Diffractive Deep Neural Networks

Abstract Conventional electronics face intrinsic bandwidth and power constraints in deep learning, fueling the pursuit of optical computing's parallel processing and energy efficiency. A critical roadblock for optical neural networks (ONNs) lies in the missing trainable‐tunable nonlinear activation mechanisms required for modeling complex data correlations and cross‐task generalization. Here, this limitation is overcome through engineered tungsten trioxide (WO₃) thin films with dynamically controllable nonlinearity. Z‐scan measurements demonstrate stoichiometry‐dependent nonlinear responses modulated via photogenerated coloration, enabling precise control of optical nonlinearities. The films' non‐volatile memory effects permit implementation of adaptive in‐memory computing architectures where activation functions are trainable and task‐specifically optimized. This reconfigurable nonlinearity framework enhances both learning capability and generalization performance in ONNs, while achieving high‐performance THz‐scale response speeds (1 THz) for real‐time adaptive computing. Integrated as programmable activation layers in optical diffraction networks, the system demonstrates classification accuracy improvements of 4.62% (MNIST), 3.29% (Fashion‐MNIST), 13.53% (KMNIST), and 12.2% (CIFAR‐10). The synergistic combination of non‐volatile tunability, ultrafast reconfiguration, and stoichiometric control positions WO₃ thin films as a disruptive materials platform, resolving long‐standing activation function limitations in photonic neuromorphic systems while enabling adaptive high‐performance optical processors through on‐demand nonlinear engineering.