Electrically tunable nonlinear transmission characteristics of graphene hybrid plasmonic waveguides for neuromorphic activation functions

Optical neural networks (ONNs), distinguished from the von Neumann computing architecture, represent a high-performance novel computing paradigm that leverages the advantages of optical transmission including large bandwidth, low latency, low power consumption, and parallel signal processing. This approach holds promise for addressing the challenges of energy consumption and computational efficiency in current artificial intelligence technology development. In recent years, on-chip integrated optical neural networks have garnered extensive academic attention due to their potential to perform vector–matrix multiplication operations through linear optics and nonlinear activation functions via nonlinear optics. However, research on implementing nonlinear activation modules in the optical domain remains immature. We design and fabricate a graphene-based hybrid plasmonic waveguide electro-absorption modulator integrated on silicon photonic platform, where its dynamic range serves as the nonlinear activation function for photonic neurons. Concurrently, we construct feedforward neural networks for image classification tasks to test the nonlinear activation functionality, validating its feasibility and high prediction accuracy. This study proposes that utilizing surface plasmon polaritons (SPPs) to enhance light-matter interactions, thereby establishing an efficient, compact, and tunable nonlinear electro-optic module for photonic neurons. It provides critical device support for future large-scale, high-density photonic neuromorphic computing systems.