Intrinsic Volatile Photonic Memristors for Optical Neuromorphic Computing

ABSTRACT Inspired by electronic memristors' success in artificial neural networks, photonic memristors are sought for next‐generation optical neuromorphic computing. However, current volatile photonic memristors require electrical feedback, increasing complexity and reducing efficiency. Here, we propose a universal theoretical framework for intrinsic volatile photonic memristors and report three implementations of this scheme based on nonlinear optical resonant cavities. Utilizing intrinsic feedback, photonic memristive behaviors with optical nonlinearity and memory dynamics are observed experimentally. On this basis, we further demonstrate photonic memristive reservoir computing (PMRC) systems for image and time‐series classification. These PMRCs operate at rates from Hz to GHz—up to three orders of magnitude faster than electronic memristors. Notably, the laser‐based scheme can theoretically reach TOPS/W‐level energy efficiency. Our approach surpasses the current limits of electronic computing via optical parallel computing with spatial and wavelength multiplexing.