In‐Situ Trained Microring‐Based Neural Networks for Scalable and Robust Photonic Computing

Abstract Photonic computing offers high speed, large bandwidth, and ultra‐low power consumption, making it a promising alternative to traditional electronic processors, especially for matrix‐vector multiplication (MVM) and convolution tasks. Among photonic architectures, microring resonator (MRR)‐based optical neural networks (ONNs) are attractive due to their compact footprint and wavelength‐division multiplexing. However, MRRs are highly sensitive to environmental disturbances and crosstalk, limiting computational accuracy. While in‐situ training has emerged as an effective method to enhance system performance by adapting weights during computation, it requires real‐valued bidirectional processing to support backpropagation—a significant challenge for noncoherent MRR‐based systems. Here, an in‐situ trained MRR‐based ONN that overcomes these limitations through real‐valued bidirectional optical computing is demonstrated. By integrating multiwavelength multiplexing with on‐chip forward and backward propagation, this architecture enables physical parameter updates via optical backpropagation without lookup table dependency. Experimental validation perfectly matches digital computing results and shows a 13.3% accuracy improvement over conventional MRR weight banks in classification tasks, with sustained precision under prolonged operation. Systematic analysis confirms the architecture's robustness against thermo‐optic crosstalk and environmental variations. This work establishes a pathway toward scalable, disturbance‐resilient photonic computing for next‐generation artificial intelligence hardware.