Large-scale integrated optoelectronic chaos for machine learning acceleration

Chaos finds widespread use in modern machine learning, yet its implementations in traditional nonlinear circuits have encountered speed bottlenecks due to ever-expanding computational needs. The optical chaotic source offers an attractive alternative, combining ultra-wideband capabilities, inherent nonlinearity, and massive parallelism. However, existing photonic schemes typically trade off between the single-channel throughput and multi-channel scalability, preventing effective learning acceleration of large-scale tasks. Here, an integrated microcomb-optoelectronic chaos engine (iMOCE) is demonstrated. By using a microcomb to optoelectronic nonlinear cavity, massively parallel channels with a 6-dB bandwidth of 25 GHz per channel are achieved, representing a two-order-of-magnitude improvement over previous approaches using microcombs. The proposed iMOCE achieves a total random-bit generation rate of 32.768 Tbps (1.024 Tbps per channel), unparalleled by existing optical chaotic sources. The chip is fabricated in a commercial foundry and is compatible with wafer-scale production, ensuring manufacturability and scalability. To showcase its learning acceleration capability, the iMOCE is applied to four learning accelerator tasks, including the multi-armed bandit problem, connect-3 game, traveling-salesman solving, and electrocardiogram trace recognition task. Compared with MCU/GPU baselines, iMOCE reduces per-inference time by about two orders of magnitude across tasks. By bridging wafer-scale integrated photonics with probabilistic computing, our iMOCE establishes a scalable, massively parallel chaos primitive for accelerating learning, decision-making, and combinatorial optimization.