Large-scale integrated optoelectronic chaos for machine learning acceleration

Abstract

Abstract Chaos has emerged as a useful resource for machine learning, yet traditional nonlinear circuits face speed bottlenecks. Optical chaos sources offer an attractive alternative with ultra-wideband operation and massive parallelism, but existing schemes must trade single-channel throughput against multi-channel scalability. Here, we demonstrate an integrated microcomb-optoelectronic chaos engine (iMOCE). By driving an optoelectronic nonlinear cavity with a chaotic microcomb, the iMOCE generates massively parallel chaos with a 6-dB bandwidth of 25 GHz per channel, representing a two-order-of-magnitude improvement over previous microcomb-based approaches. The system delivers a total random-bit generation rate of 32.768 Tbps and accelerates four representative tasks. Compared with MCU/GPU baselines, it reduces per-inference time by about two orders of magnitude. These results establish iMOCE as a scalable, massively parallel chaos primitive for machine learning acceleration.