Harnessing ultra-broadband random laser for massively parallel physical random bit generation and optical decision making

The relentless growth of data-intensive computing and artificial intelligence demands physical random number generators with unprecedented speed and parallelism to accelerate reinforcement learning and other stochastic decision-making algorithms. Despite recent advances in microring resonators, chaotic laser arrays, Fabry–Pérot semiconductor lasers, and random fiber lasers, existing schemes still suffer from inherent trade-offs between channel count and output power. Here, we demonstrate a high-power, ultra-broadband, spectrotemporally uncorrelated multi-wavelength Brillouin random fiber laser for massively parallel physical random bit generation and optical decision making. A 100-nm broadband optical frequency comb, comprising over 441 lines with a 30-dB peak power roll-off, is obtained through the synergistic enhancement of cascaded stimulated Brillouin scattering and quasi-phase-matched four-wave mixing. Notably, dispersion-shifted fiber further enhances the four-wave mixing effect, increasing the number of spectral lines. Rayleigh scattering and randomly distributed feedback break the correlations between wavelength channels, enabling direct and parallel random bit extraction without complex post-processing. Random bit sequences from 45 representative wavelength channels were experimentally characterized and verified using the NIST SP 800-22 statistical test suite. Operating at a 40 GSa/s sampling rate, each channel achieves 160 Gbps, yielding a theoretical maximum rate of 70.56 Tbps (160 Gbps × 441) for spectral lines within the 30-dB peak power roll-off. Furthermore, the proposed system experimentally supports four simultaneously acquired channels, while its scalability to at least 256 multi-armed bandit optimization problems, a representative class of stateless reinforcement-learning tasks, is demonstrated through offline reconstruction from independently acquired channel traces, providing a scalable photonic platform for decision-making acceleration.