Self-optimizing multichannel optical computing

Optical computing offers a promising route towards ultrafast and energy-efficient alternatives to conventional digital processors, though practical end-to-end realization remains a significant challenge. Among several architectural limitations, most current implementations are confined to single-channel processing, which underutilizes the inherent information capacity of light. Here we demonstrate a self-optimizing multichannel optical computing architecture based on multi-plane light conversion that natively processes RGB images and structured numerical data entirely within the optical domain. We introduce two complementary optimization strategies that enable autonomous adaptation without the need for differentiable forward models: Bayesian optimization to refine channel mixing at the input level, and a hardware-in-the-loop protocol based on self-organized criticality to navigate the high-dimensional phase landscapes. Across medical imaging, natural image classification, and regression tasks, multichannel optical preprocessing with random phase masks improves accuracy by 25–45 percentage points over raw pixel baselines, with RGB encoding consistently outperforming grayscale by 6–11 percentage points. Furthermore, our self-optimization strategies provided additional gains of 7–12 percentage points. These demonstrated capabilities establish self-optimizing multichannel optical computing as a successful proof-of-concept, paving the way for future efforts to address the system-level challenges of optical machine learning. Fatma Nur Kilinc introduce a multichannel optical computer that self-optimizes, using Bayesian tuning of input channels and a self-organized-criticality protocol to reshape its own hardware, with no differentiable model required. Multichannel processing alone lifts accuracy 25-45 points, and self-optimization adds 7-12 more.