Abstract Optical neural networks (ONNs) offer a route to low-latency, energy-efficient AI, but scaling them to modern model sizes is constrained by two practical bottlenecks: training large ONNs is computationally prohibitive, and implementing or tuning millions of optical components is highly sensitive to fabrication imperfections and alignment errors. Here we report a metasurface-based optical learning machine that bypasses these barriers by operating in an ultra-wide regime. We use an optical metasurface with 41 million optical parameters to form an ultra-wide NN, and show that, at this scale, a fixed, untrained metasurface can closely approximate a fully trained one. Residual mismatch is compensated by a compact digital backend with only 10² 10 2 – 10⁴ 10 4 trainable parameters, enabling task adaptation without retraining or re-fabricating the metasurface. Experimentally, the resulting hybrid system delivers scalable machine vision across six tasks, achieving accuracy competitive with state-of-the-art models including ResNet and Vision Transformer, while remaining tolerant to fabrication imperfections and misalignment. Our approach is highly scalable, tolerant to fabrication errors, and operates under both coherent and incoherent illumination, thus providing a practical pathway to large-scale, high-performance optical AI computing.
Luo et al. (Wed,) studied this question.