Photonic integrated circuits offer a promising platform for ultra-low-energy artificial intelligence computation, but their scalability is limited by the accumulation of analog hardware imperfections such as thermal crosstalk, phase noise, and digital-to-analog precision limits. Here we present a hardware–software co-designed framework enabling deep photonic neural computation on a single metasurface. A compact physics-informed neural network (PINN) trained through teacher–student distillation actively compensates thermal and quantization errors in programmable photonic meshes, restoring classification accuracy from 41.73% to 95.27% and achieving an extinction ratio of 63.96 dB. Building on this calibrated substrate, we introduce the Single-Layer Photonic Computing (SLiM) architecture that implements deep neural networks through temporal recurrence rather than spatial cascades. Using dual-wavelength differential detection to enable signed optical weights, the system emulates a 100-layer neural network on a single metasurface while achieving 95.20% MNIST classification accuracy and operating at 16.2 fJ per multiply-accumulate operation. These results demonstrate a scalable pathway toward fault-tolerant and energy-efficient photonic neural accelerators.
Ramisetty et al. (Sun,) studied this question.