Zeroth-Order On-Chip Training of Physics-Informed Neural Networks on Photonic Accelerators
Physics-informed neural networks (PINNs) can be trained directly on a photonic micro-ring-resonator accelerator without backpropagation by using zeroth-order optimization, which estimates parameter updates from forward inferences alone. This removes the need for a differentiable device model, pre-calibration, and backward optical hardware, and lets fabrication error and analog noise be absorbed during training, enabling an on-chip PDE solver whose accuracy is bounded by analog bit precision.
2501.00742
An optical neural network implements a neural network's linear layers in photonic hardware; here a 1x4 micro-ring resonator (MRR) weight bank multiplies wavelength-encoded inputs by tunable optical w…