Inductively Scalable, Single-Step Neural Surrogates for Wave-Scattering Inverse Problems

Details

Full-wave electromagnetic simulation is a major bottleneck in photonic chip and optical device design: FDTD solvers can take hours per evaluation, making iterative inverse design impractical. Neural surrogate models promise to replace these solvers, but prior single-step (non-recurrent) surrogates have only scaled to tens of variables. This paper from Charles Dove and Laura Waller at UC Berkeley breaks that barrier using a training strategy called adversarial curriculum learning: instead of random sampling, a parallel gradient-ascent process continuously searches for configurations where the surrogate disagrees with the ground-truth FDTD solver, feeding those hard cases back into training. Combined with source normalization and an evolving replay dataset, this trains a surrogate for 2D wave scattering with up to 41,772 controllable variables. The key result is inductive generalization: without retraining, the surrogate scales to over 3 million variables — a 73.8x increase — and achieves 1.29x to 26.5x speedups over FDTD on large forward simulations and freeform inverse design tasks (beam splitters, GRIN lenses up to 98 wavelengths wide). Performance is comparable to or better than FDTD-based designs. The approach could meaningfully accelerate the photonic design loop, particularly for AI accelerator chips where optical interconnects are increasingly relevant.