Transformer-based learned dimensional collapse for silicon photonic inverse design

Abstract

We present a deep learning framework that revolutionizes photonic device design by collapsing costly 3D electromagnetic simulations into fast, accurate 2D representations. Our dual-stage, Transformer-based architecture combines a rapid factorization-cached 2D FDFD solver with a U-Transformer network to reconstruct full 3D fields, while a dedicated phase module preserves phase integrity. Trained on 16,000 3D-FDTD simulations of random silicon photonic devices, our model achieves over 99.1% field-matching accuracy and enables inverse design optimizations that are over 100 times faster than traditional methods. Designed devices exhibit less than 0.5 dB transmission mismatch and more than 90% structural similarity to 3D-FDTD results across the 1.5-1.6 mu m wavelength range. This scalable approach enables high-throughput, rapid, and practical design workflows for next-generation photonic components.