Dual-decoder transformer for inverse design of optical multilayer films

Optical multilayer films (OMFs) are essential for modern photonics, yet their inverse design faces scalability and interpretability challenges due to high-dimensional parameters. Conventional optimization struggles with complexity, while existing machine-learning methods suffer from combinatorial explosion when jointly encoding materials and thicknesses. We present OptoFormer, a transformer network that addresses these issues via a dual-decoder design: one predicts material sequences and the other predicts thickness ones. Additionally, a material-compatible dictionary is introduced to ensures physical feasibility of the predicted OMFs. The model achieves high spectral prediction accuracy and autonomously learns physical correlations like structural inverse symmetry without explicit constraints. In simulation, we demonstrate a 36-layer omnidirectional reflector, high-Q optical cavities, and a hyperspectral light-coding array. OptoFormer provides a scalable and data-driven framework for the on-demand inverse design of advanced photonic devices.