Physical Information-Embedded Deep Learning for Forward Prediction and Inverse Design of Nanophotonic Devices

Nanophotonic devices have enabled electromagnetic manipulation with unprecedented accuracy and functionalities. With the increased functionality requirements of such devices and the progress of fabrication technologies, trends toward miniaturization and multi-functionality have emerged, making forward response prediction and inverse design increasingly complicated. In this paper, a physical information-embedded deep learning model for flexible forward prediction and multifunctional inverse design is proposed. Besides the overwhelming advantage of computing efficiency over conventional iterative optimization methods, the proposed method exhibits strong generalization on different wavelengths and polarizations by embedding the corresponding physical information into the model through effective sparse parameter method (SPM). For the forward prediction, responses can be predicted among the 1300-1550 nm wavelength band and two polarizations (TM/TE) with high accuracy. For the inverse design, with the assistance of forward prediction models and broadband training methods, multi-dimensional information for the target can be considered simultaneously to retrieve the satisfactory desired designs topologies in an unsupervised way. The multi-functionality of inverse design is demonstrated by implementing wavelength demultiplexers with small crosstalk and power splitters with arbitrary splitting ratios on a 250 nm broadband wavelength in a compact footprint of 2 × 2 μm ² without additional heavy training of the model.