Nanophotonics use the interaction between light and subwavelength structures to design nanophotonic devices and to show unique optical, electromagnetic, and acoustic properties that natural materials do not have. However, this usually requires considerable expertise and a lot of time-consuming electromagnetic simulations. With the continuous development of artificial intelligence, people are turning to deep learning for designing nanophotonic devices. Deep learning models can continuously fit the correlation function between the input parameters and output, using models with weights and biases that can obtain results in milliseconds to seconds. In this paper, we use finite-difference time-domain for simulations, and we obtain the reflectance spectra from 2430 different structures. Based on these reflectance spectra data, we use neural networks for training, which can quickly predict unseen structural reflectance spectra. The effectiveness of this method is verified by comparing the predicted results to the simulation results. Almost all results maintain the main trend, the MSE of 94% predictions are below 10-3, all are below 10-2, and the MAE of 97% predictions are below 2 × 10-2. This approach can speed up device design and optimization, and provides reference for scientific researchers.
Keywords: deep learning; device design; nanophotonic; neural networks; reflectance spectra.