A deep learning approach to the forward prediction and inverse design of plasmonic metasurface structural color

Abstract

This report details a deep learning approach to the forward and inverse designs of plasmonic metasurface structural color. Here, optimized Deep Neural Network models are presented to enable the forward and inverse mapping between metamaterial structure and corresponding color. The forward model is capable of predicting color with >96% accuracy, with a 105 order of magnitude decrease in computational time when compared to finite-difference time-domain simulations used in conventional design workflows. An inverse model is trained using a tandem autoencoder, employing the pre-trained forward model. Here, the use of synthetic training data for self-learning is reported, which results in an ≈15% improvement in training accuracy. The tightly constrained inverse model allows for the instantaneous design of metasurfaces, given a desired color, with an accuracy of >86%, making it suitable for commercial use as well as the acceleration of photonics research.