Predicting Laser-Induced Colors of Random Plasmonic Metasurfaces and Optimizing Image Multiplexing Using Deep Learning

Structural colors of plasmonic metasurfaces have been promised to a strong technological impact thanks to their high brightness, durability, and dichroic properties. However, fabricating metasurfaces whose spatial distribution must be customized at each implementation and over large areas is still a challenge. Since the demonstration of printed image multiplexing on quasi-random plasmonic metasurfaces, laser processing appears as a promising technology to reach the right level of accuracy and versatility. The main limit comes from the absence of physical models to predict the optical properties that can emerge from the laser processing of metasurfaces in which random metallic nanostructures are characterized by their statistical properties. Here, we demonstrate that deep neural networks trained from experimental data can predict the spectra and colors of laser-induced plasmonic metasurfaces in various observation modes. With thousands of experimental data, produced in a rapid and efficient way, the training accuracy is better than the perceptual just noticeable change. This accuracy enables the use of the predicted continuous color charts to find solutions for printing multiplexed images. Our deep learning approach is validated by an experimental demonstration of laser-induced two-image multiplexing. This approach greatly improves the performance of the laser-processing technology for both printing color images and finding optimized parameters for multiplexing. The article also provides a simple mining algorithm for implementing multiplexing with multiple observation modes and colors from any printing technology. This study can improve the optimization of laser processes for high-end applications in security, entertainment, or data storage.