UCLA

Reaching the true potential of nanophotonic devices requires the broadband control of spectral and angular selectivity in the absorption and emission of electromagnetic waves. To this end, previously investigated design methods for nanophotonic structures and have encompassed both conventional forward and inverse optimization approaches as well as nascent machine learning (ML) strategies. While far more computationally efficient than optimization processes, ML-based methods that are capable of generating complex

nanophotonic structures are still ‘black boxes’ that lack explanations for their predictions. In that regard, we demonstrate that well-established deep learning architectures such as convolutional neural networks (CNN), which are highly proficient at forward design, can be explained to derive unique design insights by extracting the underlying physical relationships learned by network. To illustrate this capability, we trained a CNN model with 10,000 images of selective mid-infrared thermal emitters and their corresponding absorption spectra. The trained CNN predicted the spectra of new and unknown designs with over 95% accuracy. After training the CNN, we applied the Shapley Additive Explanations (SHAP) algorithm to the model to determine features that made positive or negative contributions towards specific spectral points, thereby informing which features to create or eliminate in order to meet a target spectrum. Using this strategy, we show that a starting electromagnetic metasurface design can be selectively manipulated to create target spectral properties. Our results reveal that the physical relationships between structure and spectra can be obtained, and new designs can be achieved, by exposing the valuable information hidden within a neural network.