Revealing the hidden capacity of artificial intelligence in nanoscience: physics-driven metric learning

Abstract

We present a new approach for design of novel loss functions and introduce an optimal similarity-metric design for machine-learning-based design and knowledge discovery in nanophotonics. Machine-learning algorithms estimate the input-output relation in a photonic nanostructure by minimizing a loss function. We show that careful selection (from the available loss functions) or design of novel loss functions that are optimized for specific tasks can considerably improve the performance of machine-learning algorithms for design and knowledge discovery in photonic nanostructures. We also discuss the limitations and inefficacies of conventional loss functions that are currently being used for machine learning algorithms.