Metric Learning: Harnessing the Power of Machine Learning in Nanophotonics

Abstract

We present a novel metric-learning approach based on combined triplet loss and mean-squared error for providing more functionality (e.g., more effective similarity measures) to the machine-learning algorithms used for the knowledge discovery and inverse design of nanophotonic structures compared to commonly used mean-squared error and mean-absolute error. We demonstrate the main shortcoming of the existing metrics (or loss functions) in mapping the nanophotonic responses into lower-dimensional spaces in keeping similar responses close to each other. We show how a systematic metric-learning paradigm can resolve this issue and provide physically interpretable mappings of the nanophotonic responses while facilitating the visualization. The presented metric-learning paradigm can be combined with almost all existing machine-learning and deep-learning approaches for the investigation of nanophotonic structures. Thus, the results of this paper can have a transformative impact on using machine learning and deep learning for knowledge discovery and inverse design in nanophotonics.