Nonlinear Dimensionality Reduction for Low Data Regimes in Photonics Design

Abstract

Efficient exploration of high-dimensional parameter space is essential in modern photonic component design. Linear dimensionality reduction such as principal component analysis has proven useful in identifying lower dimensional subspace of interest in several design problems. Yet such subspaces often exhibit curvature reflecting nonlinear relationships between design parameters. For such systems linear dimensionality reduction methods can be suboptimal. We discuss how an appropriate architecture for an autoencoder neural network along with a numerically robust initialization, show improved performance compared to linear methods even in low data regimes, which are typical for photonic design problems.