Latent space-based modeling for spectral prediction in generative photonics design.

Abstract

Electromagnetic (EM) metasurfaces consist of periodic structures of sub-wavelength dimensions that exhibit the ability to manipulate light for many novel applications. Calculating the optical response of a metasurface, typically performed using full-wave EM solvers in simulation, is a time- and resource-intensive operation. To accelerate computational design, machine learning-based surrogate models are increasingly investigated. The main challenge for these models is achieving data efficiency while preserving the diversity in possible shape design choices for the nanostructures. The most common degree of freedom in metasurface design is the pattern design of the base unit cell structure that is periodically repeated. In this work, a latent representation-based encoding of this base structure is investigated in the context of creating an optical response prediction machine learning model. The latent space-based model is found to be data efficient while retaining diversity in possible shapes of the nanostructures.