Machine learning has revolutionized various fields in computer science, including computer vision (e.g. stable diffusion), and large language models (e.g. ChatGPT). Recently, there has been a growing focus on Scientific Machine Learning (SciML) due to its potential for scientific discovery. In optics and photonics, a range of data-driven approaches has been explored to solve complex optical problems.
Our research aims to develop physically reliable machine-learning models that efficiently learn the underlying physical relationship between structures and optical properties within the domain of nanophotonics. Specifically, we are currently developing 1) Equivariant neural networks, 2) Surrogate models for Maxwell’s equations solvers, and 3) Data-driven generative model mimicking fabrication processes.
A. Baucour, M. Kim, and J. Shin. Data-driven concurrent nanostructure optimization based on conditional generative adversarial networks. Nanophotonics. 11, 2865-2873 (2022).
M. Kim, A. Baucour, and J. Shin Equivariant neural networks for the design of optical nanostructures. In preparation.
M. Kim, A. Baucour, and J. Shin. Model-agnostic shift-equivariant downsampling. In preparation.