Advisory Committee Chair
Advisory Committee Members
Purushotham V Bangalore
Mary Ellen Zvanut
Date of Award
Degree Name by School
Doctor of Philosophy (PhD) College of Arts and Sciences
The ability to tune optical responses in nanophotonics makes this class of materials useful for a broad range of applications, such as drug delivery systems, molecular sensing, green energy, and telecommunications, to name a few. Optical responses that offer strong resonance at small optical mode volume are often sought, requiring manipulation of many material properties i.e., refractive index and nanostructure shape. Machine learning has recently been employed to optimize photonic elements more efficiently. Machine learning optimization of optical responses suffers from reproducibility and accuracy issues due different nanostructures producing similar optical responses. To increase reliability and accuracy in recent neural network assisted optimizations, we explore an all-dielectric core-shell nanostructure, a system that provides both high complexity in the optical response, over 170K combinations, while keeping the number of tunable parameters low enough to obtain a ground truth. Our results illustrate the objective function choice and neural network simultaneously matter in the quality and reliability of the optimization. We then show it is possible to map similarity of optical responses in a latent space. We find employing unsupervised methods designed to deal with highly similar data allows for the mapping and grouping of features (mode, amplitude, and width) in optical responses regardless of the underlying material properties. This is demonstrated using tSNE embedding of various plasmonic and dielectric spherical nanostructures in differing sizes and environments. We obtain the embedding map via addition of synthetic data added to our initial data set we find the embedding organizes the optical responses by changes to the location of features in optical responses. We explore the time evolution of optical responses by generating over 640K different kinetic traces modeled on the varying decay recombination dynamics of rates within dielectric and semiconductor materials. Applying the tSNE algorithm to explore the latent space, our results illustrate the ability to understand similarities in the kinetic traces with various charge carrier dynamics (first, second, and third order recombination modes). We have found it is possible to increase the reliability and accuracy of the optimization of optical responses can be improved by tailoring the objective function and simulation quality used in the optimization. The latent space explorations demonstrate how unsupervised methods sort materials such allowing for easy to find and substitute different material properties for the same optical response. The latent space embedding also provides underlying insights to charge carrier recombination dynamics. These methods pave the way for the tSNE mapping of full temporal and spectral optical responses.
Hoxie, David J., "Machine Learning and Latent Space Representation of Optical Responses in Nanostructures and Thin Films" (2023). All ETDs from UAB. 35.
Available for download on Friday, May 09, 2025