Lorenzo Lupano
Pronouns: He/Him
Research Mentor(s): L. Jay Guo
Research Mentor School/College/Department: Electrical Engineering and Computer Science / Engineering
Program:
Authors: Charles Sager, Lorenzo Lupano, Taigao Ma, L. Jay Guo
Session: Session 7: 4:40 pm – 5:30 pm
Poster: 98
Abstract
Thin-film structures are structures comprising various layers of different materials with thickness in the order of magnitude of nanometers, and they are widely used in filters, absorbers, photovoltaic cooling, and various other applications. Inverse design is a critical technique that can enable these applications, which seeks to use algorithms to optimize the structure and achieve desired optical performance. Recently, many machine learning (ML)-based methods have been proposed to deal with inverse design, however, it is still unknown which model performs best on this problem. Here, we aim to solve this problem by establishing a benchmark work to evaluate and compare multiple different machine learning models applied to multilayer thin-film optics inverse design, including Variational Auto-Encoders, Generative Adversarial Networks, Mixture Density Networks, and Tandem Networks. The evaluation will focus on three critical criteria, including accuracy, diversity, and robustness. To do this, we first collect the training dataset through transfer matrix method simulations and then implement and train these ML models. Finally, we analyze the results to identify the optimal model that exhibits the best performance using different evaluation metrics.
