Synergizing Nanophotonics and Artificial Intelligence: From Inverse Design to Intelligent Optical Frontends

Overview

The research advances two complementary paths: AI-facilitated nanophotonics inverse design and the deployment of nanophotonics as optical computation frontends.

The dissertation introduces a hybrid optimization framework that combines deep learning with heuristic search and incorporates physics-informed loss functions to enhance robustness against fabrication variations and mechanical deformations. The design framework demonstrates broadband polarization-splitting metasurface on flexible substrates with high efficiency. The dissertation further explores an optical storage paradigm, which trains a photonics foundation model and encodes the learned geometry–spectrum representations as additional storage dimensions, beyond standard multiplexing strategies.

This dissertation further demonstrates metasurface as a computer vision frontend, which extracts spectral information of the scene and projects it onto the sensor plane. Applications include real-time snapshot hyperspectral imaging (HSI), hyperspectral video understanding at Tb/s data rates, and dopamine sensing down to 10−8 mM. Collectively, this dissertation establishes a blueprint for the next generation of intelligent optical systems.