Machine learning driven design and fabrication of photonic devices with engineered dispersion

Abstract

This thesis combines optical physics and advanced machine learning to tackle challenges in thin film and waveguide design. By engineering material dispersion and optimizing optical systems, it bridges traditional physics methods with modern data-driven strategies, using simulations and artificial intelligence to enhance the design of complex optical structures.

Initially, a Python Transfer Matrix Method (TMM) is developed to simulate and validate the optical properties of Distributed Bragg Reflectors, bandpass filters, and Epsilon-Near-Zero materials. Planar disordered waveguides are introduced to study random scattering and Kerr effects in both linear and nonlinear regimes, with Finite-Difference Time-Domain simulations generating comprehensive datasets for machine learning model training.

Various neural network architectures—Multilayer Perceptrons, Convolutional Neural Networks, and Transformers—are explored to recognize patterns in ultrafast laser characterization, thin film stacking, and disordered waveguides. The thesis then applies Reinforcement Learning (RL) to optical design problems, illustrating fundamental RL concepts through a custom maze game and using advanced algorithms for optimization tasks.

Experimental results examine ENZ apertures made of thin silver and silica layers, forming anisotropic effective media with enhanced transmission. Ultrafast laser characterization with neural networks improves the accuracy of Second- and Third-Harmonic Generation measurements, reducing the need for repetitive experiments.

A multi-agent RL system is developed for the inverse design of thin films, optimizing parameters, material selection, and layer thickness. A novel reward function employing polynomial control allows tailored spectral responses, while regional constraints and penalties for unwanted features improve filter and reflector designs.

In disordered waveguide design, Generative Adversarial Networks are used for inverse design tasks. Three architectures are evaluated for generating waveguide geometries based on desired optical responses, addressing the complexities of sparse and disordered systems.

Finally, Large Language Model integrates with the RL-TMM framework using Retrieval-Augmented Generation, enhancing the system's reasoning and prediction capabilities for optical responses.

Date of Award	30 Jun 2025
Original language	English
Awarding Institution	University of St Andrews
Supervisor	Andrea Di Falco (Supervisor)