Sequential Learning for Nanophotonic Inverse Design

Abstract

The development of nanophotonics has enabled many important and practical applications, ranging from photovoltaics and displays, to optical communications, imaging, and sensing. Inverse design deals with finding suitable nanostructures (a combination of material and structure dimensions) that can satisfy desired optical responses. Traditionally, optimization-based methods are widely used. However, they are not efficient as we need to restart the optimization process when facing new design targets. Recently, many machine learning-based methods have been proposed to solve these problems by learning from training datasets. However, these models can only be used to design for structure dimensions due to their fixed output dimension, which may limit the design performance. This thesis explores two new methods based on sequential learning (sequential decision process and conditional sequence generation) to tackle these inverse design challenges. With these methods, a series of novel photonic applications have been demonstrated. This dissertation starts with studying the design performance of two methods that do not use sequential learning: optimization-based methods and machine learning-based methods. Here, the particle swarm optimization is used to identify ideal multilayer thin film structures for structural color metrology. Detailed simulations and analysis are provided to understand this new type of structure. On the other hand, multiple popular machine-learning based methods (tandem networks, variational auto-encoders, generative adversarial networks) are benchmarked on two different nanophotonic inverse design tasks from three evaluation metrics: accuracy, diversity, and robustness, providing a guideline for new-coming researchers to select the model that can best suit their design criteria and fabrication considerations. Next, this dissertation will explore a new inverse design method based on sequential decision process. Using the previously proposed reinforcement learning algorithm called OML-PPO, a series of novel photonic applications have been designed and demonstrated. I fabricated and experimentally demonstrated the decorative chrome coatings with good environmental sustainability and multi-functionality, and colorful solar cells for building-integration with ~90% relative energy efficiency. The broadband transparent conductor and ultra-broad reflector are also designed and verified through simulation and analysis. On the other hand, from the physics perspective, we propose to leverage the sequential decision process to figure out a way for generalized design principles. Using this method, we successfully formulate a design principle to design reflective chrome spectrum step-by-step. Finally, a new model based on conditional sequence generation will be proposed: OptoGPT, which uses the decoder in transformer model and autoregressively generates designed structures as the output based on the input spectrum. To make the model suitable for multilayer structures, I propose the idea of “structure token” and “structure serialization” to convert the multilayer structure into a sequence. A series of ideas are further proposed to eliminate design barriers regarding optical targets, material selections, incident angle and polarization states as well as design constraints in fabrication, making inverse design tasks in multilayer structures approaching “solved”. On the other hand, OL-Transformer is proposed as a surrogate simulator for multilayer thin film structures, which uses the encoder in transformer model. Results show that OL-Transformer can speed up the simulation by ~6X times while still obtaining accurate spectrum predictions.