Machine Learning for Next-Generation Metamaterials

Abstract

This research focuses on prototyping next-generation electromagnetic metamaterials, which is often restricted by the complexity of their exotic geometry in the design, simulation, fabrication, and characterization processes. By utilizing machine learning methods, I aim to address challenges in each engineering step for the development of next-generation electromagnetic metamaterials. This dissertation tackles problems from the deep inverse design of electromagnetic metamaterials while mitigating the expensive training cost of deep surrogate models through analytical methods acceleration, deep active learning, and physics-informed training.To approach these problems, I begin by exploring the neural adjoint method that facilitates the inverse design of complex metamaterials, which is shown to have over five orders of magnitude acceleration over conventional numerical simulations. I also performed a benchmark to determine the most efficient deep neural network architecture for rapid forward modeling using deep surrogate model. To further increase the efficiency of data-driven surrogate models applied in previous work, I introduced deep active learning to mitigate the data bottleneck of surrogate models, established a fundamental physical bound to limit design space, and developed semi-analytical methods to speed up the data collection time. The dissertation also provides an overview of physics-informed learning to address two key issues: data bottleneck and physics interpretation. Since physics information can be treated as an add-on to all deep surrogate models mentioned in previous work, the discussion of physics-informed learning will focus on how the additional physics information helps improve machine learning in electromagnetic metamaterials.By examining the theory, design, characterization, and fabrication of metamaterials, this dissertation not only demonstrates the role of machine learning at each step of the metamaterials prototyping workflow but also sheds light on the rapid development of exotic metamaterials.