Machine Learning For Electromagnetic Metamaterial Design
Open Access
- Author:
- O'Connor, Philip
- Area of Honors:
- Electrical Engineering
- Degree:
- Bachelor of Science
- Document Type:
- Thesis
- Thesis Supervisors:
- Douglas Henry Werner, Thesis Supervisor
Julio Urbina, Thesis Honors Advisor - Keywords:
- Machine Learning
Metamaterial Antennas
Computational Electromagnetics
Neural Networks - Abstract:
- Computational electromagnetics has found success in the design and simulation of metamaterial devices for many applications. This work explores machine learning as a tool for computationally efficiently modeling metamaterial devices. Conventional methods have proven effective, though computationally expensive and slow, for analyzing metamaterials. By using a dataset of geometric metamaterials generated by traditional computational electromagnetic methods, a neural network can be trained to generate and analyze new metamaterial designs more quickly than previous methods. This work develops and presents a method for parameterizing a dataset of geometric metamaterial patterns and electromagnetic properties for a neural network to generalize these electromagnetic properties to new patterns. The motivation for training a neural network on this computationally expensive process of analyzing the electromagnetic properties of geometric metamaterials is to ultimately utilize the quick computation of neural networks. Through parallelized GPU computing, thousands of metamaterials can be simulated by the trained neural network in the same time conventional methods can analyze a single design.