Loading...
Machine learning surrogate models for electromagnetic simulation
Hranicky, Parker A.
Hranicky, Parker A.
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2025
Date Submitted
Collections
Files
Research Projects
Organizational Units
Journal Issue
Embargo Expires
Abstract
The advent of simulation technology has transformed engineering and design, allowing for in-depth analysis and optimization. However, running numerical simulations, particularly over large domains, demands considerable computational resources and time. In the era of artificial intelligence, surrogate models offer a promising solution by providing data-driven approximations that emulate the behavior of these simulations. Such models can accelerate simulation processes by orders of magnitude, reducing runtimes from hours or days to mere minutes or seconds. This efficiency enables engineers to quickly explore and iterate through design spaces, fostering rapid innovation.
This research focuses on developing a surrogate model for EMC Plus, a 3D Finite Difference Time Domain (FDTD) solver used to resolve Maxwell’s equations. The goal is to predict the Shielding Effectiveness (SE), a measure of an enclosure’s ability to shield against electromagnetic effects. The surrogate model takes as input the enclosure’s geometry, material properties, and the illumination conditions of an incident plane wave.
To achieve this, data was generated from EMC Plus simulations of simple enclosures with apertures and constant plane wave excitations. An initial study utilized a Fourier Neural Operator (FNO) architecture, demonstrating promising results. However, to address the complexity of real-world scenarios, a graph neural network (GNN) was introduced. This GNN employs modified edge convolutions to encode features such as geometry and material properties, followed by a pooling layer and fully connected layers to predict SE values across a range of frequencies. For the GNN study, additional data was generated, incorporating more complex geometries, varied aperture configurations, and diverse excitation conditions to enhance the model's applicability to practical scenarios. This approach not only produces promising results but establishes a scalable architecture for building surrogate models for real world applications.
Associated Publications
Rights
Copyright of the original work is retained by the author.