Learning-based physics-enhanced modeling and inline material characterization for electromagnetics
Doctoral thesis, 2026
This thesis explores data-driven modeling and inline material characterization methods for electromagnetic problems. In particular, it explores possibilities to enhance the methods by combining physics-based and learning-based techniques.
The field of Computational Electromagnetics (CEM) offers many numerical methods to solve Maxwell's equations. However, they tend to be computationally demanding, especially in three spatial dimensions. In the first part of the thesis, we propose and evaluate Gradient-Informed Attentive Normalization Training (GIANT), a neural network training procedure for constructing accurate and computationally efficient surrogate models for CEM methods. GIANT consists of two main components: (i) Attentive Normalization (AttNorm), a reparameterization procedure that enables the training of very deep fully-connected neural networks; and (ii) gradient-informed training, which is used to minimize the number of samples that are needed to train the neural network. For the reciprocal microwave problems that we consider, we use continuum sensitivity analysis to compute these gradients at a very low computational cost. We demonstrate GIANT by constructing highly accurate surrogate models for a cylindrical resonator that is filled with an inhomogeneous dielectric and for an H-plane microwave filter.
Conventional material characterization techniques for electromagnetic problems require a sample of the material to be removed and placed in a separate measurement setup, which is prohibitive for many applications. In the second part of the thesis, we present two methods for material characterization that are suitable for inline applications. The first is an auto-calibration method that, given observations of uncalibrated scattering parameters, simultaneously determines both the averaged permittivity of an inhomogeneous dielectric in the measurement domain and the unknown calibration parameters of an uncalibrated measurement system. The second is a method for determining unknown parameters that describe the materials of an electric motor. The method is based on a computationally efficient physics-based model of the motor, to which we apply a Bayesian estimation framework to benefit from prior information about the sought parameters. We demonstrate the methods by applying them to synthetic data and measurement data.
The field of Computational Electromagnetics (CEM) offers many numerical methods to solve Maxwell's equations. However, they tend to be computationally demanding, especially in three spatial dimensions. In the first part of the thesis, we propose and evaluate Gradient-Informed Attentive Normalization Training (GIANT), a neural network training procedure for constructing accurate and computationally efficient surrogate models for CEM methods. GIANT consists of two main components: (i) Attentive Normalization (AttNorm), a reparameterization procedure that enables the training of very deep fully-connected neural networks; and (ii) gradient-informed training, which is used to minimize the number of samples that are needed to train the neural network. For the reciprocal microwave problems that we consider, we use continuum sensitivity analysis to compute these gradients at a very low computational cost. We demonstrate GIANT by constructing highly accurate surrogate models for a cylindrical resonator that is filled with an inhomogeneous dielectric and for an H-plane microwave filter.
Conventional material characterization techniques for electromagnetic problems require a sample of the material to be removed and placed in a separate measurement setup, which is prohibitive for many applications. In the second part of the thesis, we present two methods for material characterization that are suitable for inline applications. The first is an auto-calibration method that, given observations of uncalibrated scattering parameters, simultaneously determines both the averaged permittivity of an inhomogeneous dielectric in the measurement domain and the unknown calibration parameters of an uncalibrated measurement system. The second is a method for determining unknown parameters that describe the materials of an electric motor. The method is based on a computationally efficient physics-based model of the motor, to which we apply a Bayesian estimation framework to benefit from prior information about the sought parameters. We demonstrate the methods by applying them to synthetic data and measurement data.
neural networks
Bayesian estimation.
deep learning
material characterization
Surrogate models
statistical signal processing
TBA
Opponent: Prof. Dirk de Villiers, Department of Electrical and Electronic Engineering, University of Stellenbosch, South Africa
Author
Simon Stenmark
Chalmers, Electrical Engineering, Signal Processing and Biomedical Engineering
Stenmark, S, Rylander, T, Carlsson, J, Viberg, M, Botha, M. M. Parameter Estimation for Electric Machines with Hairpin Conductors Utilizing Hybrid Modeling
Electromagnetic waves are everywhere -- inside your microwave, in your car, in your phone. Constructing devices that rely on electromagnetic waves means solving Maxwell's equations, the fundamental equations that describe how electromagnetic waves behave. Conventional methods for solving Maxwell's equations are physics-based, in that the solution is computed from a careful description of the physics of the problem we wish to solve. Unfortunately, these methods can be difficult to apply, and they often require enormous amounts of computing power. In contrast to physics-based methods, learning-based methods can learn a problem from data, which means that a careful description of the physics is not needed. Machine learning has revolutionized many fields of science and can handle many problems that are difficult to solve using physics-based methods. However, they typically require large amounts of data, which can be difficult and expensive to collect.
In this thesis, we combine the strengths of physics-based and learning-based methods for solving Maxwell's equations. First, we construct efficient surrogate models that can mimic conventional simulation tools in a small fraction of the time. We also develop inline measurement methods that can be used for quality control directly in a production environment. We demonstrate how combining physics-based and learning-based methods allows us to tackle complicated electromagnetic problems in an efficient manner.
In this thesis, we combine the strengths of physics-based and learning-based methods for solving Maxwell's equations. First, we construct efficient surrogate models that can mimic conventional simulation tools in a small fraction of the time. We also develop inline measurement methods that can be used for quality control directly in a production environment. We demonstrate how combining physics-based and learning-based methods allows us to tackle complicated electromagnetic problems in an efficient manner.
ChaseOn Multiantenna wireless architectures for next-generation wireless systems (Mantua)
VINNOVA, 2017-01-01 -- 2021-12-31.
Modeling of RF emissions from e-axis (MORFex)
Swedish Energy Agency (P2024-00989), 2024-10-01 -- 2028-09-01.
Infrastructure
C3SE (-2020, Chalmers Centre for Computational Science and Engineering)
Chalmers e-Commons (incl. C3SE, 2020-)
Subject Categories (SSIF 2025)
Signal Processing
Artificial Intelligence
Electrical Engineering, Electronic Engineering, Information Engineering
DOI
10.63959/chalmers.dt/5921
ISBN
978-91-8103-464-6
Doktorsavhandlingar vid Chalmers tekniska högskola. Ny serie: 5921
Publisher
Chalmers
TBA
Opponent: Prof. Dirk de Villiers, Department of Electrical and Electronic Engineering, University of Stellenbosch, South Africa
Related datasets
Data for: GIANT Networks: Very Deep Fully-Connected Neural Networks Applied to Microwave Problems [dataset]
URI: https://doi.org/10.71870/8e3q-j519 DOI: 10.71870/8e3q-j519