Physics-informed machine learning for electromagnetism

Licentiate thesis, 2026

Electromagnetism underpins a wide range of modern technologies, from electric machines and antennas to optical components and photonic circuits. To design these devices, accurate and efficient modeling of electromagnetic phenomena is essential. Traditional numerical techniques such as finite-element and finite-difference methods have achieved great success over many decades, but their high computational cost becomes a bottleneck as electromagnetic devices grow in complexity.

In recent years, machine learning has emerged as a powerful tool for accelerating electromagnetic modeling by replacing costly numerical simulations with fast surrogate models. Such models can enable rapid exploration of large design spaces and thereby support the development of higher-performing devices. However, purely data-driven models often require impractical amounts of training data and may fail to respect fundamental physical laws. Physics-informed machine learning addresses these limitations by embedding prior knowledge about the governing physics into the model architecture or learning process.

In this thesis, I explore new methods for physics-informed machine learning in electromagnetism, with the aim of improving data efficiency, reliability, and interpretability of neural network-based surrogate models. The thesis centers on two complementary methods. The first exploits the resonant structure underlying electromagnetic scattering spectra, while the second utilizes symmetries in Maxwell’s equations. By developing widely applicable tools to improve surrogate models, the work contributes to advancing electromagnetic analysis and design.