Public defence in Computer Science, M.Sc. Katsiaryna Haitsiukevich

Title of the thesis: Advances in physics-informed deep learning

Thesis defender: Katsiaryna Haitsiukevich 
Opponent: Tenure Track Assistant Professor Olga Fink, The École polytechnique fédérale de Lausanne (EPFL), Switzerland
Custos: Associate Professor Pekka Marttinen, Aalto University School of Science, Department of Computer Science

Accurate modeling of physical systems is essential for engineering and scientific discovery. Reliable models help optimize manufacturing processes and deepen our understanding of complex phenomena. Physical systems — such as industrial processes or chemical reactions — are traditionally described using mathematical equations based on physical laws. While these traditional models are powerful, they can be slow and rely on oversimplified assumptions. At the same time, machine learning, particularly deep learning, offers a data-driven alternative. However, these methods typically require large datasets, which are often difficult or expensive to collect.

This thesis bridges the gap between traditional and data-driven approaches. By integrating prior physical knowledge into machine learning techniques, it develops methods that combine the strengths of traditional models with the flexibility of machine learning approaches. This enables the development of accurate and efficient hybrid models that can be applied in real-world scenarios with limited data. The findings demonstrate that even with scarce measurements, these hybrid approaches can effectively model industrial systems, unlocking new possibilities for innovation.

Keywords: neural networks, physics-informed neural networks, differential equations, physical system modeling, sample efficient modeling, prior knowledge incorporation, industrial applications of deep learning

Thesis available for public display 10 days prior to the defence at Aaltodoc.