New research teaches artificial intelligence the laws of physics
Sarvin Moradi defended her PhD thesis at the Department of Electrical Engineering on July 2nd.
Artificial intelligence can learn a great deal from data, but what if it also understood the basic laws of physics? That is the idea behind the PhD research of Sarvin Moradi that combines machine learning with physics to create more reliable AI models for complex engineering systems. The research shows that these models not only predict the behavior of systems more accurately, but also remain physically consistent and stable. This makes them a promising building block for future applications such as digital twins, predictive maintenance, intelligent control systems and energy-efficient technologies.
Modern society depends on complex engineering systems. Surgical robots, electric vehicles, power grids and industrial production processes all rely on mathematical models that predict how a system will behave. These models help engineers monitor equipment, prevent failures and operate systems safely and efficiently.
Building such models is often difficult, especially when the underlying physical processes are highly complex or not fully understood.
The limits of today's AI
Machine learning has become an increasingly popular way to create models directly from measurement data. Instead of describing every physical process with equations, AI learns patterns from real-world observations.
Although this approach can be very accurate, it also has a drawback. Many AI models function as a ‘black box’: they make predictions without explaining how they arrive at them. As a result, they can become less reliable when circumstances change—an important concern for safety-critical applications.
Teaching AI the language of physics
The research of Sarvin Moradi introduces a new method that combines machine learning with an energy-based physics framework known as port-Hamiltonian theory. Rather than relying on data alone, the AI also takes into account fundamental physical principles.
This allows the models to learn complex behavior while automatically preserving important properties such as stability and physical consistency. The approach is also suitable for systems in which multiple components interact and exchange energy, making it applicable to a wide range of engineering challenges.
Tested across different technologies
The new method was successfully tested on several types of engineering systems, including mechanical systems, viscoelastic materials, electromagnetic applications and interconnected multi-physics systems.
Across all case studies, the models accurately reproduced system behavior while maintaining the physical properties that engineers require for reliable simulations.
Building trustworthy AI
Beyond the technical advances, the research contributes to a much broader goal: developing AI that engineers can trust.
As AI is increasingly used in critical technologies, there is growing demand for systems that are not only accurate, but also reliable, interpretable and consistent with the laws of physics. By combining data with scientific knowledge, this research offers a new step towards trustworthy AI for the next generation of engineering applications.
Title of PhD thesis: . Supervisors: Dr. Maarten Schoukens, Prof. Roland Toth, and Dr. Nick Jaensson.