Peter Pavlík

Peter is a researcher focused on physics-informed machine learning and its applications in weather, climate, and geosciences. His work incorporates physical and domain knowledge into data-driven models with the aim of making them more accurate, trustworthy and less energy-demanding. He is particularly interested in environmentally responsible applications of AI, including improving renewable energy systems, enhancing early warning systems for extreme weather, and supporting climate change mitigation and adaptation strategies.

Peter holds a Master’s degree in Intelligent Software Systems from the Faculty of Informatics and Information Technologies at the Slovak University of Technology. He is currently a PhD candidate at the Faculty of Information Technology at Brno University of Technology, where his research focuses on physics-informed machine learning for forecasting tasks. As part of his doctoral work, he developed and published a novel physics-informed neural network architecture for precipitation nowcasting – LUPIN.

During his PhD, Peter has completed research visits at Delft University of Technology (Faculty of Civil Engineering and Geosciences) and at Know Center in Graz, working on interdisciplinary problems in geoscience and applied machine learning. He has presented his work at venues such as the EGU General Assembly and the ESA-ECMWF Machine Learning Workshop, among others.

Supervising team: Viera Rozinajová (KInIT), Anna Bou Ezzeddine (KInIT), Softec (industry partner)

We explore the use of deep learning to improve the efficiency and accuracy of forecasts of physical systems. Accurate forecasting of changes occurring in the real world is inseparable from modeling and simulation of underlying physical processes. This was traditionally performed using complex numerical physical simulations. Data-driven deep learning emulators are a promising alternative to these numerical approaches, showing many advantages in some areas but also lacking in others. Machine learning methods incorporating existing domain knowledge about the modeled physical processes can combine the best aspects of both paradigms. This is the premise of physics-informed machine learning (PIML). We apply it in the weather forecasting domain, working closely with the company SOFTEC and the Slovak Hydrometeorological Institute. We train models for the task of precipitation nowcasting from radar reflectivity data, with a focus on correctly estimating the precipitation quantity in the solid phase for the purposes of optimizing winter road maintenance.

Interns: