Collaboration
If you are working on related topics and think there might be an interesting connection, I would be very happy to hear from you. Feel free to reach out by email for potential collaborations, discussions, or simply to exchange ideas.
Research
Research vision
My research lies at the intersection of generative modeling, scientific machine learning, and control. I am interested in models that learn from data while respecting structure known a priori, such as physical laws or geometric constraints. The goal is to bridge data-driven learning and first-principles modeling: expressive enough to represent uncertainty and complex dynamics, but structured enough to remain meaningful in scientific and physical contexts.
Learning flexible models for scientific systems while preserving the structure that makes them meaningful.
Main themes
Theme 01
Generative Modeling for Scientific Data
Generative models are useful in scientific problems because they can represent uncertainty over structured objects: fields, trajectories, PDE solutions, and other states constrained by underlying mechanisms. I am particularly interested in flow matching or diffusion models in this context.
In many applications, the goal is not generation for its own sake, but inference under incomplete information. A learned distribution can encode which states are plausible and can then be combined with observations, constraints, or objectives to solve inverse problems, guide optimization, or quantify uncertainty.
Flow matching Diffusion models Generative priors Uncertainty Inverse problems
Theme 02
Domain-Informed and Physics-Informed AI
Scientific learning problems usually come with structure that should not be ignored. I am interested in methods that incorporate such prior knowledge into machine learning models, including physical laws, symmetries, boundary conditions, constraints, expert knowledge, or qualitative information about admissible solutions.
This is the role in which I view physics-informed AI: not as a single technique, but as part of a broader question of how domain knowledge should shape learning. The challenge is to keep models flexible and data-adaptive while making their behavior consistent with the structure of the problem, especially when data are limited, simulations are imperfect, or extrapolation matters.
Physical constraints Prior knowledge Hybrid modeling Residuals Structure
Theme 03
Control and Decision-Making in Physical Systems
Control problems make the consequences of learning explicit: predictions become actions, actions change the state, and errors can accumulate through feedback. This makes physical control a natural setting for studying how machine learning interacts with dynamics, constraints, uncertainty, and optimization objectives.
My Master’s thesis at Porsche Motorsport explored this perspective through reinforcement learning for race driver modeling. I remain interested in control both as an application area and as a conceptual lens for learned models that should not only describe data, but also support decisions in evolving systems.
Reinforcement learning Feedback systems Dynamics Closed-loop behavior Optimization
Contact
Supervision
I am also open to supervising motivated students for thesis projects. If you are interested in working with me, please send me an email and briefly describe what you are excited about in particular, what background you bring, and which kind of problem you would like to work on. You do not need to have a fully developed project proposal, but it is helpful if you already have a concrete direction or research question that genuinely interests you.