Can a model know what it does not know?
Uncertainty-aware and physics-informed machine learning.
Computational methods for scientific questions.
I develop machine-learning, optimization, and software methods that make scientific prediction more useful, efficient, and reliable.
Research direction
I work at the intersection of scientific computing, machine learning, and optimization. My research uses permanent magnets as a demanding application area for developing methods that can support well-founded decisions from complex and limited data.
At the Christian Doppler Laboratory for magnet design through physics-informed machine learning, I work on data-mining routines, predictive models, and their validation in collaboration with project partners at CNRS and Toyota.
How I approach the problem
Design tasks often involves competing objectives, limited data, and expensive evaluations such as real world experiments or high-fidelity simulations. I analyse, augment and curate datasets, develop machine learning models to find patterns and to explore design trade-offs with optimization algorithms. ML models are also useful to replace experiments when direct evalution is impractical due to resource constraints. We need reliable models to support decision-making.
So, a central question is whether a model can be trusted where it is used. My work therefore evaluates uncertainty alongside prediction error. Incorporating uncertainty-aware models into optimization workflows offers a key advantage: We can constrain the confidence of predicted designs. This helps distinguish bad candidate designs from promising solutions.
Research themes
The questions that keep my work moving.
How can expensive science become feasible?
Surrogate models for costly scientific predictions.
Which trade-offs are worth making?
Multi-objective optimization under real constraints.
What should we learn next?
Active learning for efficient computational experiments.
How to transfer computational methods to various problems?
Modular, reproducible scientific software.