I am a postdoctoral researcher at Technical University Dortmund, where I work on data-driven modelling of biophysical systems — particularly focusing on cell mechanics and cell motion. My research combines physics-based insight with modern computational approaches to understand complex biological processes.
I earned my PhD in Biophysics from the University of Cologne, specialising in advanced data analysis. During my doctoral studies, I explored innovative methods for traction force microscopy and the data-driven discovery of physical laws.
After completing my PhD, I joined Helmholtz-Zentrum Hereon as a postdoctoral researcher, developing symmetrical and physically constrained neural models for solving partial differential equations. Later, at Ludwig Maximilian University of Munich, I investigated stochastic automatic differentiation techniques for Monte Carlo processes.
My current research interests revolve around the quantification of uncertainty in inverse ill-posed problems — especially in image reconstruction and the identification of governing equations (ODEs, PDEs, and SDEs). To tackle these challenges, I integrate concepts from deep learning, Bayesian inference, optimisation algorithms, and information theory.
Beyond research, I am passionate about interdisciplinary collaboration and advancing reproducible computational science.
PhD in Biophysics, 2021
University of Cologne, Germany
MSc in Solid Mechanics, 2015
Ecole Polytechnique, France
Swipe left or right
I develop hybrid Vision Transformer + U-Net networks that reconstruct the forces cells exert on soft substrates directly from microscopy data, for 2D and 3D traction force microscopy. The models generalize across spatial scales and noise levels and can take metadata such as cell type as extra input. They build on my Bayesian inference methods with automatic regularization, released as open-source software.
I design symmetry-equivariant convolutional networks with hard physical constraints: new input and output layers for staggered grids that enforce rotation symmetry and the conservation of mass and momentum. These surrogates give accurate, stable long rollouts for shallow-water flow and turbulence and generalize beyond the training data (ICML 2025). The video compares the reference solver with our network.
Which equation produced this data? My sparse Bayesian learning algorithm with automatic thresholding selects the few terms that govern a system and quantifies the uncertainty of every coefficient, while an active-learning loop decides which perturbations to apply next. It recovers ODEs, PDEs and stochastic differential equations, from the Lorenz attractor to Turing patterns.
Neural PDE surrogates with geometric and physical constraints that give more accurate and stable long-term predictions (ICML 2025).
At the University of Hamburg, I have given the exercise sessions.
At the University of Hamburg, I have given the exercise sessions.