I'm a master's student in Computational Materials Science, and I like using machine learning to make slow, expensive simulations fast.
📍 Freiberg, Germany
- Teaching neural networks physics: surrogate models, physics-informed networks and neural operators
- Making ML results trustworthy: testing on unseen data, uncertainty, and honest evaluation
- Making things fast: multi-GPU training and inference optimisation
- multi-pde-pretraining-transfer: does pretraining on several physics problems help a network learn a new one faster? A small, controlled experiment.
- neural-pde-surrogate-benchmark: neural networks that predict how heat spreads, compared head to head (MLP, CNN, Fourier neural operator).
- JAX-PINN: a physics-informed neural network in JAX that recovers equation parameters from data.
- pytorch-ddp-scaling-lab: how well does training scale across several GPUs, and where does the time go?
- deep-learning-inference-benchmark: PyTorch vs ONNX Runtime vs TensorRT, and which runs a model fastest.
- CrystalDiff: a diffusion model that generates 3D crystal structures (research prototype).
- sem-segmentation-deep-learning: finding grain boundaries in microscope images using synthetic training data.
- Finishing my master's thesis at HZDR on ML surrogates for laser-physics simulations
- Exploring pretraining and transfer for neural operators
