Dawid Pasterny

Machine-learning surrogates for physics simulations. Machine-learning engineer at the German Aerospace Center (DLR) in Augsburg.

GitHubAll notesRSS

Recent notes

• Theories are works of art, symphonies of thought
  Note · 12 November 2026 · Science
  Theories are works of art and they should be presented as such.
• Memex
  Note · 12 November 2026 · Computer science
  Vannevar Bush's 1945 machine for associative memory, and the vault behind this site.
• Fast multipole method
  Note · 12 November 2026 · Computer science › Algorithms
  Barnes–Hut's successor: multipole and local expansions bring the n-body sum down to O(N).
• Dislocation theory
  Note · 12 November 2026 · Engineering › Material science › Dislocation theory
  Why metals bend instead of shattering: the line defects that carry plastic flow.
• Crystal plasticity
  Note · 12 November 2026 · Engineering › Material science › Plasticity
  From slip on one plane to the stress–strain curve of a polycrystal.
• Bravais lattices
  Note · 12 November 2026 · Physics › Solid state physics
  The fourteen ways to repeat a point in space, and why metals pick only three of them.

All notes →

About

I build machine-learning surrogates for physics simulations: neural models that learn from a solver and then answer in milliseconds what the solver needs hours for. My background is mechanical engineering and computational science, and I work across the stack, from CUDA kernels and finite-element solvers to training loops and the coupling between them.

I currently work on multiscale phase-field fracture at the German Aerospace Center (DLR) in Augsburg. The notes on this site come from my memex, a personal knowledge base that I write together with an LLM. Most of them are explorable essays: you change a parameter and the figure answers.

Experience

• Machine Learning Engineer, German Aerospace Center (DLR), Augsburg · 2024–present. Physics-aware neural surrogates that accelerate phase-field fracture solvers; an end-to-end framework for multiscale, multiphysics fracture simulation on FEniCSx and JAX; training and simulation campaigns on GPU and HPC clusters. Two first-author papers on multiscale phase-field fracture are in preparation.
• FEM Analyst, Avio Aero (GE Aerospace), Bielsko-Biała · 2023–2024. Structural analysis of low-pressure turbine components for a turboprop engine in Ansys APDL, from model build to design review; Python and JavaScript tools that automate model setup and post-processing.
• Guided research, Technical University of Munich · 2021–2022. How far directional message-passing graph neural networks for molecular property prediction generalise to molecules larger than, and different from, those they were trained on.
• Research assistant, Chair of Machine Elements, Technical University of Munich · 2020–2022. Reinforcement learning for automated gearbox design: DQN, A3C, DDPG, SAC and HER implemented from scratch, and a CadQuery interface for parametric 3D gearboxes.
• Aerodynamics and CFD, Formula Student team Green Team, University of Stuttgart · 2017–2018. Full-vehicle CFD in Star-CCM+ with automated setup and evaluation, composite manufacturing, competitions in Spain and Germany.

Education

• M.Sc. Computational Science and Engineering, Technical University of Munich · 2019–2023. Thesis: approximating many-electron wave functions with physics-aware surrogate models.
• B.Sc. Mechanical Engineering (Informatics in Mechanical Engineering), AGH University of Krakow · 2015–2019, with an Erasmus year at the University of Stuttgart. Thesis: topology optimisation of a Formula Student wheel carrier.

Tools

• Machine learning: PyTorch, JAX, scikit-learn; graph neural networks, physics-informed and surrogate models, reinforcement learning.
• Scientific computing: FEniCSx, Ansys APDL, Star-CCM+, CadQuery; finite elements, CFD, phase-field fracture, topology optimisation.
• High-performance computing: C/C++, CUDA, MPI, OpenMP; perf, Valgrind, SLURM.
• Languages: Polish (native), English (fluent), German (conversational).