My PhD research topic is the similarity assessment of data from numerical PDE simulations. In short, common comparison operations for this data like L¹ or L² metrics are suboptimal, as the only consider local, element-wise distances without accounting for the context or larger structures. I investigated deep learning methods to create more accurate and robust comparison methods for this task. More recently, I worked with diffusion models in the context of fluid flows and created a benchmark to investigate the strengths and weaknesses of different neural network architectures for the prediction of turbulent flows. In addition, I am interested in more general topics in physics-based simulations, machine learning, computer graphics, and game development.

I was part of the physics-based simulation group between November 2019 and January 2025, and I am now working at TNG Technology Consulting.


Contact

E-mail: georg.kohl (at) tum.de


Publications

Theses


Teaching

Supervised Theses

  • Björn Kremser, “Learned PDE Corrections for Fluid Flows with Diffusion Models”, B.Sc. Thesis, TUM, December 2023
  • Hanfeng Wu, “Perceptual Losses for Deep Learning on Fluid Simulations”, B.Sc. Thesis, TUM, September 2021
  • Benjamin Holzschuh, “Transfer Learning for Physical Simulations”, M.Sc. Thesis, TUM, March 2021