Novel deep learning methods for 3d volumetric data



The goal of this thesis is to develop a novel deep learning method for 3d volumetric data. If successful, ultimately it should be possible to learn the simulation of smoke or mixing fluids, and therefore replace the costly simulation with a cheaper prediction in many use cases. This thesis is about writing a prototype for one or two of the approaches that we have on the table. Further details are best discussed in a meeting (Zoom, Skype, maybe personal meetings will be already allowed).

We're looking for an excellent master student that wants to spend some extra time on her or his thesis. In exchange, the topic would be very research oriented, and it can result in a top tier paper that we would write together. That should put you in a very good position in case you are thinking about a PhD (here or elsewhere).


  • Willingness to crack problems
  • Tolerance to frustration if cracking doesn't work immediately
  • Knowledge of python, pytorch, gradient descent and automatic differentiation are beneficial


  • Linux or Windows
  • Python + pytorch


For more information please contact Adam Celarek (