Graph Neural Networks for Molecular Force Fields

For the simulation of molecular systems, atomic forces need to be predicted in each timestep. While those forces can be derived from quantum mechanical calculations, it is orders of magnitude more efficient to use physics-inspired approximative models, so-called molecular mechanics force fields. In the past few years, SE(3) equivariant graph neural networks have been used for constructing far more accurate force fields, however, their computational cost is still prohibitively high for many systems.

Instead, molecular mechanics can be enhanced by graph neural networks for developing highly efficient machine-learned force fields such as Grappa [1]. We plan to extend this model by more expressive physics-inspired terms and advanced graph neural network architectures.

Project type:

Master thesis in computer science or physics at Heidelberg University / KIT

Prerequisites:

  • Familiarity with Deep Learning and PyTorch
  • Basic understanding of graphs and geometry
  • No prior knowledge on molecules needed

Contact:

References:

[1] Leif Seute, Eric Hartmann, Jan Stühmer and Frauke Gräter, Grappa – a machine learned molecular mechanics force field, Chemical Science, 2025

Switch to the German homepage or stay on this page