Implicit Machine Learning Force Fields

Accelerate Molecular Dynamics Simulations

Johannes Maeß^1,2 Leon Werner^1,2 J. Thorben Frank^1,2 Winfried Ripken^1,2 Martin Michajlow^1,2 Joshua Futterer^1,2 Klaus-Robert Müller^† 1,2,3,4 Stefan Chmiela^† 1,2

¹ BIFOLD, Berlin ² Machine Learning Group, TU Berlin ³ MPI for Informatics, Saarbrücken

⁴ Department of Artificial Intelligence, Korea University, Seoul ^† Corresponding authors.

arXiv Code

1st MD step: Many iterations
to converge

2nd MD step: Reuse of fixed-point
→ 2–5 iterations

3rd+ MD step: Linear extrapolation
→ ∼1 iteration

Citation

@article{maess2026imlff,
  title={Implicit Machine Learning Force Fields Accelerate Molecular Dynamics Simulations},
  author={Mae{\ss}, Johannes and Werner, Leon and Frank, J. Thorben and Ripken, Winfried
  and Michajlow, Martin and Futterer, Joshua and M{\"u}ller, Klaus-Robert and Chmiela, Stefan},
  journal={arXiv preprint arXiv:2607.29158},
  year={2026}
}