Zum Inhalt springen
Technische Universität München

Machine Learning Augmented Molecular Simulations

Advancing model fidelity with multi-body potentials parametrized by deep learning architectures and learning dynamics of molecular systems.
Recent Papers:
Accurate machine learning force fields via experimental and simulation data fusion

S. Röcken, J. Zavadlav, Npj Comput. Mater. 2024, paper , arXiv, GitHub

The benefits of training ML potentials with both ab initio and experimental data are better agreement with experimental observations and out-of-target property generalization.

Deep Coarse-grained Potentials via Relative Entropy Minimization

S. Thaler, M. Stupp, J. Zavadlav, J. Chem. Phys. 2022, paper, arXiv, GitHub

Coarse-grained ML potentials trained with Relative Entropy result in more acurate potential energy srufaces, require less data and can be employed with larger integration timesteps compared to the conventional training with Force Matching.

Learning neural network potentials from experimental data via Differentiable Trajectory Reweighting

S. Thaler, J. Zavadlav, Nat. Commun. 2021, paper, arXiv, GitHub
received MDSI Best Paper of Year 2021 Award
selected as Editors’ Highlights paper in Applied physics and mathematics research, featured in Nat. Commun. collection Molecular Dynamics simulations and Computational methods in Life Science.

Efficient top-down parametrization of deep neural network molecular models, enabling molecular models to exactly match experimental observations.

Accelerated Simulations of Molecular Systems through Learning of their Effective Dynamics

P. R. Vlachas, J. Zavadlav, M. Praprotnik, P. Koumoutsakos, J. Chem. Theory Comput. 2021, paper,  arXiv, GitHub

A probabilistic machine learning model that can learn the dynamics of molecular systems, which accelerates simulation timescales by up to three orders of magnitude.

To top