idw – Informationsdienst Wissenschaft
Nachrichten, Termine, Experten
Nachrichten, Termine, Experten
A simple transformer-based model challenges the role of physical constraints in molecular dynamics simulations
Simulating how atoms and molecules move over time is a central challenge in computational chemistry and materials science. Classical machine learning approaches to molecular dynamics (MD) encode fundamental physical principles directly into their model architectures, most notably energy conservation and equivariance, the requirement that predicted forces remain consistent regardless of how a molecule is oriented in space. These so-called inductive biases have long been considered essential for reliable, physically meaningful MD models. But are they truly indispensable?
Max Eissler, Tim Korjakow, Stefan Ganscha, Oliver T. Unke, Klaus-Robert Müller, and Stefan Gugler have developed a novel approach that deliberately strips away these physical constraints and instead relies on a general-purpose transformer architecture to learn physical behavior from data. Their work, titled “How simple can you go? An off-the-shelf transformer approach to molecular dynamics,” was published in the Journal of Chemical Physics on 3 March 2026 as a collaboration between BIFOLD members and Google DeepMind researchers Their model, MD-ET, is built on an edge transformer (ET), an architecture that has been only minimally adapted for the MD domain, and implements neither built-in equivariance nor energy conservation.
MD-ET: State-of-the-art results without physical hard-coding
Rather than engineering physical laws into the model’s architecture, the team trained MD-ET on approximately 30 million molecular structures from the QCML database (https://www.nature.com/articles/s41597-025-04720-7 ), an enormous set of molecules that was also designed at BIFOLD in a collaboration with Google DeepMind. The key insight behind this approach is that the model may be able to approximate physical behaviour, such as equivariant force predictions, purely from exposure to large amounts of data, without those behaviors being explicitly enforced. After pretraining, the model is fine-tuned for a small number of steps on target systems.
The results are notable. On several established benchmarks, MD-ET achieves competitive or state-of-the-art performance. The model learns to predict forces that are approximately equivariant, with deviations many orders of magnitude below typical force magnitudes. Stable NVT simulations, where temperature and particle count are held constant, succeed even in a few-shot setting. The authors identify the large pretraining dataset and the relative ease of optimizing edge transformers as key factors behind these results.
However, the picture is more nuanced for NVE simulations, systems with fixed energy and no thermostat. Here, energy conservation is only approximately learned and proves sensitive to both molecular size and numerical perturbations. For larger structures, the model exhibits runaway energy increases, highlighting a real limitation of the unconstrained approach. The authors are candid about this: MD simulations using non-conservative forces should be carefully validated on a case-by-case basis.
Taken together, the findings contribute to an ongoing discussion in the field about the extent physical inductive biases are necessary components of MD models, or whether sufficiently expressive general-purpose architectures trained on large datasets can learn to respect physical principles without being explicitly constrained to do so. MD-ET suggests that for many practical applications, the latter may be achievable, while also clarifying where the boundaries of this approach currently lie.
Paper: How simple can you go? An off-the-shelf transformer approach to molecular dynamics.
Authors: Max Eissler, Tim Korjakow, Stefan Ganscha, Oliver T. Unke, Klaus-Robert Müller, Stefan Gugler.
Journal of Chemical Physics (2026): https://pubs.aip.org/aip/jcp/article/164/9/094308/3381953/How-simple-can-you-go-...
BIFOLD News: https://t1p.de/2cmw5
Dr. Stefan Gugler
Postdoctoral Researcher
stefan.gugler(∂)tu-berlin.de
Technical University of Berlin
BIFOLD – Berlin Institute for the Foundations of Learning and Data
Paper: How simple can you go? An off-the-shelf transformer approach to molecular dynamics.
Authors: Max Eissler, Tim Korjakow, Stefan Ganscha, Oliver T. Unke, Klaus-Robert Müller, Stefan Gugler.
Journal of Chemical Physics (2026): https://pubs.aip.org/aip/jcp/article/164/9/094308/3381953/How-simple-can-you-go-...
https://BIFOLD News: https://t1p.de/2cmw5
Publication Preview "How simple can you go? An off-the-shelf transformer approach to molecular dynam ...
Copyright: Journal of Chemical Physics (2026)
Merkmale dieser Pressemitteilung:
Journalisten, Wissenschaftler
Chemie, Informationstechnik, Werkstoffwissenschaften
überregional
Forschungsergebnisse
Englisch
Publication Preview "How simple can you go? An off-the-shelf transformer approach to molecular dynam ...
Copyright: Journal of Chemical Physics (2026)
Sie können Suchbegriffe mit und, oder und / oder nicht verknüpfen, z. B. Philo nicht logie.
Verknüpfungen können Sie mit Klammern voneinander trennen, z. B. (Philo nicht logie) oder (Psycho und logie).
Zusammenhängende Worte werden als Wortgruppe gesucht, wenn Sie sie in Anführungsstriche setzen, z. B. „Bundesrepublik Deutschland“.
Die Erweiterte Suche können Sie auch nutzen, ohne Suchbegriffe einzugeben. Sie orientiert sich dann an den Kriterien, die Sie ausgewählt haben (z. B. nach dem Land oder dem Sachgebiet).
Haben Sie in einer Kategorie kein Kriterium ausgewählt, wird die gesamte Kategorie durchsucht (z.B. alle Sachgebiete oder alle Länder).
