Dresden 2026 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
DY: Fachverband Dynamik und Statistische Physik
DY 14: Machine Learning in Dynamics and Statistical Physics II
DY 14.11: Vortrag
Montag, 9. März 2026, 17:45–18:00, HÜL/S186
Self-Consistent Benchmarking of Machine Learning Force Fields via Energy-Landscape Exploration — •Anand Sharma1,2, Igor Poltavskyi1, and Alexandre Tkatchenko1 — 1Department of Physics and Materials Science, University of Luxembourg, Luxembourg — 2Indian Institute of Science Education and Research Pune, India
The rapid growth of Machine Learning Force Field (MLFF) models has prompted the development of diverse benchmarks to assess their accuracy and transferability. Most existing approaches rely on predefined test datasets, introducing biases and limiting fair comparison between models.
We introduce a general, system- and model-agnostic benchmarking framework that evaluates MLFFs through self-generated datasets. For each model, molecular structures are obtained by sampling random initial configurations of atoms and relaxing them using the model’s predicted forces. The resulting datasets are analyzed through (i) comparison with the model’s original training data, (ii) validation against ab-initio reference calculations, and (iii) cross-model dataset comparison. Applied to the SO3LR [1] and MACE-MP-0 [2] models, our framework identifies gaps in their training set coverage and enables unbiased evaluation of models’ predictive capabilities. Overall, our approach provides a consistent, extensible foundation for comparing and improving next-generation broadly applicable MLFFs.
[1] A. Kabylda, et. al, J. Am. Chem. Soc. 147, 33723 (2025).
[2] I. Batatia, et. al, J. Chem. Phys. 163, 184110 (2025).
Keywords: Machine Learning Force Fields; Atomistic Simulations; Computational Chemistry