Refining machine learning potentials through thermodynamic theory of phase transitions
- 1Multiscale Modeling of Fluid Materials, Technical University of Munich, Garching, Bavaria, Germany
- 2Atomistic Modeling Center, Munich Data Science Institute, Technical University of Munich, Garching, Bavaria, Germany
Abstract
Foundational machine learning potentials can alleviate the accuracy and transferability limitations of classical force fields. They can substantially expedite material design and discovery by providing microscopic insights into material behavior through Molecular Dynamics simulations. However, insufficiently broad and systematically biased reference data affect the predictive quality of the learned models. These models often exhibit significant deviations from experimentally observed phase transition temperatures by several hundred kelvins. Finetuning is therefore necessary to achieve adequate accuracy in many practical problems. This work proposes a top-down finetuning strategy that corrects inaccurately predicted transition temperatures using experimental reference data. We introduce the Differentiable Transition Temperature Correction method to minimize the free energy differences between phases at the experimental target pressures and temperatures. Using a benchmark of pure Titanium at pressures up to 5 GPa, we show that our approach substantially improves the predicted phase diagram and liquid-state diffusion constant. Transition temperatures are within tens of kelvins of the experimental reference. Our model-agnostic approach can be supplemented with top-down training on additional experimental properties. In principle, it is also generally applicable to correct other free energy differences beyond the field of materials science. Thus, our approach can serve as an essential step towards highly accurate machine learning potentials.
Related
- Targeted fine-tuning of machine-learning interatomic potentials for phonons and phase transitions.
- Machine Learning of Band Gaps and Thermodynamic Stability in A2BB′X6 (X = F, Cl, Br, I) Double Perovskites with Uncertainty-Aware Multi-Objective Screening
- Correcting the Liquid Densities of Universal Machine-Learning Interatomic Potentials Using Local-Number-Density Repulsion
- pdb2reaction: End-to-End Reaction-Path Elucidation from PDB Structures Using Machine-Learning Interatomic Potentials
- Applicability Domain of Universal Machine Learning Interatomic Potentials at Solid-State Battery Interfaces
BibTeX
@article{fuchs2026refining,
title = {{Refining machine learning potentials through thermodynamic theory of phase transitions}},
author = {Paul Fuchs and Julija Zavadlav},
journal = {npj Computational Materials},
year = {2026},
month = jun,
doi = {10.1038/s41524-026-02195-7},
url = {https://doi.org/10.1038/s41524-026-02195-7},
note = {CC BY},
}
RIS
TY - JOUR
TI - Refining machine learning potentials through thermodynamic theory of phase transitions
AU - Paul Fuchs
AU - Julija Zavadlav
PY - 2026
DA - 2026/06/22
JO - npj Computational Materials
DO - 10.1038/s41524-026-02195-7
UR - https://doi.org/10.1038/s41524-026-02195-7
AB - Foundational machine learning potentials can alleviate the accuracy and transferability limitations of classical force fields. They can substantially expedite material design and discovery by providing microscopic insights into material behavior through Molecular Dynamics simulations. However, insufficiently broad and systematically biased reference data affect the predictive quality of the learned models. These models often exhibit significant deviations from experimentally observed phase transition temperatures by several hundred kelvins. Finetuning is therefore necessary to achieve adequate accuracy in many practical problems. This work proposes a top-down finetuning strategy that corrects inaccurately predicted transition temperatures using experimental reference data. We introduce the Differentiable Transition Temperature Correction method to minimize the free energy differences between phases at the experimental target pressures and temperatures. Using a benchmark of pure Titanium at pressures up to 5 GPa, we show that our approach substantially improves the predicted phase diagram and liquid-state diffusion constant. Transition temperatures are within tens of kelvins of the experimental reference. Our model-agnostic approach can be supplemented with top-down training on additional experimental properties. In principle, it is also generally applicable to correct other free energy differences beyond the field of materials science. Thus, our approach can serve as an essential step towards highly accurate machine learning potentials.
SN - 2057-3960
ER -