Multi-head committees enable direct uncertainty prediction for atomistic foundation models

Result code in IS VaVaI
<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11320%2F25%3A10510127" target="_blank" >RIV/00216208:11320/25:10510127 - isvavai.cz</a>
Result on the web
<a href="https://verso.is.cuni.cz/pub/verso.fpl?fname=obd_publikace_handle&handle=pYZ9mWAi.1" target="_blank" >https://verso.is.cuni.cz/pub/verso.fpl?fname=obd_publikace_handle&handle=pYZ9mWAi.1</a>
DOI - Digital Object Identifier
<a href="http://dx.doi.org/10.1063/5.0302097" target="_blank" >10.1063/5.0302097</a>

Result language
angličtina
Original language name
Multi-head committees enable direct uncertainty prediction for atomistic foundation models
Original language description
Machine learning potentials have become a standard tool for atomistic materials modeling. While models continue to become more generalizable, an open challenge relates to efficient uncertainty predictions for active learning and robust error analysis. In this work, we utilize MACE and its multi-head mechanism to implement a committee neural network potential for message-passing architectures, where the committee comprises multiple output modules attached to the same atomic environment descriptors. As with traditional committees of independent networks, the standard deviation of the predictions functions as an estimate of the model's uncertainty. We show for a range of datasets in custom-build models that the uncertainty of the force predictions correlates well with the true errors. We subsequently apply this concept to foundation models, in particular MACE-MP-0, where we train only the newly attached output heads while keeping the remaining part of the model fixed. We use this approach in an active learning workflow to condense the training set of the foundation model to just 5% of its original size. The foundation model multi-head committee trained on the condensed training set enables reliable uncertainty estimation without any substantial decrease in prediction accuracy.
Czech name
—
Czech description
—

Type
J<sub>imp</sub> - Article in a specialist periodical, which is included in the Web of Science database
CEP classification
—
OECD FORD branch
10301 - Atomic, molecular and chemical physics (physics of atoms and molecules including collision, interaction with radiation, magnetic resonances, Mössbauer effect)

Project
<a href="/en/project/GA21-27987S" target="_blank" >GA21-27987S: Accurate molecular dynamics of liquids and solvation through machine learning of ab initio interactions</a><br>
Continuities
P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Publication year
2025
Confidentiality
S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Similar results(10)