Crash testing machine learning force fields for molecules, materials, and interfaces - model analysis - TEA Challenge 2023

Aug 5, 2026

Upload by Riley

Atomistic simulations underpin research and industrial studies of molecules, materials, and their interfaces, but their accuracy depends on force fields that must capture complex interatomic interactions across spatio-temporal scales and an enormous chemical space. Machine-learning force fields trained on quantum energies and forces reach near sub-kcal/mol·Å accuracy while remaining computationally efficient. The TEA Challenge 2023 systematically evaluated widely used MLFFs, analyzing how they reproduce potential energy surfaces, cope with incomplete reference data, model multi-component systems, and handle complex periodic structures.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages18

Reading Time45 min

AccessFree to download

Introduction

Force fields and machine-learning force fields

TEA Challenge 2023 evaluation scope

Datasets and proposed challenges

Model architectures compared

Analysis of accuracy, stability, and efficiency

Open Access Article . Pu on 10 Februaryblished 2025. Down on 3/31/2025loaded 3:40: 18 PM . hicle is licensed under a Creative C mmons A 3 0 U d[ttr .](ttr .)ibution npor e nce. Chemical Science | EDGE ARTICLE | View Article Online
View Journal | View Issue | | --- | --- | |
Cite this: Chem. Sci., 2025, 16, 3720

All publication charges for this article | Crash testing machine learning force ﬁelds for molecules, materials, and interfaces: model analysis in the TEA Challenge 2023† | | have been paid for by the Royal Society of Chemistry | Igor Poltavsky, *a Anton Charkin-Gorbulin,‡ab Mirela Puleva,‡ac Grgory Fonseca,‡a Ilyes Batatia,d Nicholas J. Browning,e Stefan Chmiela,fg Mengnan Cui,h
J. Thorben Frank, fg Stefan Heinen,i Bing Huang,j Silvan Kser, k Adil Kabylda, a Danish Khan,il Carolin Müller, m Alastair J. A. Price,no Kai Riedmiller, p Kai Töpfer, k Tsz Wai Ko,q Markus Meuwly, k Matthias Rupp, r Gbor Csnyi, d
O. Anatole von Lilienfeld, fginost Johannes T. Margraf,u Klaus-Robert Müller fgvwx and Alexandre Tkatchenko *ac | | Received 26th September 2024 Accepted 25th December 2024
DOI: 10.1039/d4sc06529h | Atomistic simulations are routinely employed in academia and industry to study the behavior of molecules, materials, and their interfaces. Central to these simulations are force ﬁelds (FFs), whose development is challenged by intricate interatomic interactions at diﬀerent spatio-temporal scales and the vast expanse of chemical space. Machine learning (ML) FFs, trained on quantum-mechanical energies and forces, have shown the capacity to achieve sub-kcal (mol−1 Å−1) accuracy while maintaining computational eﬃciency. The TEA Challenge 2023 rigorously evaluated commonly used MLFFs across diverse applications, highlighting their strengths and weaknesses. Participants trained their models using provided datasets, and the results were systematically analyzed to assess the ability of MLFFs to reproduce potential energy surfaces, handle incomplete reference data, manage multi-component systems, and model complex periodic structures. This publication describes the datasets, outlines the proposed challenges, and presents a detailed analysis of the accuracy, stability, and eﬃciency of the MACE, SO3krates, sGDML, SOAP/GAP, and FCHL19 * architectures in molecular dynamics simulations. The models represent the MLFF developers who participated in the TEA Challenge 2023 . All results presented correspond to the state of the ML architectures as of October 2023 . A comprehensive analysis of the molecular dynamics | | [rsc.li/chemical-science](rsc.li/chemical-science) | results obtained with diﬀerent MLFFs will be presented in the second part of this manuscript. | aDepartment of Physics and Materials Science, University of Luxembourg, L-1511 Luxembourg, Luxembourg. E-mail: [alexandre.tkatchenko@uni.lu](alexandre.tkatchenko@uni.lu); igor.poltavskyi@ [uni.lu](uni.lu) bLaboratory for Chemistry of Novel Materials, University of Mons, B-7000 Mons, Belgium cInstitute for Advanced Studies, University of Luxembourg, Campus Belval, L-4365 Esch-sur-Alzette, Luxembourg dDepartment of Engineering, University of Cambridge, Trumpington Street, Cambridge, CB2 1PZ, UK eSwiss National Supercomputing Centre (CSCS), 6900 Lugano, Switzerland fMachine Learning Group, Technical University Berlin, Berlin, Germany gBIFOLD, Berlin Institute for the Foundations of Learning and Data, Berlin, Germany hFritz-Haber-Institut der Max-Planck-Gesellscha􀀁, Berlin, Germany iVector Institute for Arti􀀁cial Intelligence, Toronto, ON, M5S 1M1, Canada jWuhan University, Department of Chemistry and Molecular Sciences, 430072 Wuhan, China kDepartment of Chemistry, University of Basel, Klingelbergstrasse 80, CH-4056 Basel, Switzerland lChemical Physics Theory Group, Department of Chemistry, University of Toronto, St George Campus, Toronto, ON, Canada mFriedrich-Alexander-Universitt Erlangen-Nürnberg, Computer-Chemistry-Center, Nge

FAQs

What problem does the TEA Challenge 2023 address for machine-learning force fields?

It evaluates commonly used ML force fields across diverse applications to identify strengths and weaknesses in reproducing potential energy surfaces and handling challenging data and system conditions.

Which aspects of MLFF performance are analyzed in the publication?

The study analyzes accuracy, stability, and efficiency, including performance in molecular dynamics simulations and behavior for incomplete reference data and multi-component or periodic structures.

Which MLFF architectures are covered in the detailed analysis?

The publication presents analysis for MACE, SO3krates, sGDML, SOAP/GAP, and FCHL19 architectures in the context of molecular dynamics simulations.

Aug 5, 2026

Upload by Riley