Universal Machine Learning Potentials under Pressure

Antoine Loew, Jonathan Schmidt, Silvana Botti, Miguel A. L. Marques · 2025

Read Universal Machine Learning Potentials under Pressure by Antoine Loew, Jonathan Schmidt, Silvana Botti, Miguel A. L. Marques (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-17792.txt.

Introduction

Computer models that predict how atoms push on each other have become workhorses of materials science, promising one tool for nearly any substance. Antoine Loew, Jonathan Schmidt, Silvana Botti, and Miguel Marques stress-test that promise under extreme pressure in this 2025 paper. They benchmark the leading universal models from zero to one hundred and fifty gigapascals, far beyond everyday conditions.

At normal pressure the models shine, but their accuracy fades steadily as pressure climbs. The authors trace the fade not to the methods themselves but to gaps in the data the models learned from, which barely samples crushed matter. A round of extra training on high-pressure examples restored much of the lost accuracy with ease. The lesson, they argue, is to hunt for blind regimes like this one before calling a model universal.

Catalog record

CREATORAntoine Loew, Jonathan Schmidt, Silvana Botti, Miguel A. L. Marques
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-17792

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Universal Machine Learning Potentials under Pressure Antoine Loew, Jonathan Schmidt, Silvana Botti, Miguel A. L. Marques Abstract Universal machine learning interatomic potentials (uMLIPs) represent arguably the most successful application of machine learning to materials science, demonstrating remarkable performance across diverse applications. However, critical blind spots in their reliability persist. Here, we address one such significant gap by systematically investigating the accuracy of uMLIPs under extreme pressure conditions from 0 to 150 GPa. Our benchmark reveals that while these models excel at standard pressure, their predictive accuracy deteriorates considerably as pressure increases. This decline in performance originates from fundamental limitations in the training data rather than in algorithmic constraints. In fact, we show that through targeted fine-tuning on high-pressure configurations, the robustness of the models can be easily increased. These findings underscore the importance of identifying and addressing overlooked regimes in the development of the next generation of truly universal interatomic potentials. I Introduction High-pressure materials science has emerged as a promising pathway for discovering new materials with valuable applications, ranging from superconductivity to nanostructuring McMillan 2002 ; Zhang et al. 2017 ; McMahon and Nelmes 2006 . At elevated pressures, materials undergo dramatic structural, electronic, and reactive changes, such as phase transitions, metallization, polymerization, and the formation of exotic compounds Loubeyre et al. 2020 ; Moog et al. 2021 . These phenomena are not only central to understanding planetary interiors, where pressures exceed hundreds of gigapascals, but also critical for synthesizing advanced materials with tailored properties, such as ultra-hard nano-materials or energy-dense polymorphs Grochala et al. 2007 ; Popov et al. 2002 ; Eremets et al. 2004 ; Tang et al. 2023 . For instance, the metallization of hydrogen under extreme compression Eremets and Troyan 2011 , the discovery of non-molecular CO 2 phases Iota et al. 1999 ; Giordano and Datchi 2007 , and the pressure-induced polymerization of organic molecules Li et al. 2021 exemplify how high-pressure environments enable the creation of novel states of matter with groundbreaking applications. Traditional computational approaches to investigating high-pressure chemistry have relied extensively on quantum mechanical methods, particularly density functional theory (DFT), which have provided invaluable fundamental insights into pressure-induced phenomena Liu et al. 2009 ; Sun et al. 2005 . Recent comprehensive reviews highlight the extensive application of ab initio simulations across the gigapascal pressure regime, demonstrating their crucial role in predicting and understanding high-pressure behavior Napiórkowska et al. 2023 ; Zeller et al. 2024 ; Rychkov 2020 . However, these quantum mechanical approaches suffer from computational costs that scale prohibitively with both system size and simulation timescales, strongly limiting their utility for exploring complex multi-component systems, large-scale structural transformations, or kinetically controlled processes under dynamic pressure conditions. The classical limitations in computational cost associated with ab initio methods have undergone a paradigmatic shift with the emergence of machine learning interatomic potentials (MLIPs), which bridge the accuracy of quantum mechanical approaches with the efficiency of classical force fields Behler 2016 ; Thompson et al. 2015 ; Graser et al. 2018 ; Schmidt et al. 2019 ; Unke et al. 2021 . With the rapid advancement of MLIPs, various computational frameworks were developed to integrate these methods Schütt et al. 2019 ; Wang et al. 2018 , while an increasing number of ab initio databases have begun to offer extensive datasets of high-quality and consistent DFT calculations Jain et al. 2013 ; Kirklin et al. 2015 ; Curtarolo et al. 2012 ; Schmidt et al. 2024 ; Scheidgen et al. 2023 . This combination of software and data led to the development of universal MLIPs (uMLIPs), i.e. foundation models that are capable, without further training, of describing a diverse materials space covering the whole periodic table. One of the first examples was M3GNet Chen and Ong 2022 , an extension of MEGNet that incorporated three-body interactions within its graph neural network architecture, thereby enriching the local chemical environment representation provided to the model. Following this development, numerous uMLIPs have been proposed Neumann et al. 2024 ; Park et al. 2024 ; Liao et al. 2024a ; Deng et al. 2023 ; Choudhary and DeCost 2021 ; Yang et al. 2024 , trained on increasingly larger and more diverse datasets, comprising up to millions of compounds and tens to hundreds of millions of DFT calculations Barroso-Luque et al. 2024 ; Schmidt et al. 2024 ; Deng et al. 2023 ; Peng et al. 2025 , significantly broadening the applicability and robustness of ML-driven materials materials science. Despite the impressive scale and diversity of existing datasets, many universal MLIPs still fail to generalize to specific regimes of interest, such as high pressures. In such cases, the lack of targeted training data can lead to significant inaccuracies, limiting the applicability of these models. In this context, we argue that the development and validation of uMLIPs specifically adapted to high-pressure environments is crucial. Such models have the potential to bridge the gap between quantum mechanical accuracy and computational efficiency, enabling reliable predictions in pressure regimes that remain challenging for conventional approaches. In this work, we tackle this problem by introducing a new dataset comprising 190 thousand compounds with a total of 32 million atomic single-point calculations under varying pressure. Using this dataset, we demonstrate the limitations of state-of-the-art uMLIPs in this regime and explore strategies for fine-tuning these models to restore their predictive capability under high-pressure conditions. II Results II.1 Dataset of calculations under pressure The database of ab initio calculations under pressure used in this work is summarized in Table 1 . Our dataset extends the existing Alexandria database Schmidt et al. 2024 , which has been used to train several state-of-the-art uMLIPs Neumann et al. 2024 and develop the largest training datasets freely available at the moment Yang et al. 2024 ; Barroso-Luque et al. 2024 . The 0 GPa dataset consists of 190 thousand distinct structures covering the complete periodic table. Each of these structures was then re-evaluated using DFT with the Perdew-Burke-Ernzerhof (PBE) exchange correlation functional Perdew et al. 1996 , consistently with the calculation setup used for the training data, across a predefined set of pressure values ranging from 0 GPa to 150 GPa. For each pressure value, full ionic relaxations were performed to obtain equilibrium volumes, atomic positions, and total energies. The resulting high-pressure dataset comprises a total of 32 million single-point DFT calculations. All calculations were performed using the same exchange-correlation functional and computational parameters as in the original Alexandria dataset to ensure consistency. The final dataset (see Table 1 ) includes, for each pressure and material, the relaxed crystal structure, total energy, atomic forces and stress tensors, together with the atomic configurations along the relaxation path. Table 1: Summary of the pressure dataset develop in this work. The missing materials at ambient pressure are due to older calculations for which we do not have the geometry optimization paths in Alexandria . The missing materials under pressure concern calculations that did not converge. Pressure (GPa) 0 25 50 75 100 125 150 Materials [x10 3 ] 162 188 190 188 188 188 187 Atomic configurations [x10 6 ] 3.2 4.0 6.0 4.0 4.7 5.1 5.2 As illustrated in Fig. 1 , the distribution of first-neighbor distances in the dataset systematically

Read the complete text