MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures

TL;DR

MatterSim addresses the challenge of predicting materials properties across the entire periodic table under diverse temperatures and pressures by combining deep graph networks with active learning and large-scale first-principles supervision. It delivers a zero-shot atomistic emulator capable of energies, forces, and stresses across broad chemical and thermodynamic space, and extends to Gibbs free energy and phase-diagram predictions with near-first-principles accuracy. The framework supports active learning, uncertainty quantification, and fine-tuning to higher levels of theory, enabling data-efficient customization and end-to-end property prediction, demonstrated on MatBench and liquid water scenarios. Together, these capabilities yield state-of-the-art performance on discovery tasks and provide a scalable platform for continual learning and domain-specific tailoring in materials design.

Abstract

Accurate and fast prediction of materials properties is central to the digital transformation of materials design. However, the vast design space and diverse operating conditions pose significant challenges for accurately modeling arbitrary material candidates and forecasting their properties. We present MatterSim, a deep learning model actively learned from large-scale first-principles computations, for efficient atomistic simulations at first-principles level and accurate prediction of broad material properties across the periodic table, spanning temperatures from 0 to 5000 K and pressures up to 1000 GPa. Out-of-the-box, the model serves as a machine learning force field, and shows remarkable capabilities not only in predicting ground-state material structures and energetics, but also in simulating their behavior under realistic temperatures and pressures, signifying an up to ten-fold enhancement in precision compared to the prior best-in-class. This enables MatterSim to compute materials' lattice dynamics, mechanical and thermodynamic properties, and beyond, to an accuracy comparable with first-principles methods. Specifically, MatterSim predicts Gibbs free energies for a wide range of inorganic solids with near-first-principles accuracy and achieves a 15 meV/atom resolution for temperatures up to 1000K compared with experiments. This opens an opportunity to predict experimental phase diagrams of materials at minimal computational cost. Moreover, MatterSim also serves as a platform for continuous learning and customization by integrating domain-specific data. The model can be fine-tuned for atomistic simulations at a desired level of theory or for direct structure-to-property predictions, achieving high data efficiency with a reduction in data requirements by up to 97%.

Figures (53)

• Figure 1: MatterSim is a deep learning atomistic model for predicting materials properties with high predictive accuracy across chemical elements, temperatures and pressures, enabling a wide range of applicability and functionality.
• Figure 2: MatterSim is developed on an enriched materials space.(a) A data explorer employed in MatterSim for generating datasets covering wide potential energy surface; Histogram of the stress (GPa) and effective temperature (K) of (b) the generated materials in this work, (c) the MPF2021 dataset and (d) the Alexandria dataset. (e) Comparative performance metrics of MatterSim across six tasks: energy prediction on MPF-TP and random-TP datasets, phonon properties including max frequency and density of states (DOS), Bulk Modulus, and inverse F1 score in MatBench-Discovery leaderboard. Lower scores indicating superior performance for all tasks. Refer to main text and supplementary information for task details.
• Figure 3: MatterSim as a zero-shot emulator empowering materials discovery.(a) and (b) are the contribution of each dataset to the combined convex hull formed by Alexandria-MP-ICSD dataset (see text) and RSS-generated materials; (c) Elementwise appearance distributionriebesell_pymatviz_2022 of the 852 RSS-generated materials found be to on the combined convex hull formed by the Alexandria-MP-ICSD and RSS-generated materials. The materials containing H, Si, N, Sb, O, S, Se, Te, F, Cl, Br, I are removed due to potential issue with how anion corrections are implemented in Materials Project when applied to hypothetical materialsgithubstrict_anionsOption. (d) exhibits examples of materials found to be lower than the Alexandria-MP-ICSD hull, with the corresponding space group in the parentheses.
• Figure 4: MatterSim as a zero-shot emulator for predicting lattice dynamics and thermodynamic properties.(a), (c) and (e) are parity plots of maximum phonon frequency, bulk modulus and computed free energy difference between 0 and 300K, respectively; (b) is the phonon dispersion of ZnSe, (d) is the temperature dependent bulk modulus of AlN and (f) is the predicted B1-B2 phase boundary of MgO with a comparison to first-principles studies and experimental measurements.
• Figure 5: MatterSim as a zero-shot molecular dynamics (MD) engine.(a) Examples materials selected for running molecular dynamics; (b) the success rate of molecular dynamics with increasing temperature and pressure for various categories of materials; (c) analysis of the stopping temperature and pressure of the molecular dynamics trajectories, with the temperature and pressure profile of the trajectory shown in the inset; (d) the potential energy of a MOF material under increasing temperature and NVT ensemble, with the inset being the radial distribution function of the Ag-Ag atoms $g_\mathrm{Ag-Ag}(r)$ at 0, 200 and 400ps of the trajectory; and (e) the potential energy of a bulk inorganic material under increasing pressure and NPT ensemble, with the inset being the radial distribution function of the Mg-Mg atoms $g_\mathrm{Mg-Mg}(r)$ at 0, 200 and 400ps of the trajectory.