[Paper Review] MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures

Han Yang, Chenxi Hu+20 more|arXiv (Cornell University)|May 8, 2024

Machine Learning in Materials Science74 citations

TL;DR

MatterSim is a universal deep learning atomistic model trained with active learning from first-principles data, predicting energies, forces, and stresses across elements, temperatures (0–5000 K), and pressures (0–1000 GPa) with near-first-principles accuracy and enabling zero-shot MD and phase diagram predictions.

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%.

Motivation & Objective

• Address the need for accurate, fast predictions of materials properties across the periodic table under realistic temperature and pressure conditions.
• Develop a universal ML-based interatomic model trained via active learning on large-scale first-principles data.
• Enable zero-shot simulations and continuous learning to improve generalizability and data efficiency.

Proposed method

• Use deep graph neural networks (M3GNet and Graphormer backbones) to model atomistic systems under 0–5000 K and 0–1000 GPa.
• Employ active learning with a first-principles supervisor (PBE+U) and an ensemble uncertainty monitor to curate diverse, high-coverage datasets (~17M structures).
• Provide zero-shot machine-learning force fields (MLFF) for energies, forces, and stresses with high accuracy (MAE ~36 meV/atom on MPF-TP).
• Benchmark against public datasets and new benchmarks (MPF-TP, Random-TP, MatBench Discovery) for energetics, phonons, and mechanical properties.
• Demonstrate fine-tuning/fine-tuning to adapt to different theory levels (e.g., rev-PBE0-D3 for water) with data-efficient customization.

Experimental results

Research questions

• RQ1 Can MatterSim reliably predict energies, forces, and stresses for materials across nearly all elements under wide temperature and pressure ranges?
• RQ2 How does active learning and diverse data coverage improve generalization beyond relaxation-trajectory datasets?
• RQ3 What is MatterSim’s accuracy in predicting thermodynamic properties (e.g., Gibbs free energy) and phase diagrams under finite T and P?
• RQ4 How effective is MatterSim for zero-shot MD, materials discovery, and end-to-end property prediction?

Key findings

• MatterSim achieves up to ten-fold accuracy improvement over prior universal MLFFs on high-temperature/high-pressure datasets (MPF-TP and Random-TP).
• Zero-shot predictions for energies, forces, and stresses attain MAEs around 36 meV/atom on MPF-TP, with near-first-principles fidelity in free energy predictions (sub-10 meV/atom up to 1000 K) and MAE 15 meV/atom against experiments.
• Active learning yields ~17 million labeled structures spanning 0–5000 K and 0–1000 GPa, with broader chemical and structural coverage than relaxation-based databases; discovered 16,399 structures on or below the convex hull, including 1,974 new structures.
• MatterSim enables rapid materials discovery, showing 5,213 RSS-derived materials on the hull (71% of the hull contribution) and high-throughput screening capability.
• Phonon prediction MAE ~0.87 THz against PhononDB; bulk modulus MAE ~2.47 GPa at 0 K and ~0.97 GPa for temperature dependence up to 1000 K; phase boundary MgO B1–B2 in agreement with experiments and first-principles data across wide T–P ranges.
• Gibbs free energy predictions achieve sub-10 meV/atom accuracy up to 1000 K, with MAE ~15 meV/atom versus experiments; phase diagrams demonstrated under QHA for MgO and Si.