A Foundational Potential Energy Surface Dataset for Materials: High-Quality MatPES Dataset Drives Universal Machine Learning Interatomic Potentials
In March 2025, a multi-institutional consortium released the MatPES dataset, containing approximately 400,000 structures carefully sampled from 281 million molecular dynamics snapshots, covering 16 billion atomic environments. Universal machine learning interatomic potentials (UMLIPs) trained on this dataset achieve performance comparable to or surpassing models trained on larger datasets on benchmarks for equilibrium, near-equilibrium, and molecular dynamics properties. A high-fidelity potential energy surface dataset based on the r²SCAN functional was also released, improving the description of interatomic bonding.
The MatPES dataset emphasizes data quality over quantity, requiring only 400,000 structures to train UMLIPs comparable to those trained on larger datasets. This work challenges the current paradigm relying on DFT relaxation data by sampling non-equilibrium structures from molecular dynamics trajectories, enhancing the generalization of potentials to non-equilibrium states. The introduction of the r²SCAN dataset further improves the accuracy of bonding descriptions.
The MatPES dataset samples approximately 400K structures from 281M molecular dynamics snapshots, covering 16B atomic environments. The sampling strategy ensures structural diversity, including equilibrium and near-equilibrium configurations. UMLIPs trained on this dataset (e.g., MACE, CHGNet) perform excellently on multiple benchmarks (e.g., formation energy, lattice constants, elastic constants, phonon spectra), comparable to or better than models trained on larger datasets (e.g., MPtrj). The r²SCAN dataset provides a more accurate potential energy surface than PBE, particularly improving descriptions of covalent bonds and weak interactions. The paper does not provide specific numerical comparisons but claims state-of-the-art performance on multiple properties.
This dataset has significant implications for materials discovery and design. High-quality UMLIPs can accelerate the screening and property prediction of new materials, reducing reliance on DFT calculations. Industries such as pharmaceuticals, energy, and electronics can benefit from faster materials simulations. The open-source dataset promotes community collaboration and model standardization.
It is recommended that materials simulation software companies (e.g., Materials Design, Schrödinger) integrate UMLIPs trained on MatPES. Investment opportunities exist in AI+ materials startups. In engineering, the MatPES dataset can be used for pre-training, followed by fine-tuning for specific material systems to improve prediction accuracy.
Key areas to watch: 1) Coverage of more elements and compounds in the dataset; 2) Performance of UMLIPs on complex defects and interfaces; 3) Dataset expansion combined with active learning; 4) Validation of the r²SCAN dataset in industrial applications; 5) Model transferability and uncertainty quantification.