MPtrj

The Materials Project contains a large set of relaxation data across its entire database. These relaxation "trajectories," along with single-point data, for the PBE GGA and GGA+U calculations are stored in the tasks collection of Materials Project. Deng et al. collated the energies, forces, stresses, structures, and on-site magnetic moments (when available) from this ionic step data, and used it to train a graph neural network, CHGNet [1], for the potential energy surface of most of the periodic table.

The collected dataset is called MPtrj in the CHGNet [1]. Subsets of MPtrj were used in the earlier universal machine learning interatomic potentials (MLIPs) M3GNet [2] and later ones such as MACE-MP-0 [3].

[WIP] The original MPtrj dataset is hosted on MPContribs (explorer) (bulk download) with a parquet format bulk download. An example notebook for working with parquet data is also available.

References

[1] B. Deng, P. Zhong, K.J. Jun, J. Riebesell, K. Han, C. J. Bartel, and G. Ceder. "CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling". Nature Mach. Intell., vol. 5, pp. 1031–1041, yr. 2023. (DOI) (original figshare)

[2] C. Chen and S.P. Ong, "A universal graph deep learning interatomic potential for the periodic table". Nature Comput. Sci., vol. 2, pp. 718–728, yr. 2022. (DOI)

[3] I. Batatia et al., "A foundation model for atomistic materials chemistry". arXiv:2401.00096v3, yr. 2025. (DOI)