Datasets for 'Energy Underprediction from Symmetry in Machine-Learning Interatomic Potentials'

This collection contains datasets associated with paper "Energy Underprediction from Symmetry in Machine-Learning Interatomic Potentials". There are nine pretrained machine-learning interatomic potentials (MLIAPs), CHGNet, MACE, ORB-MPtrj, SevenNet, eqV2, eqV2-DeNS, MatterSim, eSEN, and eSEN-MPtrj, used for a test structure dataset with 153,235 structures queried from the Materials Project (database release v2023.11.1, licencse: CC BY 4.0, no GNoME structure involved). Details please find in the paper (preprint version). pandas_dataframe_pandas_id_strc_spg.json.gz: A pandas dataframe in json format can be read using Python `pandas.read_json`. There are three columns: material_id, json-serialized pymatgen.core.Structure object (need to be decoded by monty decoder), space group. This data was queried from the Materials Project using the SummaryDoc. pandas_dataframe_pandas_id_SpgSym_SpgNum_DoFs.json: A pandas dataframe in json format can be read using Python `pandas.read_json`. There are columns on material_id, space group number and various degree of freedom (DOF). The DOF computation is achieved by PyXtal API. pandas_dataframe_monty_dump_ThermoDoc_GGA_GGA+U_153243_retrieved_ehull.json.gz: A gzip-compressed json file for pandas dataframe that was dumped using `monty.serialization.dumpfn`, and has to be read using `monty.serialization.loadfn`. This file contain most of fields that were queried from the Materials Project using the ThermoDoc. The first few columns, 'material_id', 'chemsys', 'elements', 'nelements', 'nsites', 'composition', 'formula_pretty', 'entry_task_id', 'crystal_system', 'space_group_number', 'space_group_symbol', 'symprec_queried', 'energy_type', 'uncorrected_energy', 'correction', 'energy', 'uncorrected_energy_per_atom', 'correction_per_atom', 'energy_per_atom', 'formation_energy_per_atom', 'energy_above_hull', 'is_stable' are data recorded in the Materials Project. While, 'ef_uncorrected_retrieved', 'ehull_uncorrected_retrieved', 'ef_corrected_retrieved', 'ehull_corrected_retrieved' are retrieved formation energy and energy above hull using the uncorrected and corrected DFT energy. The retrieval was implemented using pymatgen phase diagram. \${mliap_name}_mp_relax.tar.gz: They are tar gzip files for ~13.8 millions calculations input and output files (structures, run logs, post-process data, etc.). Only CHGNet has fix-symmetry (symmetry constraint) calculations. The directory of the above \${mliap_name}_mp_relax.tar.gz looks like, \${mliap_name}/no-symmetry/batch{1...101}.tar.gz; each batch{1...100}.tar.gz would have 1,532 directories (one for each material_id); under each material_id, there are 2 or 3 directories for cell choices (as-queried, primitive, conventional); under each cell choice dir, there are three dirs denoting the relax type (no-relax, pos-relax, vc-relax), and the run logs and in/out structures are given under this dir. Each \${mliap_name}/no-symmetry/ also have refs_hull containing the MLIAP-calculated formation energy and enegy above the hull. The MLIAP-calculated energy columns follow this convention: {ef or ehull}_{cell choice}_{relax task} csv_structure_matcher.tar.gz: A separate tar.gz file provides the StructureMatcher output for DFT- and MLIAP-relaxed structure pairs, including match or not, RMSD, and max paired distance. All data are suggested to be read using Python.