Cleavage Energy Benchmarking Dataset for Surface Stability Modeling with Universal Machine Learning Interatomic Potentials

Description

This dataset contains DFT-calculated cleavage energies for 36,718 surface structures along with predictions from 19 state-of-the-art universal machine learning interatomic potentials (uMLIPs).

Dataset Statistics

Total structures: 36,718

DFT functional: PBE

Number of MLIPs: 19

Date created: 2025-08-27

Column Descriptions

Original Dataset Columns (preserved from DFT calculations)

All original columns from the DFT database are preserved

`cleavage_energy_DFT`: Original DFT cleavage energy (eV/Ų)

`slab_energy_dft`, `bulk_energy_dft`: Raw energies from DFT in eV

`structure_slab`, `structure_bulk`: Pymatgen dictionary of atomistic structures

`mpid`: Materials Project ID

`pretty_formula`: Chemical formula

`miller`: Miller index of surface

`term`: Surface termination index (0, 1, ...)

`shift`: Surface shift parameter

`surface_area`: Surface area (Ų)

For other entries, refer to the explanation in the original database at /10.5281/zenodo.10381505

MLIP Prediction Columns (added)

`cleavage_energy_[MODEL]`: MLIP-predicted cleavage energy (eV/Ų) where [MODEL] is one of the 19 uMLIP model names

Models Included

1. AlphaNet-oma-v12. EquiformerV2-153M-OMAT243. EquiformerV2-153M-OMAT24-MP-sAlex4. EquiformerV2-153M-S2EF-OC20-All+MD5. GRACE-2L-MP-r66. GRACE-2L-OAM7. GRACE_2L-OMAT8. MACE-MP-0b39. MACE-MPA-010. MACE-OMAT-011. MatterSim-v1.0.0-5M12. SevenNet-mf-ompa13. eSEN-30M-MP14. eSEN-30M-OAM15. eSEN-30M-OMAT2416. orb-v217. orb-v3-conservative-inf-mpa18. orb-v3-conservative-inf-omat19. uma-m-1p1

How to Load the Database

By running the following Python code:

import jsonimport pandas as pdwith open("cleavage_energy_benchmark_DFT_MLIPs.json", "r") as file:    json_data = json.load(file)

metadata = pd.DataFrame.from_dict(json_data["metadata"])data = pd.DataFrame.from_records(json_data["data"])