Domain-bridging datasets and DFT results for "Optimizing Cross-Domain Transfer for Universal Machine Learning Interatomic Potentials"

Updated (2026. 02. 06)SevenNet-Omni is a universal machine learning interatomic potential (uMLIP) trained based on the SevenNet-MF architecture, using 15 different open datasets across material domains of molecules, crystals, and surface systems. This item includes the SevenNet-Omni checkpoint, benchmark results, the DBS dataset, and the DFT calculation results described in the SevenNet-Omni paper. The checkpoint can be downloaded from github asset or directly from the SevenNet package.benchmark_results/alex_scanComparison of SCAN formation energies of Alexandria databasebenchmark_results/LPSC_MDComparison of force accuracies on LPSC r2SCAN MD trajectoriesDBS/dbs_total.extxyzThe domain-bridging set used for training the SevenNet-Omni.(All other databases comprising the COSMOS dataset are open-source.)Geometries are sampled from OC20, OC22, MatPES, ODAC23, OMol25, and QCML databases.Single-point calculations were performed using VASP, with pseudopotentials selected to ensure compatibility with the Materials Project database.The field 'db_label' informs the source database, sampling type, and index, formatted as {db_src}_{type}_{idx}.During training, entries labeled as 'random' and 'spare' were split into training and validation sets with a ratio of 9:1, while all 'elem' entries were used in the training.DFT_resultsContains in-house DFT calculation results presented in the main figures of the paper.Refer to DFT_results/README for detailed descriptions.Potential usageUse the latest version of SevenNet. SevenNet-Omni is supported from v0.12.0.Please visit the SevenNet repository and also the documentation for any detailed instructions forHow to use ASE calculatorHow to run LAMMPS MD simulation via TorchScript or ML-IAP packageHow to install tensor product accelerators such as FashTP or CuEquivarianceHow to enable GPU-accelerated D3 calculator for ASE and LAMMPS.