Data and scripts accompanying Accelerating Mechanistic Understanding and Prediction of Chemical Selectivity with Foundation Machine-Learned Interatomic Potentials.
SELECT50 contains 50 pairs of competing transition structures for comparing calculated reaction selectivities. The repository also includes transition-state recovery calculations, two reaction landscapes, six larger molecular systems, and calibration against experimental nickel- and palladium-catalyzed selectivities.
Reference geometries were optimized at ωB97X-D3(BJ)/def2-TZVP. The electronic-energy reference is ωB97M-V/def2-TZVPD on those geometries. Comparison methods include eSEN-small, MACE-OMOL, UMA-M-1.1, UMA-S-1.2, g-xTB, r²SCAN-3c, PBE0-D3(BJ)/def2-SVP and ωB97X-D3(BJ)/def2-SVP. The reaction-landscape calculations use UMA-S-1.1.
| Dataset | Contents | Location |
|---|---|---|
| Core benchmark | Reference structures, metadata, energy differences and geometry comparisons | Data guide |
| Optimized geometries | Method-specific transition structures | XYZ files |
| Raw energies | Electronic and Gibbs energies, including MACE-OMOL thermochemistry | Energy tables |
| Transition-state recovery | 300 perturbed-geometry calculations | Calculations, summary |
| Reaction landscapes | Stationary points, scans, IRCs and pathway comparisons | Cycloaddition, palladium cyclization |
| Larger systems | Six transition-state pairs, DFT comparison structures, energies and timings | Structures and tables |
| Experimental calibration | Ni and Pd datasets with 100 training/test partitions each | Nickel, palladium |
Coordinates and geometric errors use ångströms; absolute energies use hartree and relative energies use kcal/mol. Major/minor roles are fixed by the reference electronic energies, with ΔΔE = E(minor) − E(major) and the same structure assignments for ΔΔG. See the reference definitions for roles and exclusions.
Clone the repository and check the data:
git clone https://github.com/apounder/SELECT50.git
cd SELECT50
python3 scripts/validate_data.pyThe validator and calibration analysis use only the Python standard library. Reproduce both calibration datasets from their recorded partitions with:
python3 calibration/refresh_saved_calibration.pyTo recalculate heavy-atom RMSDs and run all tests, install NumPy:
python3 -m pip install numpy
python3 scripts/calculate_heavy_atom_rmsd.py --output rmsd_recalculated.csv
python3 -m unittest discover -s testsThe RMSD guide describes atom matching and alignment. Calculation examples include a MACE-OMOL driver and ORCA/UMA templates for minima, transition states, scans and IRCs. The driver documents its calculation dependencies and command-line options; run it with -h for help.
Please cite the repository when using these data or scripts. Record the version with git rev-parse HEAD and include that commit hash with the repository URL.
Scientific data and documentation are licensed under CC BY 4.0; source code, tests and repository configuration are licensed under the MIT License. See LICENSE for scope. Separately identified third-party material retains its original license and attribution.
Contributions are welcome through GitHub issues and pull requests. See CONTRIBUTING.md.