Reweighting free energy profiles between universal machine learning interatomic potentials for fast consensus building

# Dataset -- Reweighting Free Energies Supporting data for "Reweighting free energy profiles with machine learning interatomic potentials for fast consensus building," covering both systems studied: Li+ ion transport in a water-solvated zeolite, and the 5-hexenyl radical intramolecular cyclization (added in this revision as a second, small-molecule validation case).## Folder Contents ### `zeolite_structure/`CIF file of the studied system: Li+ ion in a water-solvated zeolite framework (1 Al, 1 Li, 55 H2O, 601 atoms total). ### `umbrella_sampling_data/`Output data from umbrella sampling (US) simulations used to construct the PMFs directly:- MACE-MATPES US simulation log files (54 umbrella windows, z = -6.61 to 6.64 Angstrom)- Pre-computed PMF from MACE-MP0 umbrella sampling (`pmf_april9_umb_mace-mp0.npz`) ### `data_540000_frames/`Processed frames used for free energy reweighting (10,000 downsampled frames x 54 windows = 540,000 frames total). Contains per-MLIP subdirectories with source and target potential energies for 7 target MLIPs:- MACE-MATPES, MACE-MP0-D3, MACE-MPA, MACE-finetuned, UMA, ORB, PET-MAD ### `data_parity_plots/`DFT and MLIP energies and forces for 1,100 structures used in the parity plots (8 MLIPs vs DFT PBE+D3 reference). POSCAR files for the DFT calculations are provided in `dft_li.zip`. ### `MACE-finetuned/`The finetuned MLIP developed in this work, along with the training data used. ### `MBAR_weights/`The MACE-MP0 unbiased MBAR weights obtained for the 540,000 downsampled frames. ### `hexenyl_parity_plot_data/`DFT (wB97M-V/def2-TZVP) and MLIP energies and gradient norms for 5,881 structures used in the parity plots (Fig. S4/S5 in the SI: 5 target MLIPs vs. DFT -- MACE-OFF, MACE-OMol, UMA-OMol, PET-OMol, PET-SPICE).- DFT reference energies and gradient norms (`dcd_frame_weights_matched.npz`, `dft_gradnorm_matched.npz`)- Per-MLIP energies and gradient norms on the same frames (`singlepoint_{mlip}.npz`, `gradnorm_{mlip}.npz`) ### `hexenyl_eabf_pmf_data/`Final PMFs from independent eABF (extended-system Adaptive Biasing Force) campaigns, used as the "source" and "independently simulated target" curves in the main-text hexenyl PMF figure.- MACE-OMol eABF PMF, 800 ps -- the reweighting source (`hexenyl_omol_eabf_final_800ps.npz`)- UMA-OMol eABF PMF, 800 ps -- the independently-simulated target reference (`hexenyl_uma_eabf_final_800ps.npz`) ### `hexenyl_reweighting_data/`Cross-MLIP reweighting inputs and results feeding the main-text hexenyl PMF figure, the hexenyl JSD/barrier-scatter figure (Fig. S11), and the region-wise-variance table (Table S1).- The 20%-downsampled MACE-OMol trajectory frames used as the reweighting source set (`maceomol_frames_for_umaomol_reweight_20pct.npz`)- Each target MLIP's energy on those same frames -- MACE-OFF, PET-OMol, PET-SPICE, UMA-OMol (`singlepoint_{mlip}_on_maceomol_traj_20pct.npz`)- Combined 6-curve comparison for MACE-OMol -> UMA-OMol: source/target/DFT/3 estimators (`hexenyl_6curve_zero_anchored.npz`)- MACE-OMol -> UMA-OMol reweighting PMF curves (`maceomol_to_umaomol_pmf_curves_20pct.npz`)- Energy-only reweighted PMFs to all 4 targets, from the MACE-OMol source (`energyonly_results_from_maceomol.npz`)- Pairwise JSD and barrier scatter data for the hexenyl system, Fig. S11 (`hexenyl_jsd_barrier_from_maceomol.npz`) ## Reference DFT data (hexenyl) The reference DFT trajectory, topology, and single-point energies the hexenyl system is built on are **not included here** -- they are taken from Dietschreit et al., *J. Chem. Phys.* **156**, 114105 (2022), and are available directly from the official tutorial repository:[github.com/learningmatter-mit/Tutorial_ActivationFreeEnergy](https://github.com/learningmatter-mit/Tutorial_ActivationFreeEnergy)(MIT license). Please cite the original paper if reusing that data.