Supplementary data for "Generalized Machine Learning Potentials for Water Adsorption in Al-Based Metal–Organic Frameworks Considering Local Flexibility"

This Zenodo includes five parts, DFT-Dataset: DFT-dataset-DZVP.zip and DFT-dataset-TZVP-zip Training dataset for MACE-AlMOFs-68 model Training dataset for MACE-AlMOFs-v1 model Test datasets including 420 MOFs Datasets to evaluate energy offsets for 352 MOFs and H2O. Train and test data for KMF-1 We provide two versions of the DFT datasets, both calculated using the PBE-D3(BJ) functional but utilizing different basis sets: DZVP-MOLOPT and TZV2P-MOLOPT. Please note that in our associated paper, we exclusively used the TZV2P-MOLOPT basis set to minimize Basis Set Superposition Error (BSSE). However, we have included both versions here to provide maximum flexibility for researchers looking to use this data to train their own models. Scripts: scripts.zip CP2K calculation script MACE training script MLP models: models.zip MACE-AlMOFs-68 model MACE-AlMOFs-v1 model We also provide two versions of finetuned MACE models using different basis sets: DZVP-MOLOPT and TZV2P-MOLOPT. Transferability indicator model: transfer_indicator.zip To help you use our MACE potentials effectively, we provide a simple tool to predict the transferability of our models. This folder contains a working example that demonstrates this process using the MACE-AlMOFs-68 model. In this folder, it is easy to reproduce all the related results in the paper. You can easily adapt this workflow for any other model. To do so, simply gather all the .cif files from the target model's training dataset and run our script to quickly evaluate its transferability. Dataset: dataset.zip The AlMOFs dataset: A collection of 420 unique aluminum-based metal-organic frameworks (AlMOFs) constructed from 439 distinct ligands. The CIF files for 439 isolated ligands KMF-1: A recently reported MOF not included in our original dataset, added here for external validation. Optimized structures and energy comparisons: The structures of the MOFs optimized using the MACE-AlMOFs-v1 model, alongside a comparison of their DFT-calculated energies against the original structures. ("opt" folder contains optimized structure. "original" folder contains the structure from the public dataset, the best folder means the structure with the lowest energy from these two datasets)