Final dataset generated using the active learning scheme and the test set used to evaluate the models - PALIRS

This repository provides data supporting the publication:https://doi.org/10.1038/s41524-025-01827-8

It includes the initial geometries of 24 organic molecules (both DFT-optimized and non-optimized), the datasets generated through the active learning (AL) process for training machine-learned interatomic potentials (MLIPs), and the test sets used to evaluate their performance and transferability.

For details on model reproduction, please visit:https://gitlab.com/cest-group/PALIRS

Contents

Total_AL_data.xyz – 16,485 structures generated through active learning cycles for MLIP training. Extxyz format.Test_AL_data.xyz – 480 clustered structures from MLMD trajectories for MLIP validation. Extxyz format.Test_data_transferability.xyz – 400 clustered structures for assessing MLIP transferability. Extxyz format.WO_Opt_initial_str.zip – Non-optimized initial geometries of 24 organic molecules. Zip archive containing separate FHI-aims geometry.in files.Opt_initial_str.zip – DFT-optimized geometries of the same 24 molecules. Zip archive containing separate FHI-aims geometry.in files.