PALIRS2: data for "Optimization of active learning strategies for infrared spectra prediction in catalysis"

This repository provides the data supporting the publication:<PAPER DOI ONCE AVAILABLE> This work enhances PALIRS (Bhatia et al., npj Comput. Mater. 11, 324, 2025) via transfer learning and extends it from C0–C2 to C0–C5 molecules.It includes the trained MLIPs and dipole models, the initial geometries of the 30 molecules in the training set and 9 molecules outside it, the datasets generated through the active learning process, and the reference and production trajectories obtained with AIMD and MLMD, from which all IR spectra reported in the paper are computed. For details on model reproduction, please visit: https://gitlab.com/cest-group/PALIRS Abbreviations MLIP - machine-learned interatomic potentialAIMD - ab initio molecular dynamicsMLMD - machine-learning-driven molecular dynamicsIR - infraredDFT - density functional theoryFS - from-scratch (training strategy)TL - transfer learning (multihead fine-tuning training strategy)TL+S - hybrid strategy, retrained from scratch on the dataset generated during TL Contents Sizes below are uncompressed. Models.zip (381 MB) - trained MACE MLIPs and dipole models: the final FS and TL ensembles, the hybrid TL+S model, the FS and TL dipole models, and the initial MLIPs used to start active learning. Geometries.zip (312 KB) - FHI-aims geometry.in input files for the 30 molecules in the training set and the 9 molecules outside it, both DFT-optimized and non-optimized. AL_data.zip (85 MB) - the three initial datasets (5400 configurations each), the final datasets after 15 active learning iterations for the FS (9729 configurations) and TL (9728 configurations) strategies, and the T:C3-C5 test set (600 configurations). Extended XYZ format. AIMD_IR_run.zip (7.9 GB) - 54 ps AIMD reference trajectories at 300 K for the 30 molecules in the training set, computed with FHI-aims. AIMD_IR_run_Temp.zip (3.0 GB) - the same at 100, 500, 700 and 900 K for the three selected molecules in the training set. AIMD_IR_run_Transf.zip (813 MB) - the same for the 9 molecules outside the training set. MLMD_IR_run.zip (18 GB) - 55 ps machine-learning MD production trajectories and predicted dipole time series at 300 K for the 30 molecules in the training set, using both the from-scratch and transfer-learned models. MLMD_IR_run_Temp.zip (6.8 GB) - the same at 100, 500, 700 and 900 K for the same three selected molecules as in case of AIMD, for both FS and TL models. MLMD_IR_run_Transf.zip (6.2 GB) - the same for the 9 molecules outside the training set and methanol, for both FS and TL models. MLMD_IR_run_MACE4IR_MACEOFF.zip (2.0 GB) - the same for three of the molecules outside the training set, driven by the MACE4IR and MACE-OFF23 foundation models.