Data and scripts for "Δ-HGPR: A Hierarchical Gaussian Process Framework for Active-Learning Δ-Potentials"

Abstract

Data and scripts underlying "Δ-HGPR: A Hierarchical Gaussian Process Framework for Active-Learning Δ-Potentials" (T. Nakajima, submitted to J. Chem. Theory Comput.). This version corresponds to the revised manuscript and supersedes version 1. The archive contains the inputs, labeled data, trained difference potentials, trajectories, run and analysis scripts, and analysis outputs underlying every table and figure of the revised manuscript and its Supporting Information. README.md maps each directory to the corresponding tables, figures, and sections; MANIFEST.tsv lists every item with its description, and SHA256SUMS gives checksums of all files. Contents production_ _ .tar.gz (8 files): malonaldehyde (PM7, ANI-2x, MACE-MH-1 bases; 6 × 10 ps) and H5O2+ (ANI-2x, MACE-MH-1 bases; 6 × 3 ps), with DFT-only, base-only, and Δ-HGPR-MD production trajectories, final models, active-learning logs, aggregated metrics, the matched many-body-only control for H5O2+ with ANI-2x, and NTChem input files labeled.tar.gz: labeled structures (positions and DFT − base force differences) of every active-learning round lr_toy_regression.tar.gz, lr_toy_md_*.tar.gz: long-range toy system (Na+/Cl− ion pair with two water molecules), with fixed-set regression tests and MD with a weak ion tether si.tar.gz: data for the Supporting Information (joint vs sequential fit, water toy system, trajectory-length convergence, f_std calibration, free-energy-perturbation model error, force verification tests) code.tar.gz: run, aggregation, statistics, and verification scripts Changes from version 1 The analytical derivative of the SOAP descriptor now includes the full contribution of the displacement of the central atom; all SOAP-based results were recomputed with the corrected code. The ANI model used throughout is ANI-2x. The coefficients of the two terms are determined by a single joint fit in all production calculations. Malonaldehyde production trajectories are 10 ps (six per condition). New: long-range toy system, matched many-body-only control for H5O2+, free-energy-perturbation diagnostics, implementation tests. Software: the Δ-HGPR program itself is not included in this record. Reference calculations were performed with NTChem (B3LYP/def2-SVP for malonaldehyde, B3LYP-D4/def2-SVP for H5O2+); NTChem input files are included and output files are not. Base calculators are ANI-2x (TorchANI), MACE-MH-1 with the matpes_r2scan head (the model file is distributed by the MACE developers and is not included), and PM7 (MOPAC).