PubChemQC COSMO σ-Profile Database with Recalibrated COSMO-SAC Parameters

This dataset provides the first large-scale open-source σ-profile database for the COSMO-SAC activity coefficient model, comprising 361,583 organic molecules drawn from the PubChemQC dataset. σ-profiles are computed using the open-source PySCF package at the BP86/def2-SVP level of theory in the conductor-screening (COSMO) limit, with surface-charge averaging following the Hsieh, Sandler, and Lin (2010) procedure that splits each profile into three sub-profiles (non-hydrogen-bonding, hydroxyl, and other hydrogen-bonding components). At 361,583 molecules, this database is approximately 162 times larger than the NIST cCOSMO σ-profile database of Bell et al. (2020), 6.8 times larger than the CHAOS database, and 206 times larger than the LVPP database, making it the largest open-source σ-profile compilation released to date. The deposit includes both the verbatim individual .sigma text files (sigma_profiles_main_raw.tar.gz, 255 MB) and a derived analysis-convenience parquet (sigma_profiles_main.parquet, 117 MB) containing the same 361,583 profiles with metadata. Integrity has been verified through row-count checks, deduplication, 10-CID round-trip comparisons, and 5-member tarball spot-checks (all checks passed, see integrity_results.json). Recalibrated COSMO-SAC exchange-energy parameters are provided in sequential_final_params.json. These were obtained by a chemistry-stratified, three-stage sequential nonlinear least-squares refit against 24,522 infinite-dilution activity coefficient (γ∞) measurements from the NIST/TRC ThermoData Engine, partitioned at the binary-pair level. Validation on a held-out test set of 9,748 γ∞ records (1,559 binary pairs), 12 binary vapor-liquid equilibrium systems (256 P-x-y data points), and 22 (solute, solvent) pairs for solid-liquid equilibrium (1,684 data points) shows that the open-source pipeline matches or slightly exceeds the established DMol3-based reference: γ∞ RMSE 0.898 (vs 0.961 default), VLE mean pressure deviation 3.12% (vs 3.92%), and SLE log10(x_sat) RMSE 0.412 (vs 0.458). The accompanying source code, validation scripts, and quick-start examples will be distributed in an associated open-source GitHub repository (MIT License) to be released alongside the paper. The GitHub URL will be added to this metadata record once the repository is public. File format: 51 σ-bins from −0.025 to +0.025 e/Ų in steps of 0.001 e/Ų, three sub-profiles concatenated for a total of 153 data rows per molecule. σ_hb = 0.0084 e/Ų, f_decay = 3.57 (Hsieh averaging). See README.md and MANIFEST.txt for full format specifications. Intended applications: classical COSMO-SAC phase-equilibrium prediction (activity coefficients, VLE, SLE, partition behavior), training data for foundation models and hybrid physics–machine learning approaches to chemical property prediction, and thermodynamics-informed process-engineering workflows including solvent selection, separation design, and crystallization process development. Citation: This dataset accompanies a manuscript currently under review at AIChE Journal. Please cite both this Zenodo deposit (DOI: 10.5281/zenodo.20672192) and the paper once published. License: Data are released under Creative Commons Attribution 4.0 International (CC-BY-4.0). The associated source code (in the GitHub repository) is released under the MIT License.