Data and code for: Thermodynamically Consistent Machine Learning of CO2 Solubility in Molten Polymers with PC-SAFT

Data and toolbox-free GNU Octave/MATLAB code accompanying the article "Thermodynamically Consistent Machine Learning of CO2 Solubility in Molten Polymers with PC-SAFT" (I. Ushiki, D. Imai; submitted to Chemical Engineering Journal, 2026). Data (CC BY 4.0): a database of 632 swelling-corrected CO2 solubilities in 12 molten or rubbery polymer species from 20 sources with metadata; isotherm-wise binary interaction parameters (k_ij) for two PC-SAFT parameter sets and for the Sanchez-Lacombe equation of state; pure-component parameter tables with sources; point-wise leave-one-polymer-out predictions of all models; random-forest settings per fold; all tables and figure data of the article, including the predicted isotherms on fine pressure grids. Code (MIT): PC-SAFT and Sanchez-Lacombe solubility calculation, isotherm-wise k_ij regression, random forest, the leave-one-polymer-out evaluation of the hybrid, direct and residual-correction models and of the baselines, the nested descriptor selection, the fine-grid consistency analysis, the bounded pure-component regression and reproduction tests. REPRODUCTION.md states what each script reproduces from which files; under GNU Octave 8.4 the scripts reproduce the results of the article to the printed precision of the data files, apart from the last printed digit of a few equation-of-state solutions, which depends on the BLAS library.