Integrating Molecular Simulations and Machine Learning for Predicting HFC Adsorption in Zeolites
Year: 2026
Type: preprint
Abstract: This study re-audits and models a simulation-derived database for the adsorption of fluoromethane (R41), difluoromethane (R32), trifluoromethane (R23), and tetrafluoromethane (R14) in zeolites 13X, 3A, 4A, and 5A. The fully documented primary population contains 1,536 adsorption-loading observations (4 HFCs × 4 zeolites × 6 temperatures × 16 pressures) and 1,440 modeled isosteric-heat values; the 96 source heat zeros occurring at P = 0 were preserved in the audit data but were not used as evaluated heat targets. Ensemble regressors were compared under leakage-aware validation. Model selection was performed only within the development pool using five-fold isotherm-grouped cross-validation, followed by evaluation on a predefined balanced HFC–zeolite-pair holdout. Random row-level splitting was retained only as an interpolation benchmark. Histogram Gradient Boosting was selected for adsorption loading and achieved R² = 0.845, RMSE = 0.512 mmol g⁻¹, and MAE = 0.371 mmol g⁻¹ on the balanced pair holdout. Extra Trees was selected for isosteric heat and achieved R² = 0.885, RMSE = 0.617 kJ mol⁻¹, and MAE = 0.559 kJ mol⁻¹. Model attribution indicates that pressure and temperature are important for loading, whereas heat prediction depends more strongly on correlated framework and molecular descriptor families. Leave-one-HFC-out and leave-one-zeolite-out analyses show that the present surrogate is reliable mainly for interpolation within the studied chemical domain and is not yet validated for unseen HFC or zeolite identities. The revised workflow therefore provides a transparent and reproducible surrogate while defining its applicability limits.
Source: Research Square
Institution: King Fahd University of Petroleum and Minerals
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204180916
