Data-driven modeling using machine learning to investigate the desulfurization performance by zeolitic adsorbents

Abstract

This work introduces an experimentally validated, data-driven machine learning (ML) framework for predicting the adsorptive desulfurization (ADS) performance of zeolite-based materials. A curated dataset of 700 entries was compiled from diverse sources, incorporating key structural and operational parameters such as Brunauer–Emmett–Teller (BET) surface area, total pore volume (TPV), temperature, contact time, and molecular weight of sulfur compounds (MW-S). Seven ML models were developed and compared, with Extra Trees Regressor (ETR) achieving the best performance (R² = 0.9979, MAE = 0.0308), followed by Random Forest (RF) (R² = 0.9932, MAE = 0.0524). Feature importance analysis and shapley additive explanations (SHAP) identified molecular weight and BET surface area as the most influential descriptors. For better interpretability and generalizability, the zeolite type was excluded as an input feature and replaced by physicochemical properties. Furthermore, the top-performing model was integrated with a genetic algorithm (GA) to optimize operating conditions, resulting in a predicted maximum adsorption capacity of 131.63 mg S/g. Model robustness was also confirmed using an independent test set. Overall, this study provides a reliable and interpretable framework for accelerating ADS system design and can be extended to other adsorption-based separation processes.