Explainable machine learning for carbon dioxide adsorption on porous carbon

Abstract

The ever-increasing carbon emission requires the development of advanced carbon capture and storage technology. As one of the most promising approaches, the adsorption of CO2 over porous carbon (PC) has been intensively studied due to the low cost, easy synthesis, and high adsorption capacity of PC. However, further improvement in its CO2 adsorption capacity is hindered by the unclear relationship between each physicochemical property and the adsorption capacity. Recently machine learning (ML) has achieved giant progress in predicting the adsorption capacity of PCs. However, due to the intrinsic “black box” property of ML, the quantitative relationship between each physicochemical property and corresponding adsorption capacity cannot be derived. Herein, first random forest was well trained to predict the CO2 adsorption capacity. Then, Shapley Additive Explanations, a powerful visualization tool, was adopted to provide insights into such an uncovered relationship. The results showed that textural property is more important than chemical composition in affecting CO2 uptake. Furthermore, SHAP indicated that high CO2 adsorption capacity appeared at low pressure, temperature of 298 K with specific surface area of 1100–1300 m2/g, micropore volume of 0.4–0.6 cm3/g, mesopore volume of < 0.1 cm2/g, carbon of 72–86 wt%, hydrogen of 1–4 wt%, oxygen of < 16 wt%, and nitrogen of < 1 wt%. This study sheds light on the understanding of the CO2 adsorption mechanism on PC and guides the design of next-generation high-performance PC for carbon capture and storage.