Incorporating Mechanistic Insights into Structure-Property Modeling of Metal-Organic Frameworks for H2 and CH4 Adsorption: A CGCNN Approach

Abstract

In this paper, we introduce a novel approach for understanding gas adsorption mechanisms in Metal-Organic Frameworks (MOFs) by combining a tailored Crystal Graph Convolutional Neural Network (AG-CGCNN) with game theory through Shapley Additive Explanations (SHAP). Our technique identifies the root causes of suboptimal performance in crystal structures through the analysis of softmax probabilities of performance outcomes, alongside the marginal contributions of crystalline and geometric features to these outcomes. Specifically, our approach discerns whether poor performance stems from inefficient atom-atom interactions indicating suboptimal gas molecule capture or from the subpar distribution of geometric characteristics, including surface area, pore limiting diameter (PLD), or largest cavity diameter (LCD). Additionally, the modeling framework adopted in this study clearly captures how intrinsic atomic features facilitate excess adsorption while volumetric spatial characteristics account for bulk absorption in complex porous MOF networks. Furthermore, we have introduced an innovative technique to quantitatively evaluate the amount of both excess gas absorbed within site interactions and bulk gas within the void spaces in a single MOF unit cell using a modified CGCNN regression model and SHAP. Additionally, we highlight the excellent performance of three 100 K-trained AG-CGCNN and demonstrate how its architecture can be used to determine the efficiency of a MOF gas capture system.