Harnessing confinement effect and interpretable machine learning to predict alkane diffusion in zeolite catalysts

Diffusion within zeolites is a critical determinant of their performance in catalysis and separations. Nevertheless, mechanistic insights remain largely system-specific, and a universal framework connecting topological features to diffusion properties is still lacking. Here, we address this challenge by combining high-throughput molecular simulations across approximately 100,000 zeolitic frameworks from the database of zeolite structures with interpretable machine learning. Guided by confinement-diffusion mechanisms, we propose a set of topology-informed descriptors (including tortuosity and cross-sectional variance) and construct a highly accurate, transferable model that quantifies the contribution of each structural factor. Our findings indicate that although pore-limiting diameters primarily facilitate diffusion, channel tortuosity and cross-sectional heterogeneity act as major transport barriers. Transfer learning further demonstrates that the framework trained on methane diffusion can be efficiently extended to other small organic molecules (e.g., ethane, ethene, and methanol). This study provides a large, systematically curated dataset for zeolite diffusion and establishes a machine learning-based framework to elucidate confinement-governed mass transport, thereby accelerating the rational design of tailored zeolite materials.