Integrated Machine Learning with Molecular Simulation for Screening MOF Adsorbents toward High-Performance Separation of CO2/CH4

Abstract

The existing machine learning (ML) models exhibit insufficient predictive performance for CO2/CH4 separation on metal–organic frameworks (MOFs). Here, we leveraged a high-throughput screening strategy integrating molecular simulations with ML and successfully screened MOFs for CO2/CH4 adsorption separation with high performance. A novel four-layer ensemble ML was trained, and the structure-performance relationship was elucidated. Shapley Additive Explanations and Partial Dependence Plot revealed that the priority factors for CO2/CH4 separation were pore-limiting diameter, metal type, and accessible surface area. An optimal pore-limiting diameter of 3.7–4.2 Å was identified, and MOFs with helical channel topology structures were advantageous. Additionally, nitrogen-containing MOFs with above 30% oxygen synergistically enhance CO2 adsorption.