Combining Interpretable Machine Learning and Molecular Simulation to Advance the Discovery of COF-Based Membranes for Acid Gas Separation

Abstract

Membrane technology can effectively remove acidic gases (H 2 S and CO 2 ) from natural gas. Covalent organic frames (COFs) have been widely used as membrane materials due to their large pore size and pure organic properties. This work combines machine learning (ML) and molecular simulation (MS) to develop a method for rapidly screening and discovering high-performance COF-based membranes. The ML model is first trained on MS data, using the structural and chemical features obtained from 20 calculations as inputs. Characteristic contributions were obtained through interpretable analytical models, and nearly 70,000 COFs were quickly screened. Finally, the top 10 high-performance COFs were selected by MS under the mixed gas condition, and the properties of the mixed matrix membranes (MMMs) obtained by combining them with six polymers were analyzed. The results show that the highest acid gas permeability of COF-based membranes reaches 8 × 10 5 Barrer, and the void fraction is the key factor determining the separation performance. The top COFs serve as effective fillers to enhance the performance of polymer membranes, which surpass the capabilities of existing MOF fillers. This paper provides an efficient and rapid method for the discovery of COF-based membranes for natural gas deacidification.