Machine Learning-Assisted High-Throughput Screening of Covalent Organic Frameworks for Propane/Propylene Adsorption-Separation
Huan Li, Junjie Leng, Cailong Zhou, Hao Wu, Lichun Dong, Zemin Feng
Abstract
A high-throughput computational screening (HTCS) approach for identifying covalent organic frameworks (COFs) for C 3 H 8 /C 3 H 6 adsorption-separation was developed by combining machine learning (ML) and grand canonical Monte Carlo (GCMC) simulation. COFs from two experimental data sets were first evaluated using GCMC simulation to measure the performance metrics of C 3 H 8 adsorption capacity and C 3 H 8 /C 3 H 6 selectivity and to extract 28 geometric and chemical feature descriptors. The obtained values for the feature descriptors and performance metrics were used to train various ML models. The best-performing interpretable ML model was determined by using the AutoML method. This selected ML model was used to screen potential COF candidates from a hypothetical data set. The most promising COFs were then validated and analyzed using GCMC simulation, revealing that gravimetric surface area and void fraction (ϕ) are the two primary factors affecting the C 3 H 8 /C 3 H 6 adsorption-separation performance of COFs, and that the extra trees regressor model exhibits the highest accuracy in predicting COF performance. By using the developed HTCS approach, 20 top-performing COFs with strong adsorption-separation efficiency and regenerability were identified, and most of them are 3D COFs with qdl, clh, and qzl topologies. Finally, 4 of the top-performing COFs were experimentally synthesized, with the best achieving a C 3 H 8 uptake of 60.5 mL g –1 and a separation factor of 3.67. This study demonstrates that the ML-assisted HTCS approach can enable the efficient identification of high-performance COFs for C 3 H 8 /C 3 H 6 separation and establish a generalizable framework for high-throughput screening of functional materials.