Machine Learning-Guided Equations for Super-Fast Prediction of Methane Storage Capacities of COFs

Abstract

Covalent organic framework (COF) is a prominent class of nanoporous materials under consideration for vehicular methane storage. However, evaluating a COF for its methane capacity involves multiple experimental or computational steps, which is expensive and time consuming. Consequently, the discovery of high-capacity COFs for methane storage is very slow. Here we developed equations for super-fast prediction of deliverable methane capacities of COFs from a small number (3 to 7) of physically meaningful and measurable crystallographic features. We provided a set of equations with different fidelities for on-demand predictions based on the accessibility of crystallographic features. We found that an equation with only three crystallographic primary features, as variables, can predict deliverable capacities of 84,800 COFs with a root-mean-square error (RMSE) of 10 cm3 (standard temperature and pressure, STP) cm-3 and mean absolute percentage error (MAPE) of 5%. However, the highest fidelity equation developed here contains seven crystallographic primary features of COFs with RMSE and MAPE of 8.1 cm3 (STP) cm-3 and 4.2%, respectively. With that, we predicted methane storage capacities of 468,343 previously unexplored COFs using the highest fidelity equation and identified several hundred promising candidates with record-setting performance. CUBE_PBB_BA2, a hypothetical COF not yet synthesized, sets the new record of balancing gravimetric (0.396 g g-1) and volumetric (221 cm3 (STP) cm-3) deliverable methane storage capacities under the pressure swing between 65 and 5.8 bar at 298K. Also, 3D-HNU5, a previously synthesized COF, has shown the potential to achieve the gravimetric and volumetric methane storage U.S. Department of Energy target (0.5 g g-1 and 315 cm3 (STP) cm-3) simultaneously with uptakes of 0.755 g g-1 and 334 cm3 (STP) cm-3 at 100 bar/270 K.