Machine-learning-enhanced symbolic regression for methane storage prediction in covalent organic frameworks

Covalent organic frameworks (COFs), recognized for their potential in vehicular methane storage, typically require comprehensive and resource-intensive evaluations for capacity determination. Herein, we introduce a novel approach that synergizes symbolic regression with machine learning, augmented by descriptive statistics, systematic feature selection, and high-throughput computational techniques, to develop predictive equations for COF methane storage capacities. By strategically selecting a concise set of three to seven features based on their interpretability and measurability, we have developed a spectrum of symbolic regression-based equations to meet diverse precision requirements in forecasting COF methane capacities. The fidelity of these equations varies: an equation with three variables achieves a root mean square error (RMSE) of $10\phantom{\rule{4pt}{0ex}}\mathrm{c}{\mathrm{m}}^{3}$ (standard temperature and pressure, STP) per $\phantom{\rule{0.16em}{0ex}}\mathrm{c}{\mathrm{m}}^{3}$ and a mean absolute percentage error (MAPE) of 5%, while expanding to seven variables fine-tunes the RMSE to $8.1\phantom{\rule{4pt}{0ex}}\mathrm{c}{\mathrm{m}}^{3}$ (STP) per $\mathrm{c}{\mathrm{m}}^{3}$ and the MAPE to 4.2%. Our structured approach ensures accurate predictions that correspond to the resolution of available data. Upon applying our most precise equation to an extensive dataset of $449\phantom{\rule{0.16em}{0ex}}468$ COFs, we identified numerous COFs with exceptional methane storage capacities. The adaptability of our equations to integrate new data broadens their applicability beyond COFs, making them versatile tools for optimizing methane storage across a range of nanoporous materials.