CO2 from flue gas is central to mitigating fossil-fuel-derived emissions, where adsorbent performance directly dictates process energy efficiency and process cost. Although machine learning (ML) has emerged as a powerful tool for accelerating adsorbent discovery, its predictive accuracy is fundamentally limited by the physical reliability of the underlying training data, a manifestation of the "garbage in, garbage out" (GIGO) problem. Most existing CO2 adsorption databases rely on Lennard-Jones (LJ) force fields, whose deficiencies in describing CO2-CO2 and CO2-framework interactions, particularly at high pressures, introduce systematic bias into the ML models. To address this, we developed a physically accurate van der Waals force field based on an Exp-PE potential and constructed a high-fidelity CO2 adsorption database. Building on this data set, we introduce quadrupole-responsive descriptors that explicitly capture the anisotropic electrostatics of CO2, leading to improved ML predictive accuracy. This framework identifies high-performing COF/MOF adsorbents, including COF-50 (ΔNCO2 = 13.58 mol/kg) and COF-364 (ΔNCO2 = 12.43 mol/kg), whose working capacities exceed those of current reported porous materials.
Lang et al. (Mon,) studied this question.