Carbon Capture Phenomena in Metal-Organic Frameworks with Neural Network Potentials

Abstract

Diamine-appended metal-organic frameworks are ultraporous materials exhibiting selective and cooperative CO_{2} adsorption mechanisms, leading to chemically tunable step-shaped isotherms and isobars that enable a large fraction of their full CO_{2} capacity to be captured or released with a modest change in temperature or pressure. Although progress has been made in elucidating carbon capture phenomena in this system, its thermal properties are poorly understood. Here, we develop density-functional-theory-derived neural network potentials for amine-appended Mg_{2}(dobpdc), metal-organic frameworks made up of Mg cations and dobpdc [dobpdc^{4−} = 4,4’-dihydroxy-(1,1’-biphenyl)-3,3’-dicarboxylic acid] linkers. These potentials are constructed with an active learning approach where their accuracy and transferability are improved through iterative generation of training datasets based on molecular dynamics. Our potentials predict adsorption energy, mechanical properties, vibrational and thermal properties with and without CO_{2}, reaching ab initio accuracy at a fraction of the computational cost. We compute the temperature-dependent heat capacity and lattice thermal expansion of Mg_{2}(dobpdc), with and without amine appendages and with and without CO_{2}, quantitatively capturing measured trends where available and explaining their molecular origins. Furthermore, we show that these potentials can be incorporated into a simulated annealing approach to identify starting structures for energy minimization needed to compute accurate binding enthalpies of these and other complex systems. Our density-functional-theory-derived neural network potentials explain the temperature evolution of the structure and properties of an important class of metal-organic frameworks at molecular length scales, while providing a necessary foundation for future studies of the chemical dynamics of carbon capture.