(738f) Active Learning for Molecular Simulations of Adsorption in Metal-Organic Frameworks

Metal-organic frameworks (MOFs) are porous, crystalline materials comprised of inorganic nodes and organic linkers. Their porous nature along with the potential to tailor the chemistry of the pores, makes them very attractive for various adsorption applications. However, the sheer number of combinations of building blocks leads to limitless possible structures, raising questions as to the feasibility of characterizing them across the thermodynamic conditions needed to evaluate them in various applications. Herein I will present our efforts to accurately and efficiently determine adsorption in databases of MOFs using active learning (AL) techniques combined with molecular simulations of adsorption. We use AL to sequentially determine the next molecular simulation to be performed while simultaneously developing a surrogate model capable of describing adsorption in the feature space selected. We first established and validated the AL technique while varying temperature, pressure, and composition of the adsorbates, finding that AL can accurately describe the adsorption space with significantly less data. We then expand the analysis to include the MOF space and adsorbate space within the AL framework and find significant data savings. Lastly, we explore the use of reinforcement learning (RL) for adsorption AL campaigns in MOF databases.