(669e) Surface Area Prediction of Nanoporous Materials Using Machine Learning Methods

The surface area of a porous material plays a critical role in a broad range of science and engineering applications, as many chemical and biological events occur near or on the surface. Surface areas of porous materials, such as metal-organic frameworks (MOFs), are commonly characterized using the Brunauer-Emmett-Teller (BET) method. The BET areas depend significantly on the linear region (the region of the isotherm) selected for the BET analysis, which is chosen via four consistency criteria as per the state-of-the-art. However, it has been shown that even with the fulfillment of these criteria, the BET method does not always provide an accurate measure of the monolayer area, often overestimating it significantly, especially for high-surface-area MOFs. In this work, we extend this argument to systematically show that even if we use the best possible consistency criteria, the BET method cannot always predict the monolayer areas accurately. To this end, we propose, for the first time, a data-driven approach to accurately predict the surface areas of MOFs. Machine learning (ML) is employed to train models based on adsorption isotherm features of over 300 diverse MOF structures to predict a benchmark measure of the surface area known as the true monolayer area. The true monolayer area is computed by counting the number of adsorbate molecules present in the first layer from the framework in molecular simulation snapshots. We demonstrate that the ML-based methods can consistently predict true monolayer areas significantly better than the BET method, especially for high surface area materials, showing great promise as a more accurate alternative to the BET method in the structural characterization of porous materials.