Ensemble Learning of Partition Functions for the Prediction of Thermodynamic Properties of Adsorption in Metal–Organic and Covalent Organic Frameworks

We develop several ensemble learning models to predict the partition function of fluids adsorbed in nanoporous materials. The partition functions so obtained allow rapid access to all thermodynamic properties of adsorption and provide an assessment of the performance of these materials over a wide range of conditions, without performing additional simulations. We examine different types of applications related to the environment and energy, including carbon capture and storage in IRMOF-1, hydrogen storage in COF-102, as well as the separation of light hydrocarbons, methane, and ethane in COF-102 and COF-108. A diversity ensemble learning approach, in which predictions obtained from neural networks with different architectures are averaged, is found to provide the most accurate predictions. Comparisons with simulation results obtained using grand-canonical Monte Carlo simulations and flat-histogram methods establish the versatility of the ensemble learning model. Most notably, the new insights brought by the ensemble learning predictions include the rapid determination of the selectivity and of the desorption free energy over a wide range of operating conditions (pressure or mole fraction of the gas feed). This, in turn, provides a means to assess rapidly the performance of nanoporous materials for separation purposes, as well as of the energetic cost to regenerate the adsorbent.