(285a) Prediction of Adsorption Thermodynamics in MOFs and Cofs Via Ensemble Learning

We discuss how ensemble learning (EL) allows for a rapid prediction of the partition function for fluids adsorbed in nanoporous materials. For this purpose, we carry out flat histogram simulations for the adsorption process and train EL models on the simulated datasets. Once partition functions are predicted by the EL models, all thermodynamic properties of adsorption can be rapidly determined for a wide range of conditions, without any additional simulations. We apply various EL approaches to carbon capture and storage, hydrogen storage and gas separation in MOFs and COFs. In particular, we find that the diversity ensemble learning approach, obtained from neural networks with different types of architectures, provides the most accurate predictions, when compared to the experiment and to Expanded Wang-Landau simulation results. The EL models so obtained give access to the materials selectivity in the case of the dsorption of mixtures and to a number of key thermodynamic properties, such as the desorption free energy which measures the energetic cost to regenerate the adsorbent.