Adsorption Prediction in Porous Media Through Active Machine Learning and Advanced Molecular Simulations: From Diverse Gas Mixtures to Water Vapor Separation

Adsorption and separation of molecular species from the gas phase is key to the technological progress as well as to resolve many of the emergent crises of the modern age. The most important property in designing such adsorption systems is the adsorption amount of the species on the adsorbate at the desired thermodynamic condition. In this direction, molecular simulations have proven to be an excellent tool to model such systems, besides laboratory-based studies. However, such simulations need many hours to weeks of computational time. Moreover, for learning the adsorption landscape for a range of thermodynamic features, it can be too resource intensive for the researcher. Further regular passive machine learning frameworks are best suited when a large dataset is available. Adsorbed solution theories such as Ideal-adsorbed solution theory also has limitations for several systems. These concerns can be overcome if an active learning strategy is applied to build machine learning model, which can balance training set size along with model performance. The present thesis addresses this challenge and explore how an active machine learning paradigm can be deployed build surrogate models for adsorption predictions. Different aspects of the active learning methodology such as initial training set design, kernel combinations, model regularization, performance scaling, are studied and analyzed. Most notably, the findings suggest that active learning protocol with gaussian process regression (GPR) models can build very accurate surrogate model for a small training set size. Algorithm performance scaling is also studied with feature space, namely mole fraction along with thermodynamic features such as pressure and temperature. Finally, a part of thesis applies molecular simulations to inform material design. First is material screening for designing sensor for dilute gas mixtures where the CoRE-2019 porous material database is explored. Most notably, materials with high selectivity are reported and their relationship with node and linker component of the top materials are analyzed. Another project deals with modeling water vapor adsorption and is motivated by how water adsorption can be enhanced on porous material. As water resources around the globe decline, atmospheric water vapor harvesting might provide a solution for easy and cheap access to fresh water. This research work probes: can electrostatic configurations, along with charge strength and pore size influence water vapor adsorption. A unique patterning is found where positive and negative charges are switched along the axial axis as well as alternated from a fully positive charged inner ring to an oppositely charged outer ring, which showed adsorption at much lower pressure and more uptake water volume than others. Analysis of energetic interactions, ordering, and hydrogen bonding reveals unique water relaxation phenomena for the alternating along that contributes to a high degree of order compared to other configurations.<p></p>