Machine Learning-Aided Computational Study of Metal–Organic Frameworks for Sour Gas Sweetening

Abstract

Nanoporous materials, such as metal–organic frameworks (MOFs), have shown great potential as adsorbents for separations in a wide variety of energy- or environment-related applications. One promising application is sour gas sweetening; a raw natural gas contains small amounts of H2S that can be detrimental to the efficient utilization of the energy source. However, the large database of nanoporous materials has made the discovery of optimum materials significantly demanding. While molecular simulations can play a complementary role in facilitating the materials search, their brute-force utilization still requires a vast amount of computational resources. In this study, we incorporate a machine learning algorithm with structural and chemistry descriptors as inputs for efficient screening. Specifically, the random forest regressor, which can also be useful for elucidating structure–property relationships, is employed. For reliable predictions with machine learning, the choices of features play considerably important roles. In addition to commonly adopted geometrical and chemical features, we propose and incorporate a set of newly designed features for training the model. These new features represent preferential binding sites of open-metal sites and dense framework atoms on the pore surface. Our analysis shows that the inclusion of the newly designed features greatly improves the machine learning performance. Our work can pave the way for the future design of nanoporous materials for sour gas sweetening. These newly designed features can also be used for the development of machine learning models for other applications, especially those involving molecules with strong dipole and/or quadruple moments, such as carbon capture.

Monte Carlo simulations for water adsorption in porous materials: Best practices and new insights

“…The structure AFOVIB has only two adsorption sites per unit cell with very strong binding of approximately À83 kJ/mol. This strong binding is most likely due to the openness of the metal atom site which is quantified to be À17.5 per the definition of Cho et al, 77 while the interactions at other locations in the structure are much more unfavorable. Given that the ratio of strong to weak adsorption sites is small, the hostadsorbate interaction contribution to the system energy is nonconstant as a function of loading and in fact weakens considerably as the relatively unfavorable sites are occupied.…”

Recommendation System to Predict Missing Adsorption Properties of Nanoporous Materials

“…Machine learning plays an important role in the discovery and deployment of NPMs [26][27][28][29][30][31][32]. Supervised machine learning models have been widely used to predict the adsorption properties of NPMs [33][34][35][36][37][38][39][40][41] from vectors of hand-crafted structural features [42,43] or from a graph representation [44]. Unsupervised machine learning methods have been used to embed NPMs into a lowdimensional "material space" [45] and cluster together NPMs with similar structures [46][47][48].…”

Nanoporous Material Recognition via 3D Convolutional Neural Networks: Prediction of Adsorption Properties

“…Nanoporous materials, such as zeolites and metal–organic frameworks, have been investigated for various energy- and environmental-related applications including methane adsorption, carbon capture, or sour gas sweetening. − However, the total number of nanoporous materials can be as much as hundreds of thousands, which makes material discovery via purely trial-and-error or brute-force-based searching extremely challenging. − To this end, predictions with machine learning models have recently drawn considerable attention for facilitating material discovery. − For these studies, a few geometrical features of the materials are computed and are used to train machine learning models to predict adsorption properties. − These features, such as crystal density, fractional free volume, surface area, and pore diameters, represent the structure geometry, have clear physical meaning, and are relevant to the adsorbate–adsorbent interaction and adsorption properties. − Although some useful predictors have made the predictions decently successful, they may impose human bias and require human expertise on the exploration and selection of the features. Furthermore, these studies require multiple steps of preprocessing, feature construction, prediction, and feature selection, requiring human labor to proceed step-by-step for modeling.…”

“…Because of the environmental burdens brought about by fossil fuels, natural gas or landfill gas can be great alternatives . However, these energy sources may be contaminated with acidic gases. ,− Furthermore, methane has very low volumetric energy density, as low as one-third of that of gasoline, which makes transportation and storage of gases difficult . For this purpose, nanoporous materials can be beneficial for methane separation and storage.…”

“…It is important to stress at this point that the CNN approach introduces a new way of making efficient predictions of adsorption properties for nanoporous materials. Previously, geometrical information, ranging from simple properties such as pore diameter, surface area, void fraction, and others to rather complicated mathematical techniques to encode them, such as persistent homology, have been designed and constructed by human expertise to serve as features in machine learning algorithms. ,,,− In the CNN approach, the structure is directly employed as the input and the CNN self-learns useful geometrical information most relevant to adsorption properties.…”