Application of machine learning (ML) methods for the determination of the gas adsorption capacities of nanomaterials, such as metal–organic frameworks (MOF), has been extensively investigated over the past few years as a computationally efficient alternative to time-consuming and computationally demanding molecular simulations. Depending on the thermodynamic conditions and the adsorbed gas, ML has been found to provide very accurate results. In this work, we go one step further and we introduce chemical intuition in our descriptors by using the “type” of the atoms in the structure, instead of the previously used building blocks, to account for the chemical character of the MOF. ML predictions for the methane and carbon dioxide adsorption capacities of several tens of thousands of hypothetical MOFs are evaluated at various thermodynamic conditions using the random forest algorithm. For all cases examined, the use of atom types instead of building blocks leads to significantly more accurate predictions, while the number of MOFs needed for the training of the ML algorithm in order to achieve a specified accuracy can be reduced by an order of magnitude. More importantly, since practically there are an unlimited number of building blocks that materials can be made of but a limited number of atom types, the proposed approach is more general and can be considered as universal. The universality and transferability was proved by predicting the adsorption properties of a completely different family of materials after the training of the ML algorithm in MOFs.

In the present study, we propose a new set of descriptors that, along with a few structural features of nanoporous materials, can be used by machine learning algorithms for accurate predictions of the gas uptake capacities of these materials. All new descriptors closely resemble the helium atom void fraction of the material framework. However, instead of a helium atom, a particle with an appropriately defined van der Waals radius is used. The set of void fractions of a small number of these particles is found to be sufficient to characterize uniquely the structure of each material and to account for the most important topological features. We assess the accuracy of our approach by examining the predictions of the random forest algorithm in the relative small dataset of the computation-ready, experimental (CoRE) MOFs (∼4700 structures) that have been experimentally synthesized and whose geometrical/structural features have been accurately calculated before. We first performed grand canonical Monte Carlo simulations to accurately determine their methane uptake capacities at two different temperatures (280 and 298 K) and three different pressures (1, 5.8, and 65 bar). Despite the high chemical and structural diversity of the CoRE MOFs, it was found that the use of the proposed descriptors significantly improves the accuracy of the machine learning algorithm, particularly at low pressures, compared to the predictions made based solely on the rest structural features. More importantly, the algorithm can be easily adapted for other types of nanoporous materials beyond MOFs. Convergence of the predictions was reached even for small training set sizes compared to what was found in previous works using the hypothetical MOF database.

No abstract

No abstract

In the present study, we propose a new set of descriptors, appropriate for machine learning (ML) methods, aiming to predict accurately the gas adsorption capacities of nanoporous materials. The present work focuses on systems with nonnegligible electrostatic interactions between the materials' framework and the guest gas. For that, the CO 2 , H 2 , and H 2 S gases are examined. The present approach is a generalization of our recent development for guest gases with no electrostatic interactions, such as CH 4 . For both types of systems, as ML descriptors we consider the adsorption probabilities by the materials' framework of a small number of probe atoms with different van der Waals diameters. After examination and evaluation of various numerical schemes, probe atoms that carry in their centers an electric dipole are found to be the most appropriate for systems with electrostatic interactions. The accuracy of the present approach is assessed by comparing the ML predictions with a data set of reference results obtained after performing grand canonical Monte Carlo (GCMC) simulations. More specifically, the CO 2 , H 2 , and H 2 S adsorption capacities of the computation-ready, experimental (CoRE) MOFs at several different thermodynamic conditions are considered. The low computational cost for the calculation of the proposed set of ML descriptors allows the screening of very large databases.