High-Throughput, Multiscale Computational Screening of Metal–Organic Frameworks for Xe/Kr Separation with Machine-Learned Parameters

Abstract

Accurate evaluation of adsorbent materials’ performance requires carrying out process simulations that take an analytical isotherm model as an input. In this work, we report a machine learning (ML) approach to approximate the saturation loading of nanoporous materials, an essential parameter for modeling the adsorption-based process simulation. Large-scale grand canonical Monte Carlo (GCMC) simulations were carried out to compute the single-component isotherms for Xe and Kr from the Computation-Ready Experimental Metal–Organic Framework (CoRE MOF) Database 2019. The generated data were used to fit the Langmuir model equation to obtain the saturation loading parameters, which were used as a basis to train several ML models. The performance of trained ML models was then compared with the pore volume-based approach, typically used in the literature, to approximate the saturation loading of the adsorbent material. Ideal vacuum swing adsorption (IVSA) simulations were carried out to screen a large number of MOFs. We found that the ML model better estimates the saturation loading from the curve fitting compared to the pore volume approach. Finally, we carried out high-fidelity vacuum swing adsorption simulations on 15 Xe-selective MOFs. While the IVSA approach provides quantitative information about the process performance metrics, we found that the commonly used performance metrics, such as Xe/Kr IAST selectivity, work as well as the shortcut methods (IVSA simulation) in ranking the adsorbent materials for Xe/Kr separation.

Engineering Machine Learning Features to Predict Adsorption of Carbon Dioxide and Nitrogen in Metal–Organic Frameworks

“…The pressure values corresponding to the intermediate loading regime for individual adsorbates are established by examining the adsorption isotherms reported in the previous studies. 43,68,69 As depicted in Table 5, our results reveal that, across all gas adsorption systems, the combination of geometric features, Henry's law constant, and energy-based features (model 4) consistently yields better results (R 2 > 0.95 for ethane, propane, krypton, and xenon; R 2 > 0.8 for carbon dioxide). Notably, the carbon dioxide adsorption system poses greater complexity due to its inclusion of Coulombic interactions compared to the others.…”

PACMAN: A Robust Partial Atomic Charge Predicter for Nanoporous Materials Based on Crystal Graph Convolution Networks

“…A Monte Carlo simulation was then run for 100,000 cycles at 298 K to compute the histogram of interaction energy. The force field parameters of the adsorbate molecules are listed in Table , and the force field parameters for the framework are available in Table S2 . All simulations were conducted using RASPA 2.0 software. , …”

“…There are now more than 10,000 porous MOFs reported in the literature to date and the number continues to grow. − Because of this, it is difficult to experimentally evaluate all of the MOFs for their potential application of interest. As such, well-established computational high-throughput approaches based on grand canonical Monte Carlo (GCMC) simulations are employed in the literature to accelerate the evaluation of nanoporous material’s performance for gas separation and storage applications. − For gas separation applications, an accurate atomistic model is critical to model the interactions between adsorbate and adsorbent materials because the relative interaction between the adsorbate gas and the adsorbent material is related to the selectivity of the adsorbent material.…”