Integrating molecular simulations with multilayer perceptron neural networks to predict the CH4/H2 adsorption and separation of MTV-MOFs

Abstract

The secure, efficient, and cost-effective storage and separation of methane and hydrogen are crucial for their large-scale industrial applications. This work evaluates the performance of 10,995 multivariate metal-organic frameworks (MTV-MOFs) as CH₄/H₂ adsorption and separation media through molecular simulations and neural network modeling. Firstly, the structural parameters and CH₄/H₂ adsorption and separation properties of MTV-MOFs were obtained by molecular simulation, and the relationships between them were studied. The results reveal that the key structural descriptors and their optimal ranges are accessible surface area (4000–5000 m²/g), largest cavity diameter (11–14 Å), pore-limiting diameter (10–12 Å), and porosity (∼0.8). A multilayer perceptron neural networks (MLP) model was subsequently constructed to predict the CH₄/H₂ adsorption and separation properties based on the results of molecular simulations. The results of MLP model are not only comparable to those of traditional grand canonical Monte Carlo (GCMC) simulation, but also exhibit a significant enhancement in screening efficiency. This demonstrates the excellent generalization ability and robustness of the designed MLP model. We hope that this study will provide some theoretical references for the efficient screening of MTV-MOFs for CH₄/H₂ adsorption and separation applications.