Machine learning-enhanced prediction of optimal metal-organic frameworks for O$_2$/N$_2$ gas separation

Abstract

In this study, a machine learning model was developed utilizing the Computable Experimental Metal-Organic Framework (CoREMOF) dataset to predict O$_2$ and N$_2$ gas selectivity. Traditionally, gas selectivity prediction methods have focused on predicting gas adsorption and gas diffusivity values separately through machine learning and calculating gas selectivity based on these predictions. However, this approach is susceptible to cumulative errors and often deviates significantly from actual values. To address this challenge, a single integrated model was introduced. New machine learning features, such as PSD% (pore size distribution percentage), Q_{st,N2}^0(heat of adsorption of nitrogen), and the Henry's coefficient of the gas, were incorporated to enhance the accuracy of the final gas selectivity predictions. Additionally, an analysis was conducted to examine the interplay between the characteristics used in machine learning training and the resultant gas selectivity. This study advocates for the direct prediction of gas selectivity using a unified machine learning model as the most appropriate approach for predicting gas separation characteristics.