Abordagens de aprendizado de máquina para prever o equilíbrio de adsorção de gases leves em zeólitas, carvões ativados e redes metalorgânicas

Abstract

This study aimed to extract knowledge about the adsorption of light gases by microporous materials (zeolites, MOFs, and activated carbons) through data analysis and machine learning algorithms: K-nearest neighbors (KNN), Decision Trees (DT), and Support Vector Regression (SVR) of data reported in 22 articles published between 1974 and 2022. A database containing 3352 data points displaying the effects of 8 input variables (solid pore volume; solid surface area; experimental temperature and pressure; adsorption capacity measurement technique; gas polarizability, kinetic diameter, and molecular mass) on adsorption capacity was constructed. Box plots, histograms, bar charts, and scatter plots were applied (as part of exploratory data analysis) to determine how various input variables relate to each other and the performance variable. Additionally, KNN, DT, and SVR models were used for the regression of the adsorbed capacity data. Furthermore, the parametric study of these models allowed determining the relative importance of input variables and partial dependence among them to explore model interpretability (to deduce heuristics for high or low adsorption capacity). The exploratory data analysis found that pressure, temperature, gas polarizability, and molecular mass were the most significant variables affecting adsorption capacity. Additionally, combinations of input variables leading to high adsorption performance were revealed through the model analysis, which can be used as guidelines for future studies in this area.