Aprendizado de máquina aplicado à modelagem da permeação de gases em membranas visando a purificação de metano

Abstract

Machine learning models have gained prominence over the years for discovering patterns in data without the need for rule-based programming. The present work aimed to study the application of three machine learning algorithms (Random Forest, Support Vector Regression, and Artificial Neural Networks) to predict the performance of different membranes used in the permeation of pure gases. Additionally, the work aimed to investigate the main factors influencing this process. A data repository (called the Original Database) was built, comprising 1672 records referring to experimental work, from 42 different bibliographic references, focusing on those who have studied the permeation capacity of methane. For modeling, filtering was performed, resulting in a database consisting of 692 records from 19 different bibliographic references. These records were well-distributed in terms of membrane type and gas permeability. Furthermore, 3 different membrane types (Polymeric, Zeolite, and MOF) and 9 different feed stream gases (He, H2, CO2, O2, N2, CH4, ethylene, ethane, and SF6) were considered. Initially, 16 input attributes were used, with permeability as the target attribute. All machine learning models evaluated were adequate to predict the permeation performance of different membranes under various circumstances of the database. Different objective functions were investigated in the process of optimizing the model's hyperparameters. The main attributes that influenced the prediction were the Average Pore Size, the Feed Gas Kinetic Diameter, and the Specific Surface Area. Most other attributes (Average Thickness, Total Pore Volume, Age, Temperature, Pressure Difference, Sweep Gas Kinetic Diameter, Feed Gas Molar Mass, Feed Gas Polarizability, and Gas Fraction) can be considered of low relevance to model prediction. To simplify the modeling and remove the noise variables, some attributes were removed using the Random Forests (FA) model. In this process, six attributes were removed, and the prediction performance remained unchanged with a Mean Absolute Percent Error of 4.9 and 8.3% and an R2 of 0.98 and 0.96 for the training and test sets, respectively. Finally, the selectivities of four binary systems (He/CH4, H2/CH4, CO2/CH4, and N2/CH4) were estimated using the previously trained FA model, which also proved to be adequate with an R2 above 0.8 and an RMSE below 0.18 in all cases.