Membrane preparation assisted by integration of machine learning and response surface methodology for CO2 separation

Abstract

The separation of carbon dioxide (CO₂) is presented as a current challenge in the environment and energy sector. The primary reason for this is to control the emissions of this gas into the atmosphere and the upgrading of biomethane. In this context, the membrane separation technology seems to be a very sustainable promising tool for such tasks. This work presents a machine learning (ML) study, based on a database created from membrane preparation conditions and gas separation records from the literature, achieved for the CO₂/N₂ and CO₂/CH₄ mixtures using dense membranes of thermoplastic elastomer Pebax® 1657. A comparative analysis of three different ML models was carried out: multiple linear regression, decision tree and random forest. This last algorithm demonstrates the best performance in statistics terms of coefficient of determination and root mean square error. In addition, the combination of the ML random forest with a method based on the design of experiments with response surface methodology (RSM) allowed to identify the favorable conditions for the membrane synthesis, with the objective of enhancing the CO₂ separation performance. This resulted in prepared membranes in the laboratory considering the proposed conditions by RSM with CO₂ permeability and CO₂/X selectivity values of 115 Barrer and 43.5 and 132 Barrer and 16.4 for the CO₂/N₂ and CO₂/CH₄ mixtures, respectively, at 35 °C.