Data-Driven Modeling and Optimization of Polymeric Membranes for CO2 Separation: A Machine Learning Perspective

Abstract

Polymeric membranes have become a promising technique for the reduction of greenhouse gas emissions, as they provide energy-efficient, scalable, and simple processes for the separation of carbon dioxide (CO2). Nevertheless, the design of high-performance polymer membranes is a challenging task since permeability and selectivity tend to be at odds. Recently, machine learning (ML) methods have been widely studied in order to accelerate membrane material discovery through the learning of the structure–property relationships of materials in experimental and computational data sets. This study presents a thorough survey of recent ML approaches used for gas separation material property prediction and optimization in polymer membranes. The performance, strengths, and limitations of a wide array of algorithms namely Random Forest, Support Vector Regression, Gaussian Process Regression, Deep Neural Networks, XGBoost, CatBoost are critically examined in relation to membrane properties. In addition, several challenges associated with ML applications, such as small amounts of data, model interpretability, uncertainty quantification, and generalization issues are discussed. Building on the aforementioned insights, an architecture design for solving membrane material selection problems is proposed, combining techniques such as molecular fingerprint representations, ensembles learning, interpretable ML, and uncertainty-aware screening strategies.