Permeation Properties of Polymer Nanocomposite Membranes for <scp> CO ₂ </scp> Separation: An Experimental and Machine Learning‐Based Optimization Approach

ABSTRACT The study aimed to optimize the gas permeation properties of polymer nanocomposite membranes through a combination of experimental investigations and machine learning (ML) techniques. Initially, polyacrylonitrile (PAN) and PAN‐polyethyleneimine (PAN‐PEI) nanocomposite membranes were fabricated and experimentally characterized to determine their CO 2 permeability and CO 2 /N 2 selectivity. The experimental data were supplemented with values reported in the literature for various nanocomposite membranes, creating a comprehensive dataset that included polymer type, nanofiller type, size, loading amount, membrane surface area and thickness, temperature, and feed pressure. Three supervised machine learning models: deep neural networks (DNN), random forest (RF), and extreme gradient boosting (XGBoost), were developed to predict and optimize membrane gas separation performance. Feature importance analysis was performed to identify the most influential parameters affecting CO 2 permeability and CO 2 /N 2 selectivity. Hyperparameter tuning was carried out using grid search with cross‐validation (GridSearchCV) and a genetic algorithm (GA) for the RF and XGBoost models, while the DNN model was optimized using both the Adam and AdamW optimizers with manual (trial‐and‐error) hyperparameter tuning. Model training utilized two datasets: (i) literature data alone and (ii) a combined dataset integrating in‐house experimental and literature data, to evaluate the impact of experimental augmentation on model performance and generalizability. Model evaluation employed metrics such as the coefficient of determination ( R 2 ) and root mean squared error (RMSE) to validate predictive accuracy. Among the models, the GA‐optimized XGBoost model achieved the highest accuracy, with R 2 values of 0.95 for CO 2 permeability and 0.85 for CO 2 /N 2 selectivity, alongside a training duration of 11–24 s, surpassing both the RF and DNN models. Subsequently, the trained XGBoost‐GA model was combined with Bayesian optimization and genetic algorithm strategies to identify the optimal process parameters that maximize CO 2 permeability and CO 2 /N 2 selectivity. The GA‐based optimization produced a maximum CO 2 permeability of 21,130 Barrer and a CO 2 /N 2 selectivity of 43, compared to the Bayesian optimization results of 17,246 Barrer and a selectivity of 40.