Experimental characterization and machine learning optimization of polymer nanocomposite membranes for carbon capture systems

Abstract

The study aimed to characterize the CO2 capture capabilities of Polyacrylonitrile (PAN) nanocomposite membranes by reinforcing them with multi-walled carbon nanotubes (MWCNT) and silica (SiO2). These membranes were made using the electrospinning manufacturing method. The nanoparticles were functionalized using Gum Arabic (GA) to improve nanoparticle distribution, which further improved the capture efficiency. The morphological techniques were used to examine the nanoparticle structures after functionalization to optimize the functionalization parameters. Experimental results showed that increasing nanoparticle concentrations enhanced CO2 permeability while maintaining stable N2 permeability, resulting in favourable CO2/N2 selectivity ratios. The 4 wt. % MWCNTs nanocomposite membrane exhibited the best CO2/N2 separation with a CO2 permeability of 289.4 Barrer and a selectivity of 6.3, while the 7 wt.% SiO2 nanocomposite membrane achieved a CO2 permeability of 325 Barrer and a selectivity of 7. These results indicated significant CO2 permeability and selectivity improvements compared to pure PAN membrane. The Maxwell mathematical model was used for validation, and the experimental results exceeded the predicted values, possibly due to well-dispersed nanoparticles and functional groups. Based on the CO2 capturing results from the previous experiments, a second experiment study focused on enhancing the CO2 capture capabilities of PAN membranes by modifying them with polyethyleneimine (PEI), a polymer with high CO2 absorption capacity. PAN was modified with three weight fractions of PEI (25%, 40%, and 50%) and then reinforced with various weight fractions of MWCNT, SiO2, and alumina (Al2O3) nanoparticles. The reinforced PAN-PEI nanocomposite membranes were produced using an electrospinning technique. The morphological characterization techniques confirmed that the PEI has effectively modified PAN polymer, which has improved the distribution of nanoparticles within the nanocomposite membranes. Gas permeation tests revealed that the 40 wt.% PEI-modified membrane achieved the best CO2/N2 separation, with a CO2 permeability of 509.4 Barrer and selectivity of 7.4. The PAN with 40% PEI was then reinforced with 1, 4, 7, and 10 wt. % nanoparticles and the highest performance was observed with 7 wt.% Al2O3, resulting in a CO2 permeability of 849 Barrer and selectivity of 9.6. The results were validated using mathematical models (Resistant Model Approach and Effective Medium Approach), confirming the effectiveness of nanoparticle infusion for CO2 separation. Finally, this research developed and applied three machine learning (ML) techniques, Deep Neural Networks (DNN), Random Forest (RF), and XGBoost models, to analyze the CO2 permeability and CO2/N2 selectivity of nanocomposite membranes. The datasets for CO2/N2 separation were sourced from our experimental results and the published experimental data available in the literature. Key performance metrics such as polymer type, nanofiller type, size, loading amount, membrane surface area and thickness, temperature, and feed pressure were analyzed. Feature importance plots provided insights into the most influential parameters for the material design. The study involved hyperparameter tuning of the DNN, RF, and XGBoost models to achieve optimal performance. Each model was tested using literature data and combined experimental and literature data to validate the models and assess the impact of incorporating experimental data. Performance metrics were evaluated to establish the research's credibility and generalizability. The XGBoost optimized ML model achieved the best prediction performance, with R2 values of 0.93 for CO2 permeability and 0.83 for CO2/N2 selectivity, highlighting the effectiveness of using ML for optimizing nanocomposite membranes.