Machine learning based prediction and optimization of thin film nanocomposite membranes for organic solvent nanofiltration

In this study, machine learning was used to form prediction models for thin film nanocomposite (TFN) organic solvent nanofiltration (OSN) membrane performance evaluation in terms of relative permeability (RP) and relative selectivity (RS). Twenty references including 9252 data points were collected to form four different models: linear, support vector machine (SVM), boosted tree (BT), and artificial neural network (ANN). Among the four models, BT exhibited optimal prediction accuracy in terms of root mean square error (RMSE) and coefficient of determination (R2) values for membrane RP (RMSE: 0.295, R2: 0.918) and RS (RMSE: 0.053, R2: 0.849) performance prediction. Parameter contribution analysis indicated that nanoparticle loading, amine concentration, chloride concentration, water contact angle, solvent viscosity, and molar volume are the main parameters influencing RP performance. For RS performance, nanoparticle loading, amine concentration, chloride concentration, and solute molecular weight play important roles. Partial dependence analysis indicated that the optimal conditions for TFN-OSN membrane fabrication are nanoparticle loading less than 5 wt%, the amine concentration around 2 wt%, and the chloride concentration around 0.15 wt%. In addition, membrane with super-hydrophilic or super-hydrophobic surface property exhibited higher RP performance based on different feed solvent types. Overall, this work introduces new ways both for TFN-OSN membrane performance prediction and for higher performance membrane design and development.

Automated summary

In this study, machine learning models were used to predict the performance of thin film nanocomposite (TFN) organic solvent nanofiltration (OSN) membranes in terms of relative permeability (RP) and relative selectivity (RS). Among the four models tested (linear, support vector machine, boosted tree, and artificial neural network), the boosted tree model demonstrated the highest prediction accuracy for membrane RP and RS. Parameter analysis identified the key factors influencing RP performance as nanoparticle loading, amine concentration, chloride concentration, water contact angle, solvent viscosity, and molar volume. For RS performance, nanoparticle loading, amine concentration, chloride concentration, and solute molecular weight played important roles. The study also highlighted the optimal conditions for TFN-OSN membrane fabrication and the impact of membrane surface properties on RP performance. Overall, this work contributes new insights into TFN-OSN membrane performance prediction and membrane design.