Machine learning for design of thin-film nanocomposite membranes

In this study, a novel machine learning approach is proposed for estimation of the permeate flux and foulant rejection in nanocomposite filtration membranes. Nine independent variables are fed to artificial neural networks (ANNs) including support, nanoparticles concentration, concentration of organic phase trimesoyl chloride (TMC) in-n-hexane (TMC in n-hexane), operation pressure, contact angle, thin layer thickness, location of the nanoparticles (NPs), post-treatment temperature and duration, and permeate flux and foulant rejection were derived as the outputs of the ANNs. The proposed method was evaluated on two datasets across training, validation and test datasets, and an unseen dataset. 2250 different initial weights and number of the neurons in the hidden layer for the proposed ANN models were considered and compared to find the optimized ANN models. The mean squared error (MSE) and coefficient of determination (R²) were employed to select the best 20 ANN models for further analysis. The proposed ANN models resulted in accurate estimates for both permeate flux and foulant rejection with R² of 0.9958 and 0.9412 in all data included in the training, validation and test datasets and R² of 0.9938 and 0.9811 in unseen dataset, respectively. In addition, results of sensitivity analysis revealed that post treatment temperature and contact angle were found the most important input variables for estimation of permeate flux and foulant rejection. The proposed method can provide valuable insights for formulating permeate flux and foulant rejection and considering the effects of each experimental condition on nanocomposite filtration membranes without doing real experiments, which is time-consuming and expensive.