Artificial neural network-based prediction of PVA thin-film composite membrane fabrication using statistically screened input parameters

Abstract The performance of polymeric thin-film composite (TFC) membranes is strongly influenced by fabrication parameters associated with both the polymeric porous support and the upper selective layer. The complex interactions among these parameters reduce the effectiveness of traditional experimental analyses. In this study, artificial neural network (ANN) models were developed to correlate key fabrication parameters of the support and polyvinyl alcohol (PVA) selective layer with the performance of PVA-based TFC membranes. Under the limited data size condition, a radial basis function (RBF) network was employed with leave-one-out cross-validation (LOOCV) to optimize the model and improve its reliability. The models demonstrated high predictive accuracy, with R 2 values of 0.9923 for permeability and 0.9986 for salt rejection. A reduced-dimension ANN (R-ANN) was further implemented by statistically screening the input variables, which improved model efficiency while maintaining high accuracy. Analysis of key fabrication parameters revealed that cross-linking the PVA layer with a cross-linking agent and heat curing critically influenced salt rejection, which exhibited complex nonlinear behavior in relation to other fabrication parameters. The proposed ANN successfully identified these nonlinear trends by considering the co-effects of multiple parameters. Although predictive capability is limited for unseen conditions, ANN outputs were consistent across the examined parameter range.