Interpretable artificial neural network modeling of forward osmosis performance in cellulose triacetate membranes

초록

Forward osmosis (FO) is a promising membrane technology for sustainable water treatment; however, permeate flux is governed by complex nonlinear interactions among multiple operating parameters, making experimental optimization challenging. In this study, 440 data points were compiled from the literature, comprising key operational inputs and permeate flux as output. An AI-enabled artificial neural network (ANN) model was developed to predict FO performance in CTA membrane systems based on solution concentrations, temperatures, draw-solution molecular weight, and hydrodynamic conditions. The ANN model demonstrated good predictive performance, with a test R2 of 0.7545 and a Pearson correlation coefficient (r) of 0.8686. Based on the optimized model, a user-friendly GUI-based decision-support tool was developed to enable rapid prediction and process optimization without requiring machine learning expertise. Single- and two-variable sensitivity analyses were performed to evaluate nonlinear dependencies and interaction effects. At the same time, an index of relative importance (IRI) was used to quantify variable influence under low, medium, and high flux conditions. ANNguided optimization predicted maximum permeate flux values of 80.20 LMH (AL-FS) and 80.69 LMH (AL-DS), which are higher than the maximum experimental value (72.0 LMH) contained in the compiled dataset. These results highlight the potential of the ANN model to identify promising operating conditions for future experimental investigation. However, the predicted optimum operating conditions have not been experimentally validated and should therefore be interpreted as model-based predictions. The proposed framework provides a practical AI-enabled approach for prediction, optimization, and decision support in FO systems, contributing to efficient and sustainable water management. The developed graphical user interface (GUI) allows researchers, engineers, and practitioners with little or no machine learning expertise to easily utilize the ANN model for rapid prediction, process evaluation, and operating-condition optimization.

키워드