Comparative modelling of membrane filtration performance using response surface methodology, artificial neural network, and random forest regression: Implications for food and bioprocessing applications

Abstract

Membrane filtration plays a critical role in the food and bioprocessing industries. Despite its widespread application, accurately predicting membrane performance remains challenging due to the non-linear and inter-dependent influence of multiple operational parameters, including transmembrane pressure (TMP), crossflow velocity (CFV), temperature, pH, and solute concentration. This study evaluates the predictive performance of response surface methodology (RSM), artificial neural network (ANN), and random forest (RF) regression for permeate flux prediction. The models were trained and validated using 46 experimental data points compiled from a published study on water and wastewater filtration. Although derived from a different domain, the underlying flux dynamics and fouling mechanisms are conceptually analogous to those in food and bioprocessing systems, making this dataset a valid proxy for benchmarking model architecture and predictive capability. The RSM model identified TMP and CFV as the most statistically significant predictors and achieved an excellent goodness-of-fit (R² = 0.9993). The machine learning models also demonstrated strong predictive accuracy, with the ANN achieving an R²of 0.995 under cross-validation and the RF model reaching an R² of 0.933. The ANN effectively captured complex non-linear relationships in the data, whereas the RF model exhibited good generalization capability and interpretability. Collectively, these findings offer a conceptual framework for future membrane modelling applications in food and bioprocessing applications, with predictive accuracy expected to improve as larger, domain-specific datasets become available.