Permeate Flux Prediction in Dead-End Ultrafiltration of Different Types of Juices Using ANN and KNN Algorithms

Abstract

The main objectives of this study were to investigate the efficiency and model the dead-end ultrafiltration (UF) process applied to natural juices (chicory, beets and alfalfa) using two nonlinear methodologies. Artificial intelligence, specifically multiple linear perception (MPL) artificial neural networks (ANN), and k-nearest neighbours (KNN) were employed to forecast the permeate flux in two dead-end UF systems: Amicon vessel (AD) and rotating disc membranes (DRDM). The prediction was based on six key input variables, namely volumetric recovery rate (VRR), membrane porosity, transmembrane pressure (TMP), rotating velocity (Ω), density (ρ), and filtration device as a new input qualitative parameter. Permeate flux was used as outlet. The trial-and-error approach was utilized to find the architecture that produced the most suitable model, based on two key statistical metrics: the root mean square error (RMSE) and coefficient of determination (R²). The results obtained indicated that the ANN model demonstrates an ability to forecast the permeate flux in UF of juices with a value of R² of 0.919 and an RMSE of 7.71. Conversely, for the KNN model with K set at 3, the R² and RMSE values were found equal to 0.72043 and 7.6097, respectively. Consequently, ANN yields a superior value of R² compared to KNN, despite the latter exhibits marginally lower RMSE values. The advantage of this research is its effectiveness in predicting filtration results. It also saves time and effort by evaluating two types of dead-end filtration mechanisms under various experimental conditions.