Machine Learning Models for the Prediction of Transmembrane Flux in Ultrasonication Assisted Microfiltration of Pretreated Orange Juice

ABSTRACT The study investigates microfiltration with and without ultrasonication through experiments and machine learning models for evaluation and prior prediction of filtration performance. Clarification of orange juice using a 0.2 μm polyvinylidene fluoride membrane was carried out with various pre‐treatment methods applied prior to microfiltration. The pre‐treatments include centrifugation, enzyme treatment followed by centrifugation, and enzyme treatment followed by adsorption (bentonite, gelatin) along with centrifugation. Physicochemical analyses such as pH, total soluble solid (TSS), color, clarity, titratable acidity, viscosity, and density were performed for the fruit juice before and after the microfiltration process based on the best performance. Then, different machine learning models such as multilinear regression, polynomial regression, support vector regression, kernel ridge regression, artificial neural network, and random forest were used to predict the permeate volume of orange juice. Among the developed models, the random forest was found to yield an excellent prediction ( R 2 = 0.89 ± 0.054) of permeate volume. Further, the time interval at which permeate volume was collected, the transmembrane pressure, was identified using Shapley analysis as the most influential feature contributing to model predictions predicting the filtration efficiency. Among the physicochemical properties, color values decreased from 2.62 (control) to 0.12, and clarity increased significantly (from 3.34% to 93.63%) in all the pre‐treatments compared to control, while pH, TSS, and density almost remained the same. The findings illustrate that ultrasonication‐assisted microfiltration with the combination of different pre‐treatments using enzymatic treatment and adsorption employing fining agents has a significant impact on improving the performance of the microfiltration. Further, the practical applicability of random forest, artificial neural networks, support vector regression as well as polynomial regression has been demonstrated as an effective approach for the accurate prediction of the performance of microfiltration of orange juice.