Advancing forward osmosis predictions: A deep learning‐based surrogate modeling approach

BACKGROUND: This study presents a deep learning-based surrogate model for the rapid and accurate prediction of forward osmosis (FO) performance under diverse operating conditions. To assess the applicability of data-driven approaches, several machine learning models - decision tree, random forest, support vector machine, and deep neural network (DNN) - were developed and compared systematically using datasets of varying sizes. RESULTS: Model performance depended strongly on dataset size. The DNN-based surrogate achieved superior accuracy when trained with more than 1563 data points, whereas the random forest model performed better with smaller datasets (<1563). Shapley additive explanation (SHAP) analysis showed that the DNN model captured the physical relationships between input parameters and FO performance effectively. The DNN-based surrogate predicted experimental water flux with a normalized root mean square error (NRMSE) of 0.082, which was comparable with that of the numerical simulation model of 0.083. CONCLUSION: The proposed DNN-based surrogate model maintains high predictive accuracy while substantially reducing computational time compared with the numerical simulation approach, enabling rapid and efficient process optimization in food manufacturing.