Wastewater from food industrial sectors is often characterized by high organic loads. In this study, submerged fluidized bed ceramic membrane reactor was applied as environmentally sustainable solution to treat food industry wastewater. Furthermore, it constructs a comparative deep learning model that uses Recurrent neural network (RNN), long short-term memory (LSTM), Bi-directional long short term memory (BiLSTM), and Gated recurrent units (GRU) to estimate membrane fouling in membrane filtration systems. Our results showed that by increasing permeate flux to 50 L. m−2. h−1, organic removal efficiency reached 80.3% with significant membrane fouling. About 43% fouling reduction efficiency was observed as bulk recirculation flow rate increased. Simultaneously, organic removal efficiency decreased as bulk recirculation rate increased. As membrane relaxation and filtration were conducted intermittently, the fouling rate was controlled more effectively than that by their continuous filtration. In terms of membrane fouling predictions, GRU had the best predictive output with maximum R2 and minimum error in time series processing when compared with other models. Monte Carlo Dropout was used to perform rigorous uncertainty quantification, and 95% confidence intervals were computed which include model epistemic uncertainty. To control membrane fouling, granular activated carbon was applied as fluidized media and membrane fouling was reduced considerably regardless of its different sources. However, despite the locally purchased carbon being cheaper but it showed 44% less organic removal. The X-ray diffraction and Fourier transform infrared spectroscopy analysis confirmed that the crystalline structure was important to improve adsorption of carbon. © 2026 Elsevier Inc.

https://www.scopus.com/pages/publications/105032848493?origin=resultslist