Comparing a new semi-empirical mathematical model and a neural network for the description and forecasting of reversible fouling in membrane bioreactors

This study compares the predictive capability of membrane fouling between a new semi-empirical mathematical model and a neural network. Calibration and validation involved 21 replicated flux-step experiments with 8 filtration fluxes ranging from 9.8 to 18.9 L/m²·h and 40-hour random-flux experiments conducted in an anaerobic membrane bioreactor. The inherent variability of reversible fouling linked to cake build-up was quantified.

In addition to cake build-up and compression, the mathematical model incorporates the restoration of residual fouling not removed by backwashing, including the rearrangement of non-detached particles and colloids, initial pore-blocking, and concentration polarization. The model predicts reversible fouling across a wide range of filtration fluxes, even beyond those used for calibration. Random-flux experiments revealed fouling trends undetectable via direct transmembrane pressure inspection. However, the mathematical model fails when the effective membrane area decreases during filtration.

The neural network predicts fouling patterns independently of underlying mechanisms, offering adaptability across a broader range of operating conditions. Nonetheless, it struggles to predict reversible fouling for flux ranges outside its training dataset, particularly at fluxes significantly higher or lower than those used in training.

The present study offers insight into reversible fouling description in order to enhance the performance of membrane bioreactors.