Artificial intelligence vs. classical approaches: a new look at the prediction of flux decline in wastewater treatment

Abstract

A B S T R A C TThis study compares the performance of three different approaches to modeling namely the classical pore-blocking models, artificial neural networks (ANN) and the novel genetic programming (GP) approach. Among the available models proposed by Hermia, standard poreblocking and cake filtration models were opted because of their better fitness with experimental measurements. A feedforward backpropagation network using Bayesian Regulation as well as Levenberg-Marquardt training methods was developed based on the experimental results. Network inputs include the controlling parameters of permeate flux namely: temperature, transmembrane pressure, crossflow velocity, pH, and filtration time. The architecture and internal parameters of the network have substantial effect on the prediction performance of the ANN. Hidden layers and neuron numbers were regulated using trial-and-error approach. The individual program proposed by GP, which has satisfied the required fitness value after 500 generations, had a depth of 10 among a population of 700 individuals. Relative error with respect to experimental results was used to compare the aforementioned models. It was found that ANN outperformed pore-blocking and GP models. The GP-based model had an acceptable coincidence with the experimental data and its ability to correlate the input and target variables by a mathematical relation showed the high potentiality of GP as a modeling tool.

A Review on Membrane Fouling Prediction Using Artificial Neural Networks (ANNs)

“…Additionally, conventional models typically require knowledge of a priori information about the system, such as the nature of the foulants or the fouling mechanism, which may not always be available [ 41 ]. Therefore, AI-based models have been increasingly used to predict membrane fouling because they can learn from data, find complex patterns, and make accurate predictions [ 42 , 43 ]. They can be trained on historical data of membrane filtration processes, including operating conditions, membrane properties, and feed characteristics, to predict the fouling rate [ 17 ].…”

Flux decline prediction in dead-end ultrafiltration combining fluorescence spectroscopy and mechanism-informed machine learning

“…The challenges to describe mechanistically such a combination of the many synergistic physico-chemical phenomena occurring during fouling have steered the research community to explore purely data-driven approaches (e.g., artificial intelligence/machine learning, AI/ML) 5,17,21,22 . Niu et al 5 recently reviewed the state-of-the-art AI/ML applied to predict water flux and other performance parameters in membrane processes during fouling.…”

Flux decline prediction in dead-end ultrafiltration combining fluorescence spectroscopy and mechanism-informed machine learning

“…Several algorithms have been explored, with artificial neural networks being the most employed 5 . Focusing on ultrafiltration, different researchers previously concluded that AI-based modeling approaches provide better predictive accuracy compared to mechanistic models, such as poreblocking 21,22 and Hermia's models 23,24 .…”