Author(s): Najat A. Amin; Adnan Qamar; Henry Tanudjaja

Linked Author(s): Henry J. Tanudjaja, Najat A.Amin

Keywords: Rtificial intelligence Desalination Machine learning Membrane biofouling Early detection

Abstract: Membrane processes are commonly applied in water treatment to obtain high-quality water from seawater and other sources. Reverse osmosis (RO) is a pressure-driven technology that effectively removes salts and ensures reliable water production. However, a key challenge in RO systems is membrane biofouling, which increases pressure drop and reduces water quality. Biofouling can lead to higher operational costs, increased energy consumption, more frequent chemical cleanings, and reduced membrane lifespan. Therefore, developing (bio) fouling monitoring strategies with early detection of biofilm formation on membrane systems is critically required. A comprehensive data generation and collection approach was adopted to establish a toolbox for automatic biofilm thickness detection and its correlation to the hydrodynamics parameters in the feed channel. Several methods will be used to generate a database, including differential pressure (∆P) sensors across the membrane, in which the increase in pressure drop occurred with the biofouling development (i. e., in terms of thickness and density), resulting in an increase in energy requirements for the filtration process. Also, feed and process parameters were extracted by using a multiparameter water quality probe composed of multiple water qualities. Another advanced approach to generate data for this study is non-destructive in-situ Optical Coherence Tomography (OCT) equipment that is installed to scan the membrane inside the MFS cell at a specified location and time to detect the growth of (bio) fouling during the operations. OCT is recently thrived as an effective microscopy technique to monitor the growth of biofilms in membrane fouling simulators (MFS) attached to RO pressure vessels. This imaging tool is capable of producing either two-dimensional (2D) or three-dimensional (3D) scans of the biofilm structure at a relevant spatial scale. Herein, we aim to establish direct links between the operation process, feed quality parameters, and the OCT scan data via employing deep neural networks (DNN). The networks, trained with this data, can be integrated into an intelligent governance framework for real-time performance evaluation and decision-making. Therefore, the database of biofilm images generated from an OCT scan of growing biofouling on the membrane surface is used to train a Convolutional Neural Network (CNN) to instantly predict biofilm thickness in RO systems. Moreover, a Non-Linear Regression-Deep Neural Network (NLR-DNN) is trained to accurately predict the non-linear relationship between operation and feed quality parameters and their association with growing biofilm in the filtration process. It shows that with proper training, the proposed machine learning model has the potential to assess how changes in input parameters over time affect the performance of RO membranes using a large-scale data-driven approach. The developed framework will be migrated to pilot facilities and ultimately to desalination plants for better decision-making and preventive controls to reduce water production costs in an artificial governance environment.

DOI:

Year: 2025