Utilising Artificial Intelligence to Predict Membrane Behaviour in Water Purification and Desalination

Abstract

Water scarcity is a critical global issue, necessitating efficient water purification and desalination methods. Membrane separation methods are environmentally friendly and consume less energy, making them more economical compared to other desalination and purification methods. This survey explores the application of artificial intelligence (AI) to predict membrane behaviour in water purification and desalination processes. Various AI platforms, including machine learning (ML) and artificial neural networks (ANNs), were utilised to model water flux, predict fouling behaviour, simulate micropollutant dynamics and optimise operational parameters. Specifically, models such as convolutional neural networks (CNNs), recurrent neural networks (RNNs) and support vector machines (SVMs) have demonstrated superior predictive capabilities in these applications. This review studies recent advancements, emphasising the superior predictive capabilities of AI models compared to traditional methods. Key findings include the development of AI models for various membrane separation techniques and the integration of AI concepts such as ML and ANNs to simulate membrane fouling, water flux and micropollutant behaviour, aiming to enhance wastewater treatment and optimise treatment and desalination processes. In conclusion, this review summarised the applications of AI in predicting the behaviour of membranes as well as their strengths, weaknesses and future directions of AI in membranes for water purification and desalination processes.

1. Introduction

Clean water remains a critical global issue. Based on recent data, an estimated 2.2 billion individuals lack access to properly managed drinking water services. Projections indicate that by 2025, approximately half of the global population will be residing in water-stressed regions. These statistics underscore the critical need to develop and implement effective water management strategies. There are several methods available to address clean water scarcity. When it comes to water purification, techniques such as reverse osmosis, ultraviolet (UV) treatment and advanced oxidation processes (AOPs) are widely utilised, each with its own distinct set of advantages and limitations [1]. In the field of desalination, seawater can be converted into potable water through processes like multi-stage flash distillation and electrodialysis [2]. However, desalination faces challenges such as high energy consumption and environmental impact. Another remedy for clean water issues is water recycling and reuse, which involves treating wastewater for reuse in agricultural, industrial and even potable applications, thereby reducing the demand for freshwater sources [2,3]. However, energy consumption, environmental impacts and economic viability are the limitations of these methods [4]. When it comes to new technologies, different methods like membrane separation, multi-effect desalination (MED) and multi-effect desalination thermal vapour comparison (MED-TVC) are employed for water purification and desalination, ensuring the provision of clean water for different purposes such as drinking, agriculture and industry. While the MED and MED-TVC methods demand more energy and resources to operate, membrane separation methods are more environmentally friendly [5]. Membrane separation methods include reverse osmosis (RO), forward osmosis (FO), nanofiltration (NF), ultrafiltration (UF) and microfiltration (MF). These techniques vary in their operational principles and applications, making them suitable for different purification needs [6]. Water flux refers to the speed at which water permeates through a membrane [7]. Multiple factors, including membrane properties, operating conditions and feed water quality influence this process. Membrane fouling is the next item to describe which occurs when unwanted materials accumulate on the surface or within the membrane’s pores, leading to decreased performance [8]. Fouling can be caused by various substances, such as suspended solids, scaling, organic matter and biological fouling. Various traditional methods are available for predicting the behaviour of membranes in water purification and desalination. These methods include empirical models, mechanistic models, transport phenomena, mass transfer models, hydraulic models, fouling models and analytical techniques such as scanning electron microscopy (SEM) and Fourier-transform infrared spectroscopy (FTIR) [9]. On the other hand, recent advancements in AI have created new opportunities for optimising membrane operations, particularly through the integration of data obtained from Computational Fluid Dynamics (CFD) simulations. CFD provides important time-dependent variables that are essential for machine learning models [10]. The water treatment processes can be influenced by various factors that can be simulated using CFD. These factors include temperature concentration polarisation (CP), External Concentration Polarisation (ECP), Internal Concentration Polarisation (ICP), specific energy consumption (SEC), Temperature Polarisation Coefficient (TPC) and Feed Concentration Polarisation (CP) [11]. The flow rate and configuration of the FO system have a significant impact on External Concentration Polarisation (ECP), which can reduce the effective osmotic pressure difference. Internal Concentration Polarisation (ICP) is affected by membrane structure and supports layer properties, leading to a reduction in water flux [12]. The specific energy consumption (SEC) and power density for the PRO process range from 0.2 to 0.8 kWh/m³ and up to 10 W/m², respectively [13]. In direct contact membrane distillation (DCMD), the Temperature Polarisation Coefficient (TPC) values range from 0.9 to 0.97 for channels with spacers and the Feed Concentration Polarisation (CP) significantly impacts the permeate flux [14].

In the field of water purification and desalination, AI can play a critical role in designing, predicting behaviour and optimising process operations [15]. Machine learning (ML), a subset of AI, encompasses the process of training algorithms to acquire knowledge from data and subsequently make informed predictions or decisions [16]. By identifying patterns and utilising minimal human intervention, ML models find application in recommendation systems, image and speech recognition and predictive analytics [17]. artificial neural networks (ANNs), a category of ML model, are inspired by the structure and function of the human brain [16,18]. Comprising interconnected layers of nodes (neurons), including an input layer, one or more hidden layers and an output layer, ANNs adjust the weights of their connections as they learn from data. Utilising AI platforms, such as ML and artificial neural networks (ANNs), can effectively model water flux and fouling behaviour in water purification and desalination operations [15,19].

Table 1. Comparison of conventional and AI models in water purification.

Model Type	Method	Performance Metrics	Results
Empirical [15]	RO	Water flux, fouling rate	Moderate accuracy, high energy consumption
Mechanistic [20]	Multi-Stage Flash Distillation	Energy efficiency, salt rejection	High energy consumption, reliable performance
Transport Phenomena [20]	Electrodialysis	Ion removal efficiency	Effective for specific ions, moderate energy use
AI/ML (ANN) [21]	FO	Water flux, fouling prediction	High accuracy, low energy consumption (R2ANN = 0.98036, R2RSM = 0.9408)
AI/ML [22]	FO	Permeate quality, fouling prediction	High performance of 0.997, mean square error of 0.04

shows a comparison between conventional and AI models in desalination and water treatment.

As Table 1 shows, AI models, including machine learning (ML) and artificial neural networks (ANNs), have demonstrated superior predictive capabilities compared to traditional methods.

Previous research has demonstrated the potential of AI in optimising membrane processes, as seen in the work by Jasir et al. [23] which developed a multilayered neural network for FO process performance evaluation with nine input variables. The performance of the ANN model surpasses that of a transport-based model, underscoring its capacity to comprehend intricate relationships between inputs and outputs. Attarde et al. [24] formulated the Spiegler-Kedem (SK) model to delineate local mass transport within the active layer of the membrane. The proposed model and the inferred parameters were validated using comparing module performance across diverse operational scenarios. Alam et al. [25] introduced a recurrent neural network (RNN) system tailored for a desalination plant. The RNN model underwent optimisation leveraging historical data to ensure consistent and reliable desalination process performance. The developed model was integrated into MATLAB and its simulation results yielded enhanced hybrid system performance in terms of resource utilisation, battery storage and minimised energy loss. Shah et al. [26] conducted a study on the dynamics of the Reynolds nanofluid model for Casson fluid flow (RNFM-CFF) over a slandering sheet utilising computational intelligence with Levenberg–Marquardt networks (CILMNs). The model effectively predicted crucial data and the accuracy of the estimated solution was verified through rigorous training and testing processes. Fetanat et al. [27] developed a feed-forward artificial neural network (ANN) model with one hidden layer and Bayesian regularisation to overcome commercial challenges of integration of nanoparticles into polymeric ultrafiltration membranes. The ANN accurately predicted the performance of ultrafiltration nanocomposite membranes, including solute rejection, flux recovery and pure water flux. Rezakazemi et al. [28] conducted a study aimed at developing a 2D mathematical model to simulate the seawater purification process using a DCMD system. The model coupled conservation equations for water molecules in three domains of the contactor module. The simulation results showed that the DCMD system could effectively purify seawater. The model accurately predicted the temperature and concentration profiles within the system. Alrefaai et al. [29] evaluated the desalination performance of different types of ceramic-based hollow fibre membranes in a DCMD system. The CFD model was developed and validated with experimental results over a 500 min period. The hydrophobic membrane (C8-HFM) showed a 30% and 9% higher initial mass flux compared to the flower-like (C8-FL/TiO₂) and rod-like (C8-RL/TiO₂) omniphobic membranes, respectively. Over the 500 min period, the flower-like omniphobic membrane (C8-FL/TiO₂) had the lowest flux reduction (11%), followed by the rod-like omniphobic membrane (C8-RL/TiO₂) with a 15% reduction and the hydrophobic membrane (C8-HFM) with a 23% reduction. Abrofarakh et al. [30] examined the impact of various factors such as inlet temperatures, velocities, channel heights, salt concentration and membrane characteristics on the efficiency of the DCMD process. The CFD model was rigorously validated against established studies, ensuring its reliability. In addition, over 1000 data points were used to train and validate the ANN, achieving high accuracy with near-zero mean squared errors and R values close to one. The results showed that a higher membrane porosity increased water vapour flux and thicker membranes reduced water vapour flux. Furthermore, salt concentration, channel dimensions, inlet temperatures and velocities significantly influence the distillation process. In this study, a mathematical model was proposed for water vapour flux as a function of key input factors. The study highlighted that salt mole fraction and hot water inlet temperature have the most significant effects on water vapour flux.

In light of the increasing global water scarcity and the pressing need to improve the efficiency of water purification and desalination membranes, this review presents a comprehensive analysis of the potential of artificial intelligence (AI) in predicting membrane behaviour. This study’s uniqueness lies in its thorough assessment of AI tools’ adaptability and predictive accuracy, particularly in the areas of forecasting and optimising membrane fouling, flux modelling and micropollutant prediction. By focusing on these aspects, the review aims to enhance the separation process and maximise the production of clean water. Furthermore, it delves into the intricate relationship between AI and membranes and identifies future research directions in this field.

2. Membranes in Water Purification and Desalination

Currently, a variety of membrane desalination techniques are employed in water treatment processes. This approach entails subjecting water to pressure through a semipermeable membrane to affect the separation of salts and contaminants [31]. Reverse Osmosis (RO) is the most widely used desalination process due to its high efficiency, lower energy consumption, reliability and minimal negative environmental impacts. Within the domain of RO desalination, four distinct types of membrane modules are deployed: plate and frame, tubular, spiral wound and hollow fibre [32].

The plate and frame type encompasses flat sheet membranes interposed between plates. In this configuration, feed water traverses the channels between the plates, while purified water is gathered on the opposite side [33]. Notably, this design is uncomplicated, facilitating effortless maintenance and cleaning, rendering it well-suited for small-scale applications. Tubular modules, on the other hand, encase the membrane within a tube. Here, the feed water is propelled through the tube, with the purified water permeating through the membrane to the exterior [34]. Recognised for their robustness, tubular modules excel in managing high levels of suspended solids, thereby proving ideal for treating wastewater and other challenging feed streams. Spiral wound modules comprise flat sheet membranes coiled around a central collection tube in a spiral configuration [35]. Delivering a substantial surface area in a condensed form, this design is markedly efficient for large-scale desalination and water purification processes. Nonetheless, a susceptibility to fouling necessitates meticulous pre-treatment of the feed water. Lastly, hollow fibre modules encompass numerous small, hollow fibres through which the feed water flows. The purified water permeates through the walls of the fibres. This design yields an exceptionally high surface area-to-volume ratio, lending to its high efficiency [36]. Hollow fibre modules are commonly deployed in applications mandating high purity, including medical and pharmaceutical industries. Industrial applications necessitate RO membrane modules that achieve a high-packing density (the ability to fit a large membrane surface area into a relatively small volume) to accommodate a significant membrane area within a compact volume, facilitating the removal of contaminants through pore sizes ranging from 0.0001 μm to slightly larger than 0.001 μm [37]. Various membrane filtration methods, such as MF, UF and NF, serve different purposes in removing particles from water [38]. MF utilises a low-pressure membrane process designed to selectively remove larger particles using pores sized between 0.1 and 10 μm, which results in minimal fouling due to the larger pore size [39]. This characteristic makes MF particularly suitable for pre-treatment stages in water purification systems, where it effectively reduces the load on subsequent filtration processes, thereby enhancing overall system efficiency and longevity. Additionally, the low-pressure operation of MF systems contributes to lower energy consumption and operational costs, making it an economically viable option for various industrial and municipal applications. UF, operating under medium pressures, selectively removes smaller particles using membranes with pores ranging from 0.001 to 0.1 μm [40]. Compared to microfiltration membranes, its main advantage is the high rejection of larger contaminants while allowing water and smaller solutes to pass through. [41]. This selective permeability is crucial for applications requiring the removal of bacteria, viruses and colloidal particles, thereby ensuring a high level of water purity. The medium-pressure operation of UF systems strikes a balance between effective contaminant removal and energy efficiency, making it a cost-effective solution for both industrial and municipal water treatment processes. Furthermore, UF membranes are often employed as a pre-treatment step in RO systems, enhancing the overall efficiency and lifespan of the RO membranes by reducing fouling and scaling. The robustness and versatility of UF technology make it an integral component in the production of potable water, wastewater treatment and various industrial processes requiring high-quality water. NF, with pores ranging from 0.001 to 0.01 μm, acts as a link between UF and RO [42]. By acting as a link, NF bridges the gap between UF and RO, providing a gradual transition in the filtration process. This stepwise approach enhances the overall efficiency and effectiveness of water purification systems. Selective removal, energy efficiency, resistance to fouling and versatility make NF a viable option for treating brackish water, removing hardness ions such as calcium and magnesium and treating coloured water [43]. NF offers partial salt rejection with higher permeability than RO, but it can be expensive due to energy requirements and maintenance costs [44]. Additionally, it is important to consider factors such as the disposal of concentrates and energy consumption [45].

To consider environmental impacts, various membrane desalination systems exist, including the self-heated vacuum membrane distillation (MD) system. This system integrates interfacial heating and conventional vacuum MD (VMD) to facilitate energy-efficient seawater desalination, sustainability, cost reduction and improved performance [46]. It employs a graphene–PVDF flat sheet membrane Joule heater to directly heat the feed stream at the feed/membrane interface [47]. The self-heated vacuum membrane distillation system offers numerous advantages, such as eliminating the need to preheat the feed solution and reducing temperature polarisation (TP), enhancing permeate flux and reducing energy requirements [48,49]. Another solar-driven membrane desalination system is the nanophotonic-enhanced solar membrane distillation (NESMD), which is a small-scale system that desalinates seawater and high-salinity feedwaters using sunlight as the primary energy source [50]. It represents a promising approach for decentralised water desalination and can turn almost any source of water into clean water, making it a sustainable solution [51]. The second solar-driven system discussed is membrane distillation coupled with solar energy [52]. MD is a non-isothermal process with simultaneous mass and heat transfer that can be implemented on small- and medium-scales [53,54]. Coupling MD with solar energy offers an affordable alternative for desalinating seawater or highly saline sources, with good energy efficiency and low-cost [55].

Table 2. Different environmentally friendly membrane desalination systems.

Methods	Features
Self-heated vacuum membrane distillation (MD)	- Eliminating the preheating of feed solution and temperature polarisation [48,56]; - Enhancing permeate flux and reducing energy requirements [48]; - Impressive performance [57]; - The lowest specific heating energy consumption [48,56]; - The highest gain output ratio [55]; - Suitable for high-purity water production [56,57].
Nanophotonic-enhanced solar membrane distillation (NESMD)	- A standalone small-scale system [54]; - A promising approach for decentralised water desalination [50]; - Has the ability to turn almost any source of water into clean water [51].
Membrane distillation coupled with solar energy	- A non-isothermal process with simultaneous mass and heat transfer [58]; - Fully implementable on small- and medium-scales [59]; - An energy-efficient method [55]; - A low-cost method [60].
RO [61]	- High rejection rates for a wide range of contaminants; - Energy efficient with advanced energy recovery systems; - Suitable for large-scale desalination plants; - Requires pre-treatment to prevent membrane fouling.
FO [62]	- Lower energy consumption compared to RO; - Utilises natural osmotic pressure differences; - Effective for treating high-salinity and wastewater streams; - Challenges with draw solution recovery and membrane fouling.
Pressure Retarded Osmosis (PRO) [63]	- Generates energy from salinity gradients; - Potential for integration with existing desalination and wastewater treatment systems; - High energy efficiency and low environmental impact; - Technical challenges in membrane development and optimisation.

shows a summary of eco-friendly membrane desalination systems.

3. Predicting the Membrane Behaviour by AI Models

There are different applications of AI in membranes to enhance their performance in different processes, especially water treatment and desalination plants. AI can be utilised to predict the fouling, flux modelling and micropollutant behaviour in membranes by large datasets and contributing different parameters as inputs of the training model [64,65].

An algorithm for predicting flux or fouling in membrane desalination using AI comprises several key components [21,66,67,68,69]. The first block involves data collection and preprocessing, where a dataset containing membrane properties, operating conditions (such as pressure and temperature) and corresponding water flux values is gathered and normalised to ensure consistent input feature scales.

The second block focuses on selecting the most suitable architecture type based on the problem at hand. For regression tasks, such as predicting continuous values like flux, options include feedforward neural networks (FNNs) or recurrent neural networks (RNNs). When dealing with image-based flux prediction, utilising membrane surface images, convolutional neural networks (CNNs) are applied.

The third block encompasses the design phase, involving the definition of the architecture. This includes setting up the input layer with membrane property and operating condition features, implementing hidden layers with nonlinear activation functions and an output layer with a single neuron for predicting flux. Additionally, it entails the fine-tuning of hyperparameters such as learning rate, batch size and the number of hidden layers.

The subsequent block is dedicated to the training phase, which involves splitting the dataset into training, validation and test sets. The model is trained using stochastic gradient descent (SGD) or an optimiser, with monitoring of validation set loss to prevent overfitting. Early stopping techniques are used to avoid excessive training.

The feature extraction block follows the training phase. For image data, CNN layers are used to extract features from membrane surface images, while techniques such as PCA or autoencoders are employed for other types of features to reduce dimensionality.

In the final block, the prediction phase, the trained model is utilised to predict the flux or fouling for new membrane properties and operating conditions. The model’s performance is rigorously evaluated using the test set. Figure 1 shows an algorithm for modelling and predicting flux and fouling in membranes.

3.1. Fouling Prediction in Membranes by AI-Based Models

There are different AI-based models for predicting the fouling in membranes. Due to the nonlinearity of filtration processes, standard mathematical equations are inadequate and artificial intelligent models (neural networks, knowledge-based, fuzzy systems, adaptive neuro-fuzzy) should be investigated [70]. Genetic algorithms, particle swarm optimisation and gravitational search algorithms enhance model prediction.

One of the ML integrative concepts is PCA which helps identify the most important factors contributing to fouling by transforming the original data into a new set of independent variables to capture most of the data’s variability, focusing on the most relevant factors [71]. PCA reveals underlying patterns in the data, aiding in predicting fouling behaviour by identifying dominant trends [72]. By examining the loadings of each variable on the principal components, it can be determined which features contribute most to fouling. High-dimensional data can lead to overfitting and computational challenges [73]. PCA addresses this by retaining only the most informative principal components, reducing the dimensionality. After dimensionality reduction, the transformed data (principal components) can be used for modelling. [74]. Regression models or machine learning algorithms can predict fouling rates based on the reduced feature set. The accuracy of the model depends on the quality of the original data and the relevance of the retained principal components. Deviations from expected patterns, such as sudden changes in principal component scores, can indicate fouling onset. Early detection allows for timely maintenance and cleaning, improving system efficiency. In terms of predicting fouling behaviour, PCA has developed by inputs involving transmembrane pressure, flow rates, water quality parameters and operational data and outputs involving fouling prediction, anomaly detection and performance metrics [75].

Machine learning platforms have been employed for optimising the processes [76]. Optimisation methods such as genetic algorithm (GA) and particle swarm optimisation (PSO) effectively optimise parameters related to membrane fouling, including transmembrane pressure, crossflow velocity, feed temperature and feed pH [77]. Hybrid intelligent models that combine these optimisation methods outperform single models [78].

An artificial neural network model, specifically a multilayer perceptron artificial neural network (MLP ANN), has been developed to predict fouling in membrane systems [79]. This model can be trained using a set of databases from a large-scale wastewater treatment plant, incorporating hydraulic and water quality parameters. The model is capable of calculating fouling factors, identifying optimal predictors and optimising feed content for fouling prediction. Notably, early-stage fouling is associated with total chlorine content, electrical conductance and total dissolved solids, while late-stage fouling is best predicted by turbidity, nitrate, nitrite and organic matter levels in the feed. This model can be utilised to enhance the efficiency and performance of RO systems in wastewater treatment, ultimately resulting in more effective water purification processes. The ANN model has demonstrated the ability to predict membrane fouling with high accuracy across various nanofiltration stages [80]. Key inputs for the ANN model include flow rates and feed water quality parameters. Notably, under experimental conditions, the model does not incorporate colloidal fouling and biofouling, yet it can still make precise predictions without utilising feed water turbidity and bacteria concentrations as inputs. Impressively, the model requires minimal data for training, achieving accurate predictions using just 10% of almost 650 experimental data while maintaining less than 5% absolute relative error. These straightforward ANNs excel at capturing changes in feed water quality, flux and recovery, thus addressing the challenges associated with mechanical models for long-term fouling prediction in municipal drinking water nanofiltration. ANN models can be utilised to develop and validate the impact of membrane fouling caused by polydisperse feed suspensions. These models accurately simulate the changes in specific fluxes over time for various types of feed suspensions and have been validated across a wide range of hydrodynamic parameters [81]. Figure 2 shows a common structure of ANNs in membrane modellings.

By increasing the number of hidden layers, the accuracy and performance of the neural network would increase. However, there is an optimum point as overly increasing the number of hidden layers causes issues such as overfitting. In terms of DL, a long short-term memory (LSTM) model has been utilised to examine variations in filtration performance and fouling growth [82]. LSTM models are well-suited for tasks involving time series data or sequences due to their ability to capture long-term dependencies and patterns over time. The handling of sequential data is the first reason why LSTM is suitable for predicting fouling in membranes. These models are specifically designed to work with sequences of data, making them ideal for time series analysis where historical information is essential for predicting future states. Another advantage is the memory capability of LSTMs. They feature a memory cell that can retain information for extended periods, aiding in understanding the progression of fouling over time. Additionally, overcoming the vanishing gradient problem is another advantage of LSTM models. Traditional RNNs struggle with long-term dependencies due to this issue, while LSTMs can learn from long sequences thanks to their unique architecture. The input to an LSTM model is typically a three-dimensional array (samples, time steps, features). The filtration performance task includes samples (different instances of the filtration process), time steps (sequential time points at which measurements are taken) and features (variables such as pressure, flow rate, temperature and fouling indicators). The output can vary depending on the specific task. In the single-step prediction, it predicts the next value in the sequence which is the level of fouling at the next time step. In the multi-step prediction, it predicts fouling levels over the next several time steps.

To collect data, laboratory-scale membrane fouling experiments can be established using natural organic matter (NOM) in filtration systems. Leveraging the experimental data, the LSTM model can predict permeate flux and fouling layer thickness. This study illustrates that deep learning can replicate the influence of NOMs on NF systems and other membrane processes. This model can predict the fouling thickness behaviour with high accuracy (R² = 0.9987). Sung et al. [83] utilised optical coherence tomography (OCT) to continuously monitor the fouling layer in FO membranes. Using a standard membrane cleaning method, their observations spanned four distinct stages over 21 days. They found that chemical cleaning, which removed two to three times more organic matter (OM) than physical cleaning, was notably effective. Real-time OCT observations revealed an initial thin, dense and flat fouling layer, followed by a subsequent stage characterised by a thick and rough surface. Furthermore, a deep learning convolutional neural network model accurately predicted fouling layer characteristics based on real-time fouling images, achieving high correlations for the thickness, porosity, roughness and density of the fouling layer. This innovative approach holds promise for enhancing the comprehension and management of membrane fouling in FO and other membrane processes. Sanghun et al. [84] developed a deep neural network (DNN) for modelling membrane fouling during NF and RO. They used in situ fouling image data from optical coherence tomography (OCT). The DNN model was trained using 13,708 high-resolution fouling layer images. It simulated both organic fouling growth and flux decline and achieved better predictive performance than existing mathematical models. The R² value of 0.99 for fouling growth simulation and 0.99 for flux. Bagheri et al. [85] focused on AI and ML techniques for controlling membrane fouling in filtration processes, particularly in water and wastewater treatment systems. Notably, four successful modelling techniques were identified: ANNs, fuzzy logic, genetic programming and model trees. Among these, well-known ANNs like multilayer perceptron and radial basis function exhibit impressive predictive capabilities, achieving an R² value of 0.99 and minimal error in membrane fouling prediction. Other intelligent techniques, including clustering analysis, image recognition and feature selection, also play a positive role in membrane fouling control. By building a recurrent neural network (RNN) model for time-series analysis, we can forecast the long-term performance of an osmotic membrane bioreactor, including conductivity, water flux and membrane fouling [86].

Table 3. AI methods for predicting fouling behaviour in membranes.

Method	Inputs	Outputs	Description
- GA + PSO [87]	- Feed temperature, pressure, pH	- Optimisation/fouling prediction	Integrates genetic algorithm (GA) and particle swarm optimisation (PSO) to enhance the predictive accuracy of fouling behaviour by optimising operational parameters.
- MLP + ANN [88]	- Hydraulic and water quality parameters	- Fouling prediction	Utilises multilayer perceptron (MLP) and artificial neural networks (ANN) to model and predict membrane fouling based on complex hydraulic and water quality data.
- ANNs [89]	- Polydisperse feed suspensions	- Fouling behaviour	Employs artificial neural networks (ANNs) to simulate the fouling dynamics in systems with polydisperse feed suspensions, capturing the heterogeneity of particle sizes.
- LSTM [82]	- Pressure, temperature, pH	- Fouling growth	Leverages long short-term memory (LSTM) networks to forecast the temporal evolution of fouling layers, accounting for sequential dependencies in the data.
- CNNs [90]	- Transmembrane pressure, crossflow velocity, temperature, pH	- Fouling prediction	Applies convolutional neural networks (CNNs) to analyse spatial and temporal variations in membrane parameters, providing robust predictions of fouling events.
- RNN [91]	- Operating parameters	- Conductivity, fouling and flux prediction	Utilises recurrent neural networks (RNNs) to predict conductivity, fouling and flux variations by modelling the sequential nature of operating conditions.

highlights a summary of different described AI methods for predicting fouling behaviour.

3.2. AI-Based Flux Modelling in Membranes

Several investigations have been carried out in the field of flux modelling in membranes in water treatment and desalination processes.

Support Vector Machine (SVM) is widely used for both classification and regression tasks [92]. However, its primary application lies in solving classification problems [93]. The main objective of the SVM algorithm is to find the optimal decision boundary (hyperplane) that effectively separates data points in an n-dimensional space into distinct classes [94]. This boundary allows us to correctly categorise new data points in the future. In membranes, SVM has been employed to predict permeability by inputs involving operational parameters and output of permeability of the membrane [95]. The results have shown the high accuracy of implementing this model with R² values above 0.99.

By developing a computational tool capable of simulating osmosis through an asymmetric membrane in pressure retarded osmosis mode across various scenarios, an agent-based model can be created using the NetLogo [96]. Known for its graphical visualisation capabilities and suitability for modelling complex systems, NetLogo can be used to validate the simulation results against empirical data from the literature, showing good agreement. The agents in this model represent various components of the forward osmosis process, including water molecules, solute particles and membrane elements. These agents interact based on predefined rules that simulate the physical and chemical processes involved in forward osmosis. For instance, water molecules move across the membrane from the feed solution (FS) to the draw solution (DS) driven by osmotic pressure differences. The model integrates parameters such as temperature, membrane characteristics and flow velocities to faithfully simulate the FO process. The simulation outputs include metrics like water flux, solute concentration and membrane performance, which are instrumental in evaluating the efficiency of the FO process under different conditions.

In the process of FO groundwater desalination, optimisation can be achieved through a combined Taguchi-neural network approach [97]. The input parameters, including feed solution velocity, draw solution velocity, feed solution temperature and draw solution temperature, are vital in this process. The measured outcome is the maximum reverse solute flux selectivity. The Taguchi method involves evaluating the primary optimal condition using an L16 (44) orthogonal array and examining the impact of different parameters like feed solution velocity, draw solution velocity, feed solution temperature and draw solution temperature. After determining the primary optimal conditions using the Taguchi method, the study employed neural networks to refine these conditions further. The experimental results from the Taguchi method were used to train the neural networks, which then predicted the real optimum parameter levels for the FO process. Analysis of Variance (ANOVA), a statistical tool was used to identify the significant parameters affecting the FO process. ANOVA helped in recognising the most influential factors, such as the draw solution temperature, which was found to be the most critical parameter for maximising reverse solute flux selectivity (RSFS). The experimental setup for FO involved varying the feed and draw solution velocities and temperatures. The study focused on two orientations: active layer facing feed solution (AL-FS) and active layer facing draw solution (AL-DS), to determine the optimal conditions for each orientation. Effluent quality and energy consumption reduction in wastewater treatment processes have been enhanced through the use of the adaptive Kernel function model by employing several AI models and methodologies [98]. Utilising this model requires to be integrated with other models. The first model is the Fuzzy Neural Network (FNN). The core AI model used in this modelling is the Fuzzy Neural Network, which combines fuzzy logic with neural networks to handle the uncertainties and nonlinearities in the wastewater treatment process. Adaptive Kernel Function Models are the second platform. These models are integrated within the FNN to enhance the learning and adaptation capabilities of the controller. They help in fine-tuning the control parameters dynamically based on the real-time data from the wastewater treatment process. In this platform, the third model is multi-objective optimisation. The study uses a multi-objective optimisation approach to balance different objectives such as improving operational efficiency, satisfying effluent quality and reducing energy consumption. The improved multi-objective particle swarm optimisation (MOPSO) integrates self-adaptive flight parameters and a multi-objective gradient method to minimise specific objectives. This modelling helps identify optimal set points for dissolved oxygen and nitrate in the treatment process and these set points can be managed using the adaptive fuzzy neural network controller (FNNC). The benchmark simulation model No.1 (BSM1) has been used for result evaluation.

The important results of the model are, firstly, to improve the operational efficiency of the wastewater treatment process by optimising the control parameters in real-time. Secondly, the study achieved better effluent quality, meeting the required standards more consistently compared to traditional control methods. Furthermore, the optimised control strategy led to a notable reduction in energy consumption, making the process more sustainable and cost-effective. These results demonstrate the effectiveness of combining fuzzy neural networks with adaptive kernel function models for optimising complex processes like wastewater treatment.

ANN has also been developed in reverse osmosis (SWRO) desalination plant to optimise the operation of a SWRO desalination plant by predicting its performance using an ANN model [99]. The model used five input parameters: feed temperature, feed total dissolved solids (TDS), transmembrane pressure (TMP), feed flow rate and time. The model predicted two output parameters: Permeate TDS, permeate flow rate in terms of data collection, one year of operational data (200 samples) from the Fujairah SWRO plant was used. The data were divided into training, validation and test sets to develop and validate the ANN model. The results showed that the trained ANN model showed a high degree of accuracy, with an R² value of 0.96 for TDS and 0.75 for flow rate in the test dataset. The study found that variations in feed water temperature and TMP significantly affected both permeate TDS and flow rate. In terms of optimisation, simulations suggested that a linear increase in feed temperature (from 27.5 °C to 29.5 °C) in an SWRO hybrid system with multi-stage flash (MSF) distillation could reduce permeate TDS. These findings demonstrate the effectiveness of using ANN models for optimising the operation of SWRO desalination plants, leading to improved performance and efficiency. The output of this model is the permeate TDS and flow rate. The operation data from the SWRO plant can be divided into training, validation and test datasets. According to this modelling, feed water temperature and TMP significantly affect permeate TDS and flow rate. In this model, linear temperature control can reduce permeate TDS in SWRO hybrid systems like the Fujairah plant. Figure 3 shows an ANN structure for predicting the water flux in membranes.

Furthermore, in forecasting RO plant performance, a modelling approach utilising neural networks with back-propagation and support vector regression (SVR) algorithms has shown promise [101]. SVR has been employed as a data-driven method to forecast the performance of reverse osmosis (RO) plants. SVR was selected for its capability to handle nonlinear relationships and provide robust predictions. The model utilises past data on normalised permeate flux, salt passage and operational parameters to capture short-term variations in plant performance. During model training, various model architectures, memory time intervals and forecasting times were tested using historical data from full-scale RO plants. The results demonstrate that the SVR models successfully predicted the decline in permeate flux and salt passage trends over time. This is crucial for the early identification of membrane fouling and operational diagnostics. In addition, both SVR and artificial neural network (ANN) approaches were evaluated. While ANN models performed well, SVR provided competitive results, particularly in capturing short-term variability in plant performance. The SVR models also have the potential to act as soft sensors for early detection of fouling, enabling more effective process control strategies [101]. This modelling illustrates the effectiveness of SVR in predicting complex desalination processes and lays the foundation for developing advanced diagnostic tools for RO plant operations.

In this kind of modelling, to capture time-variability in plant performance, a short-term memory time interval is applied. Past information on normalised permeate flux and salt passage serves as unique input variables in a sequential model. This model accounts for time-variation within forecasting intervals and utilises marching forecasting models to predict target values at fixed future times, resulting in good predictive accuracy for short-term memory time intervals (8–24 h) in permeate flux and salt passage forecasting. This approach can be extended for longer time scales and integrated into control strategies and process diagnostics through data-driven forecasting models. Membrane flux prediction in forward osmosis (FO) systems can be achieved using an AI-based model by detailed data gathering.

Among two tree-based, two deep learning and two miscellaneous, XGBoost achieves an impressive coefficient of determination (R²) of 0.992 and tree-based models outperformed neural networks [102]. In this modelling, sensitivity analysis highlights strong correlations between feed solution (FS) and draw solution (DS) concentrations with membrane flux. Utilising machine learning techniques such as response surface methodology (RSM), artificial neural networks (ANN) and adaptive neuro-fuzzy inference system (ANFIS) have helped to analyse the impact of initial draw concentration, feed concentration, time, pH and temperature on water flux and reverse salt flux [103]. Upon achieving optimal conditions, the performance of FO can be experimentally validated using ANN and RSM models. The prediction of water flux in Osmotic Membrane Bioreactors (OMBR) can be established for wastewater treatment [104]. This involves integrating and developing adaptive network-based fuzzy inference systems (ANFIS) and ANN models with inputs such as mixed liquor suspended solids (MLSS), electrical conductivity and dissolved oxygen. It is noted that ANFIS models exhibit a lower root mean square error compared to ANN models. DL can be utilised to model the flux of FO membranes. In the FO process, a concentrated draw solution (DS) is employed to pull water through a semi-permeable membrane from a less concentrated feed solution (FS). Accurately predicting membrane module performance is crucial for designing full-scale FO plants. Previous theoretical FO module models involved intensive numerical calculations and complex geometry fitting.

However, the deep learning model has been optimised to minimise prediction errors and address overfitting issues [105]. The model has been trained using 116 experimental datasets, which included six input variables (flow rate, pressure and ion concentration of both the draw solution (DS) and feed solution (FS)) and one output variable (flux).

The optimised deep learning model has achieved a prediction error of 3.87%, significantly better than the theoretical FO module model, which had a prediction error of 10.13%. The deep learning method applied is based on artificial neural networks (ANNs). The model was trained to predict the flux of FO membrane modules by learning from the provided experimental data. The process involved adjusting parameters such as the number of hidden layers and neurons to optimise the model’s performance.

Table 4. Inputs and outputs of different AI methods for flux modelling in membranes.

Method	Inputs	Outputs
- NetLogo [96]	- Operational parameters	- Water flux
- Taguchi ANN [97]	- Feed velocity and temperature, draw velocity and temperature	- Solute flux selectivity
- MOPSO + FNNC [98]	- Dissolved oxygen and nitrate in the treatment process	- Flux modelling
- ANNs [99]	- Feed temperature, flow rate and TDS, transmembrane pressure (TMP), time	- TDS and flow rate
- ANNs + SVR [101]	- Past data of permeate flux and salt passage serves	- Permeate flux and salt passage prediction
- XGBoost [102]	- Feed solution and draw solution concentrations	- Flux modelling
- RSM + ANNs + ANFIS [103]	- Initial draw and feed concentration, time, pH and temperature on water flux and reverse salt flux	- Water flux
- ANFIS + ANNs [104]	- Suspended solids, electrical conductivity and dissolved oxygen	- Flux modelling
- DL [105]	- Feed flow rate, pressure, ion concentration of DS and FS	- Flux modelling

represents a summary of the described AI methods for flux modelling in membranes.

3.3. AI in Micropollutant Prediction in Water Purification and Desalination by Membranes

In terms of micropollutant prediction in membranes in water desalination and treatment processes, AI is applicable, too. In recent years, micropollutants have become more prevalent in rural areas due to diverse agricultural activities and crop types [106]. These pollutants negatively impact human health and ecological systems, necessitating effective management and monitoring [107]. Figure 4 illustrates the process of using machine learning methods to predict the behaviour of micropollutants in membranes.

In a recent study [109], researchers addressed the challenge of organic micropollutants (OMPs) in wastewater by employing an interpretable ML-assisted approach to optimise the rejection of OMPs in FO systems by utilising 18 influential factors related to membrane properties, OMPs properties and experimental conditions to develop 10 ML models. The optimal model, XGBoost-18, was identified based on performance metrics and further interpreted using Shapley additive explanations (SHAP). The study found that the McGowan volume and molecular weight of OMPs, zeta potential of the FO membrane surface and osmotic pressure of the draw solution significantly influenced OMP rejection. According to the results, the SHAP-XGBoost-11 model, which used 11 representative input features selected from the original 18 variables, achieved the most accurate prediction for OMPs rejection with an adjusted R² of 98%. The findings provide a new perspective for more efficient experimental optimisation of FO systems, aiming to achieve the highest rejection of target OMPs. The study also suggests a referential logical framework for expanding the FO process to broader applications. Using Shapley additive explanations (SHAP), key factors influencing OMP rejection, including the McGowan volume of OMPs, molecular weight, zeta potential of the FO membrane surface and osmotic pressure can be identified. The SHAP-XGBoost model demonstrates accurate predictions for OMP rejection. This study offers valuable insights for efficiently optimising FO systems and proposes a logical framework for expanding FO applications.

The removal efficiency of micropollutants in functionalised RO and nanofiltration (NF) membranes has been predicted using a small dataset and taking into account the inherent characteristics of both the micropollutants and the membranes as input for machine learning tools [110]. The selected machine learning algorithms include supervised methods such as adaptive network-based fuzzy inference system, linear regression, stepwise linear regression and multivariate linear regression, as well as unsupervised methods including support vector machine and ensemble boosted tree. Feature engineering reveals that micropollutant attributes, such as maximum projection diameter, minimum projection diameter, molecular weight and compound size, significantly impact removal efficiency. Combining these key variables can result in highly accurate predictions in both supervised and unsupervised machine learning. Notably, an NF-grid partitioning (NF-GP) model can achieve a high R² value with low error metrics.

To predict the concentrations of micropollutants (MPs) in a watershed deep learning models (long short-term memory (LSTM) and convolutional neural network (CNN)) have been employed and compared these predictions with those obtained from the soil and water assessment tool (SWAT) model [111]. Comparing the performance of these DL models with the SWAT model has shown negative Nash–Sutcliffe efficiency (NSE) values due to its limitations in simulating MPs. The results show that the LSTM model exhibited the highest performance with NSE values of 0.99 for training and 0.75 for validation. In terms of error analysis, in the multi-target MP model, the error decreased as the number of simulated pollutants increased. The simulation of four pollutants had the highest error, while the six-target simulation had the lowest error. Thus, the LSTM model has significant potential to improve the prediction of MPs in aquatic systems. The SWAT model’s poor performance can be attributed to simplified equations for micropollutant simulation and ambiguous pesticide application plans, leading to negative efficiency values.

Viet et al. [108] explored using machine learning (ML)-based models to simulate micropollutant (MP) behaviour in FO membrane-based water treatment processes. FO demonstrated extremely low rejection efficiency (2–80%) for low molecular weight MPs. Out of the eight machine learning models tested, the ensembles of trees (ET), adaptive-neuro fuzzy inference system (ANFIS) and Gaussian Process Regression (GPR) were the most effective in predicting MP behaviour. The optimised ANFIS model, particularly the one with a subtractive clustering radius of 0.1 (ANFIS-SC), exhibited outstanding performance in forecasting MP removal, with an R-value of 0.99 and an RMSE of 0.56%. The study also demonstrated the potential use of the developed ANFIS-SC model to simulate the effects of operational parameters on micropollutant (MP) elimination. This simulation can contribute to better design and more efficient operation of the FO system. Such modelling significantly enhances our understanding and optimisation of FO membrane processes for water treatment, particularly in improving micropollutant removal.

Table 5. Inputs and outputs of different AI methods for micropollutants prediction in membranes.

Method	Inputs	Outputs
- XGBoost-18 + SHAP	- Volume/molecular weight/zeta potential/membrane surface/osmotic pressure	- OMP rejection
- NF-GP	- Maximum projection diameter/minimum projection diameter/molecular weight/compound size	- OMP rejection
- ANFIS	- Feed temperature/flow rate/TDS/TMP	- OMP rejection
- ET	- Feed temperature/flow rate/TDS/TMP	- OMP rejection
- LSTM	- Inlet flow/inlet pressure/inlet temperature/surface of membrane	- Predicting MPs
- GPR	- Feed temperature/flow rate/TDS/TMP	- OMP rejection
- ANFIS + SC	- Operational parameters	- OMP rejection

represents a summary of the described models.

4. The Limitations and Challenges of AI in Predicting Membrane Behaviour and Future Directions

AI platforms offer numerous benefits but also come with drawbacks. When applied to predicting membrane behaviour, AI methods encounter limitations such as dealing with large, time-consuming datasets, platform complexity, overfitting, interpretability issues, as well as simplicity and lack of precision in some processes [112,113,114,115,116]. Table 5 outlines the advantages and limitations of applying AI concepts in the context of membranes. Some AI methods may result in overfitting for small input historical data, potentially leading to overgeneralisation [117,118]. Another common problem with some AI methods is the need for large datasets [119]. These methods require substantial data for training and tuning, which can be time-consuming, costly and risky for normal operation in some situations, posing potential dangers to water treatment processes. In certain visual modelling scenarios, AI may not have limitations in imaging the choroid, such as the OCT method. CNNs is one of the platforms which can effectively segment the membrane structure. Non-parametric methods such as SVR and ANNs may lack transparency, which is a significant drawback. Furthermore, in the optimisation of AI methods, the existence of local optima poses another challenge [120]. Local optima refer to the relatively best solutions within a neighbour solution set, which may not be suitable for specific situations.

Table 6. Positives and negatives of AI methods applied in predicting membrane behaviour.

Methods	Positives	Negatives
LSTM [121,122]	Data-based modelling/High accuracy/cost reduction	Large datasets/overfitting/interpretability
CNNs [123]	Feature extraction/translation invariance/transfer learning/segmentation	Large datasets/computational demands/overfitting
RNN [116,124]	Time-series data/flexibility/accuracy	Large dataset/complexity/interpretability
MLP [125]	Nonlinear relationships/universal approximators/feature extraction/flexibility	Large datasets requirement/overfitting/complexity of tuning
NetLogo [126,127]	Visualisation/Agent-Based Modelling/user friendly	The simplicity of modelling/Limited precision/not common platform
Taguchi ANN [115,128]	Accurate optimisation/applying setting parameters/numerical modelling	Simplicity of modelling/Data requirement/time-consuming/interpretability
MOPSO [129]	Robustness/optimisation/high-speed	Tuning complexity/large data requirement/local optima
ANNs [112,114]	Nonlinear mapping/Noise handling/multitasking	Data requirement/interpretability/complexity of tuning
SVR [113]	Effective real-value estimation/spares solution/Kernel function	Large datasets/lack of transparency/high-dimensionality difficulties
XGBoost [130,131]	High accuracy/handling missing values/regularisation/scalability/	Complex tuning/spares data/expensive
RSM [132,133]	Systematic experiment design/optimisation/interpretable/	Large datasets requirement/assumption/limited to polynomial models/local optima
ANFIS [134,135,136]	Hybrid approach/efficient resource/data-driven learning/	Resource intensive/large dataset/overfitting
DL [137]	Scalable/feature extraction/high performance	Computationally expensive/large dataset/interpretability/overfitting

presents a summary of the positive and negative aspects of using AI methods in membrane modelling.

In Table 5, the limitations of AI methods were highlighted, including the need for extensive datasets, interpretability, complexity of tuning, local optima, computational complexity, simplicity of modelling and overfitting. These challenges affect the modelling and simulation of membrane behaviour, critical for improving water purification and desalination processes. Future research should prioritise addressing these issues to advance the application of AI in this field. Dealing with the dataset size requirement could decrease costs and operational risks related to training AI methods. Moreover, streamlining the tuning process of these methods would enhance user-friendliness and reduce the likelihood of human errors.

5. Conclusions

The research studied the critical issue of water scarcity and investigated various membrane separation methods for water purification and desalination, specifically focusing on RO, FO, NF, UF and MF techniques. Additionally, it explored eco-friendly membrane separation methods for these processes.

Furthermore, the comprehensive review delved into AI concepts such as ML and ANNs, emphasising their strategies and algorithms for modelling and predicting membrane behaviour. It highlighted the diverse applications of AI in membranes and underscored its significant role in predicting membrane behaviour for water purification and desalination. Using AI techniques, membrane processes can accurately model water flux, predict fouling and simulate micropollutants using operational and water quality data. This not only enhances process efficiency and accuracy but also addresses critical challenges in water treatment.

Finally, the research suggested that future efforts should prioritise refining AI models to minimise required datasets, improve the interpretability of methods and simplify the tuning of AI concepts. This approach aims to further enhance membrane performance and sustainability in water purification and desalination systems.