Machine learning-driven design of membranes for saline and produced water treatment across scales

Abstract

Membrane-based desalination technologies are critical for addressing global water scarcity; however, their performance optimization remains a complex challenge. This review explores the potential of machine learning (ML) in advancing membrane desalination, with a focus on process optimization, fouling mitigation, and performance prediction. We provide a comprehensive analysis of ML techniques, including artificial neural networks (ANNs), support vector machines (SVMs), and random forest (RF), applied to membrane systems such as reverse osmosis (RO), forward osmosis (FO), and for the treatment of produced water. The study highlights how data-driven models enhance decision-making by correlating operational parameters (e.g., pressure, temperature, feed salinity) with membrane efficiency, energy consumption, and fouling behavior. A key emphasis is placed on ANN-based frameworks for real-time monitoring and predictive control, demonstrating their superiority in modeling non-linear interactions compared to traditional mechanistic approaches. We also examine ML applications in optimizing design parameters, maintenance strategies, and renewable-energy-integrated desalination systems. Challenges such as data scarcity and model generalizability are discussed, alongside future directions for integrating ML with emerging membrane materials and hybrid processes. This review highlights ML's role in bridging theoretical and practical gaps in desalination, offering actionable insights for researchers and industry experts to deploy intelligent and sustainable water treatment solutions.

---

Uzair Ahmad

Uzair Ahmad is pursuing his PhD in Chemical and Environmental Engineering under the supervision of Dr. Faheem Hassan Akhtar at the Lahore University of Management Sciences (LUMS). His research focuses on machine learning applications in membrane-based water desalination, with an emphasis on predictive modeling and optimization of membrane performance.

Faheem Hassan Akhtar

Faheem Hassan Akhtar is currently a faculty member and Program Head of the Chemical Engineering Program, Lahore University of Management Sciences (LUMS, Lahore). He received his PhD degree in Chemical and Biological Engineering from the Advanced Membranes and Porous Materials Center, King Abdullah University of Science and Technology (KAUST, Saudi Arabia). After postdoctoral research (Water Desalination and Reuse Center), he started his independent academic career as Assistant Professor at LUMS in 2021. His research focuses on developing strategies for the rational design of high-performance membranes and composite adsorbent technologies for energy and environmental sustainability applications.

---

Water impact

Removing salt and other impurities from water remains a significant challenge in today's world. Machine learning notably helps to improve water treatment and desalination by making processes more adaptable, efficient and cost effective. Recent advancements in operational parameters, system optimization and membrane fouling prediction were explored. This study offers a holistic methodology to develop smart water technologies in real world applications and contribute to addressing global water issues by enhancing the performance of desalination systems using machine learning.

1. Introduction

Water scarcity is a significant issue impacting millions of individuals worldwide. As freshwater supplies get less, it is important to find new ways to meet the growing need for clean water. The term “water scarcity” describes a situation where there is an insufficient supply of freshwater to satisfy people's needs for things like drinking, farming, and industrial processes. Climate change, ineffective water management policies, urbanization, and an increase in population contribute to this problem. For direct human consumption, only a small amount (2.5% of Earth's water) is suitable.^1,2 At present, there exists a substantial mismatch between the global demand for clean water and its availability, leading to a situation where roughly a quarter of the world's population experiences economic water scarcity.³

Desalination, the process of purifying water by removing salt and other contaminants, can be used to lessen the impact of water scarcity. Desalination methods can be categorized into two main groups: thermal and membrane-based processes. Since the 1960s, membrane processes, such as RO, have surpassed thermal processes for new plant installations globally,⁴ where thermal desalination has remained the predominant technology. The primary disadvantage of desalination is associated with costs;⁵ 30% of the total cost of desalinated water is associated with electricity for seawater desalination using RO. Increased energy consumption is accompanied by a rise in greenhouse gas (GHG) emissions.⁶

The overall energy demand for RO seawater desalination, encompassing both pre- and post-treatment processes, generally ranges from 3 to 6 kWh m⁻³.^7,8 To enhance energy efficiency in RO applications, there has been a growing emphasis on utilizing feedback controls tied to the salinity of the incoming feedwater as a means to reduce desalination energy consumption.^9,10 In the past few years, significant improvements have been achieved in the permeability and salt rejection capacity of membranes used to treat high salinity streams. The energy utilization for RO seawater desalination remains greater than the theoretical minimal energy demands. The average energy system use is 3.1 kWh m⁻³. Developments in desalination membranes hold the potential to reduce energy consumption. New desalination technologies reduce the required feed pressure while maintaining salt rejection. The global installed desalination capacity saw a significant increase from 27.05 (million m³ per day) in 2003 to 98.93 (million m³ per day) in 2023.

In recent years, ML has gained significant importance and has shown great potential in various domains, including water treatment technologies as shown in Fig. 1. ML methods, like support vector machine (SVM), artificial neural networks (ANN), and deep learning algorithms, have demonstrated the ability to extract valuable insights from complex data and optimize system performance.

Fig. 1 Increase in publication on desalination using machine learning algorithms. The keywords used are (“machine learning” OR “artificial neural network”) AND (“water treatment” OR “desalination” OR “produced water”). The data is retrieved based on Scopus database (data retrieved on Dec 31, 2024).

The application of ML in membrane desalination offers several advantages. ML algorithms can process vast amounts of data collected from sensor arrays, monitoring systems, and historical records to identify patterns, correlations, and trends that may not be apparent through conventional analysis methods. By leveraging this data-driven approach, ML algorithms can learn and adapt to the dynamic behavior of desalination processes, enabling real-time monitoring, predictive modeling, and advanced control strategies.

ML techniques can be employed to enhance the optimization of critical operational variables, including but not limited to feedwater flow rate, pressure, temperature, and chemical dosage. ML algorithms can provide invaluable insights into the relationships between these variables and the membrane desalination system's performance by perpetually analyzing process data and environmental factors. This information can aid in decision-making and facilitate modifying operating conditions to enhance efficiency, water recovery, and fouling potential.

In this review paper, we aim to explore the current state of research and development in machine learning for process optimization and modification in membrane desalination for produced water and saline wastewater treatment as shown in Fig. 2. We will examine the various ML techniques, advantages, and limitations, demonstrating their application in real-world desalination plants. Additionally, we will discuss the challenges associated with implementing ML in desalination systems and potential avenues for future research and development.

Fig. 2 The disciplinary fields and their correspondence: our review.

2. Saline wastewater streams

Saline wastewater is a water stream containing dissolved salts and diverse contaminants. Salinity can be quantified by the total dissolved solids (TDS) concentration, expressed in milligrams per liter (mg L⁻¹), an indicator of the quantity of soluble salts present. Soluble salt concentration can be determined by measuring the electrical conductivity (EC), total dissolved solids (TDS), and salinity.

Wastewater sources include those originating from industrial (tannery, food, paper, textile, oil and gas), agriculture activities, and domestic wastewater streams.

Saline wastewater is generated from agricultural drainage originating from saline and contains fertilizer nutrients, salt ions, herbicides, and pesticides.¹¹ In the food processing industry, using brine and salt in processing plants that handle pickled vegetables, dairy products, seafood, and canned meats can lead to the generation of saline wastewater.¹²

While conducting oil drilling operations, the oil and gas sectors simultaneously produce a saline byproduct called flow back water or produced water.¹³ Shale reservoirs have the potential to produce a significant amount of water per well, ranging from 1.7 to 14.3 million liters, in the initial 5–10 years of production.¹⁴ The specific composition of the produced water is contingent upon factors such as geographical location, hydrocarbon classes, and reservoir age.^15,16

In addition to the wastewater generated by the industries mentioned earlier, landfill leachate exhibits traits typical of saline wastewater. This is primarily attributed to ammonia, inorganic salts, organic pollutants, heavy metals, and biological organisms.¹⁷ It has been reported that as landfill leachate ages, its organic constituents will become more complex and chemically stable.¹⁸ Therefore, it is extremely difficult to treat this form of wastewater. Table 1 below shows the different compositions of saline wastewater from industrial resources. Different compositions like COD, TDS, BOD and TOC are mentioned in Table 1. Oil and gas industries have the highest amount of TDS which is 100 g L⁻¹ followed by food industries (dairy wastewater) with the value of 48 g L⁻¹.

Table 1 Composition of saline wastewater from industrial sources

Sources	Constituents	Composition, g L^−1				Ref.
		TDS	COD	BOD	TOC
Leather/tannery industries, soaking process	Salts, ammonium, phosphorus, salts, chromium, organic nitrogen, sulfide	22–33	3–6			19
Textile industries, dyeing process	Dyes, nitrogen, salts, heavy metals (Cu, Sb, Pb, Cr), suspended solids	1.5–4.0	0.4–1.4			20, 21
Textile industries, bleaching process		2.5–11.0	1.2–1.6
Food industries: dairy wastewater	Soluble proteins, lactose, mineral salts, lipids	48	29			21–24
Food industries: seafood wastewater	Oils, fats, suspended solids, excess nutrients (phosphorus and nitrogen), sodium chloride, proteins		13.18	3.25
Landfill leachate	Ammonia, inorganic salts, organic contaminants, heavy metals, and biological organisms	10.7–35.8 mS cm^−1	2.04–69.6			25, 26
Paper industries	Phenols, lignin, sulfur compounds, furans, chlorides, acetic acid, phosphate	0.40–2.5	0.48–4.45			27
Oil and gas industries	Heavy metals, toluene, benzene, grease, aromatic hydrocarbons, salts, oil, carboxylic acid	100			0.5–2	28, 29

---

2.1. Membrane-based techniques for treatment of saline wastewater streams

Recent advances in wastewater treatment include emerging technology like membrane distillation (MD) and established technology like the membrane bioreactor (MBR). Additionally, various other membrane-based methods, such as ultrafiltration (UF), microfiltration (MF), nanofiltration (NF), forward osmosis (FO), RO, membrane contractors (MC), and electrodialysis (ED), have been reported for wastewater treatment applications.³⁰ It's worth noting that ultrafiltration (UF) membranes are predominantly used in MBRs and play a crucial role in wastewater treatment, as shown in Fig. 3.

Fig. 3 Desalination using membrane techniques for wastewater treatment.³¹ (Copyright 2021, Elsevier).

Given the negative effects saline wastewater has on the environment, treating it by removing salt and contaminants is essential before releasing it into the environment. If the water quality meets certain standard criteria, the treated wastewater could be further reused, offering an alternative to freshwater supplies. Because of their high permeate quality, adaptability, and desalination potential in water treatment, membrane-based technologies are currently attracting a lot of interest.³² The different membrane-based methods that have been used to treat saline wastewater are discussed below in Table 2. Different membrane processes like FO, RO, MD and MBR are discussed below. Characteristics of different saline wastewater sources are also explained, which includes COD, TOC and EC. In the case of COD and TOC rejection, the FO process shows 90% effectiveness. RO processes can achieve water recovery greater than 83%.

Table 2 Application of membrane processes in saline wastewater treatment

Saline wastewater types/source	Wastewater/feed solution characteristics	Membrane process	Performance	Ref.
Dyeing wastewater from textile and printing industries (China)	COD = 398.11 mg L^−1	FO	90% rejection for TOC and COD	33
	EC > 100 mS cm^−1
	TOC = 170.70 mg L^−1
Landfill leachate (Virginia, USA)	COD = 69 470 mg L^−1	FO	Water recovery = 36.6%	34
	EC = 35.8 mS cm^−1
Landfill leachate (anaerobic landfill in Southwest China)	EC = 28.1–31.8 mS cm^−1	RO	Water recovery >83%	35
			EC = 0.15–0.22 mS cm^−1
Textile wastewater (Turkey)	COD = 1000 mg L^−1	RO	EC = 300 μS cm^−1	36
			Water recovery 70%
	EC = 5500 μS cm^−1
Textile wastewater (from factory in Johor, Malaysia)	COD = 405–477 mg L^−1	MD	Rejection efficiency: COD = 89.6%	37
	EC = 1294–1673 μS cm^−1
			EC = 93.7%
Crude olive mill wastewater (from Morocco region)	COD = 160 g L^−1	MD	Permeate flux = 3.9 LMH	38
	EC = 23 mS cm^−1
Textile wastewater (woolen textile dyeing factory, Turkey)	COD = 750 mg L^−1	MBR	Total phenol removal 80–87%	39
	EC = 3680 mS cm^−1
	Total phenol = 1610 mg L^−1
Tannery wastewater (Tamil Nadu, India)	Soluble COD = 7560 mg L^−1	MBR	Permeate flux 6.8 LMH	40
Bleach plant effluent (from paper/pulp industry, South Africa)	Total COD = 1700 mg L^−1	MBR	Flux = 6 LMH	41
	EC = 1196 mS m^−1

---

3. Membrane based produced water treatment

From a practical point of view, the efficiency of produced water (PW) treatment is highly dependent on the PW composition and the membrane properties. Diverse strategies have been implemented to improve the efficiency of membrane-based PW desalination methods.

This section analyzes the different types of membranes which are used for the treatment of produced water.

3.1. Applications of reverse osmosis (RO) membranes in treatment of PW

As illustrated in Table 3, commercial RO membranes are currently used for the treatment of PW for desalination and removal of hydrocarbon oils. In the case of TDS rejection, the performance of the commercial RO double filtration membrane is exceptional. The rejection rate in the case of TDS removal is 99.8%. Commercial RO (RE-4040) shows the highest rejection rate of TOC which is approximately 99.9%. These membranes also show the highest value of flux which is 50–55 L m⁻² h⁻¹. Nonetheless, fouling remains the primary limitation of the RO process for desalination and most of the research efforts are focused on membrane modification to enhance antifouling and advanced cleaning strategies.

Table 3 Performance evaluation and advances of PW using RO

Membrane	Feed, ppm		Flux, L m^−2 h^−1		Rejection, %		Ref.
	Type	TDS	TOC		TDS	TOC
Commercial RO double filtration	Synthetic PW	37 500	565	—	99.8		42
Commercial RO (BW-30)	Real PW	722	68.8	30–35	87		43
Commercial RO (TFC-HR)	Real PW	5243		>15	91.4	80.4	44
Commercial RO (RE-4040)	Synthetic PW	4800–5500	9–13	50.0–55.0	96.6	∼99.9	45
PSF-PAMAM + TMC	Synthetic PW	2000 (NaCl)	2500/5000 ppm hexadecane + 250 000 ppm SDS	18.42	89.3	99	46

---

The efficiency of RO membranes is significantly influenced by the concentration of impurities, hydrocarbons and ions, in the PW feed solution. Furthermore, the extent of membrane fouling substantially affects both process costs and the required cleaning frequency. Nevertheless, when evaluating the quality of the treated water (permeate) produced through the RO process, it becomes evident that RO remains a reliable method for treating PW, especially when the membrane is appropriately customized.

3.2. Applications of forward osmosis (FO) membranes in treatment of PW

The efforts made in produced water desalination using FO are summarized in Table 4. Several strategies have been implemented to achieve high efficiency, including system optimization, membrane development, fouling resistance, and draw solution (DS) improvement. TFC (Hydration Technology Innov) membranes show the highest values of rejection both in the case of TDS and TOC which are 97% and 99.9%, respectively. The highest value of flux is achieved in the case of PES-PVA-GO nanosheet membrane which is 16.05 L m⁻² h⁻¹.

Table 4 Performance evaluation and advances of PW using forward osmosis

Membrane	Feed, ppm			Draw solution	Flux, L m^−2 h^−1	Rejection, %		Ref.
	Type	TDS	TOC			TDS	TOC
TFC HF (Asahi Kasei Corp)	Synthetic PW	14 700	46	NaCl	11		90	56
TFC (Hydration Technology Innov.)	Real PW	24 646	95	NaCl	7.2	95 (multivalent)	>80	57
PES-PVA-GO nanosheets	Synthetic PW	256 000	2.5–100	Na2 SO4	16.05		99.9	58
TFC (Hydration Technology Innov.)	Real PW			NaCl	8	97	∼99.9	49
TFC HF (Asahi Kasei Corp)	Synthetic PW			NaCl	11.2			59

---

In addition to designing an efficient forward osmosis membrane, draw solution improvement plays a key part in enhancing the overall efficiency of the FO produced water desalination.^47,48 When FO is operated as a stand-alone process, water permeates through a semi-permeable membrane, causing a rise in the concentration of the feed solution and a fall in the concentration of the draw solution. This ultimately lowers the osmotic pressure.⁴⁹ Hence, DS should undergo reconcentration to restore its osmotic pressure, ensuring a viable and environmentally sustainable operation.^50–52 Important criteria for DS^53–55 include low molecular weight (MW), high solubility, non-toxicity, and excellent compatibility with membranes; however, the cost involved and regeneration are also essential factors to consider.

3.3. Applications of membrane distillation in treatment of PW

The studies related to PW desalination using membrane distillation (MD) with direct contact membrane distillation (DCMD), vacuum membrane distillation (VMD), and air gas membrane distillation (AGMD) are shown in Table 5. In the case of MD, the salt rejection is more than 99%. PVDF functionalized with KOH shows the highest value of salt rejection, which is 99.99%, followed by PVDF incorporated with silica membranes which shows 99.98% salt rejection.

Table 5 Performance evaluation and advances of PW desalination using MD

MD configuration	Membrane type	Feed	Flux (L m^−2 h^−1)	Rejection (%) salt
DCMD	PVDF functionalized with KOH	Synthetic PW	45	99.99
AGMD	Commercial PVDF	Synthetic PW	15.5	99.2
DCMD	PVDF + PVA	Real PW	30.3	99.9
VMD	Commercial PVDF	Synthetic coal shale gas	—	99.97
DCMD	PVDF incorporated with silica	Synthetic PW	28.6	99.98

---

4. Machine learning techniques

The four primary classifications of ML algorithms are unsupervised, semi-supervised, supervised, and reinforcement learning (RL), as shown in Fig. 4.

Fig. 4 Different types of machine learning techniques used in membrane desalination.

Unsupervised learning is a computational approach that utilizes unlabeled datasets for analysis without human interventions. This method is commonly used for grouping results, identifying meaningful structures and trends, and extracting generative features. Dimensionality reduction, feature learning, density estimation, and clustering are the most common unsupervised learning tasks. It discovers hidden patterns in unlabeled data but lacks ground truth validation. Semi-supervised learning operates on both unlabeled and labeled data.⁶⁰ It can be defined as a hybrid of the unsupervised and supervised learning methods. Therefore, it lies between “without supervision” and “with supervision”. In numerous practical scenarios, the availability of labeled data may be limited, while unlabeled data is abundant, rendering semi-supervised learning useful.⁶¹ The objective of a semi-supervised learning model is to generate more accurate predictions than would be possible using the model's labeled data alone. It has applications in data labeling, fraud detection and machine translation.

The primary aim of supervised learning ML is to acquire knowledge of a function that establishes a correspondence between an input and an output, utilizing a set of input–output pairs as an example. The system employs a collection of training examples and labeled training data to infer a function. In a task-driven approach, supervised learning is carried out when specific objectives are to be achieved from a specific set of inputs that are identified.⁶⁰ The most typical supervised activities are “regression”, which fits the data, and “classification”, which separates the data.

Another method is reinforcement learning (RL), which facilitates the automatic determination of optimal behavior for machines and software agents within a given context or environment, aiming to achieve a desired objective.⁶² This form of learning relies on reward or punishment. The ultimate objective of the penalty is to use insights gained by environmental activists to enhance their efforts to maximize the reward or minimize the risk.⁶¹ “It is a powerful tool for training AI models capable of boosting automation and optimizing the performance of intricate systems like robotics and autonomous vehicles. However, it finds limited application in addressing basic or uncomplicated issues within manufacturing and supply chain logistics”.

Various ML techniques can substantially impact the development of an efficient model in diverse application domains, depending upon the characteristics of the data and the intended objectives. A comparison of different ML models is given in Table 6. Different AI models like ANN, SVM, KNN, RF and DTR are compared based on applications, data needs and computational cost.

Table 6 Comparison of different machine learning models⁶³

Model	Applications	Data needs	Computational cost
Artificial neural network (ANN)	Regression, classification	Very large	Very expensive
Support vector machine (SVM)	Pattern recognition, regression, classification	Moderate	Medium
K nearest neighbor (KNN)	Regression, classification	Large	Expensive
Random forest (RF)	Regression, classification	Moderate to large	Low to medium
Decision tree forest (DTR)	Regression, classification	Small to moderate	Low

---

5. Artificial neural networks

The artificial neural network (ANN) is a ML method that emulates the operations of biological neurons. This computational model employs a learning approach that mirrors the cognitive mechanisms of the human brain, allowing it to tackle a wide array of tasks and problems. The ANN is a commonly employed tool in the fields of AI and ML.⁶⁴ The ANN can be considered a black-box (BB) model due to its lack of adherence to physical rules, and its parameters may lack physical significance.⁶⁵ The principal objective of the ANN is to create a correlation between input variables, which are considered independent, and output variables, which are considered dependent. The establishment of this link is facilitated by a learning mechanism wherein the network is presented with a dataset; therefore, this model is data driven. ANNs have gained significant popularity as a viable alternative because of their exceptional precision, reduced computational requirements, and capacity to effectively capture non-linear correlations between input and output variables inside a system. In addition to function approximation, other applications of ANNs include forecasting, clustering, classification, image processing and pattern recognition.

Various types of ANNs exist that can be classified according to their model parameters, design, and topology. The back propagation (BP) feedforward neural network, also referred to as the multi-layer perceptron (ML P), is widely recognized as the most frequently employed ANN. Additional types of ANNs include the recurrent neural network (RNN), radial basis function neural network (RBFNN), deep neural network (DNN), and Elman neural network (ENN). These types exhibit modifications in their structural composition, model parameters, and training algorithms while still belonging to the broader category of multi-layer perceptron (ML P) ANN.

A MLP comprises three essential components: input, hidden, and output layers. The number of input parameters corresponds to the number of neurons in the input layer, while the number of output parameters matches the neuron count in the output layer. Depending on how well-designed the ANN model is, one or more concealed layers may be composed of multiple neurons. The weights provided to each neuron during training set the hidden layer and output layer values. Each neuron in the hidden and output layers also receives a weighted unit input with a constant value called the threshold value or bias. Fig. 5 shows the basic layout of the ML P.

Fig. 5 A single hidden layer ML P structure having 4 inputs, 4 hidden layers and 2 output neurons (4–4–2). (Copyright 2021, Elsevier).

The radial basis function neural network (RBFNN) modeling technique can be considered an alternative to ANNs. The network resembles an ML P, structured as a feedforward network comprising input, output, and hidden layers. The output of each neuron is dictated by a basis function that exhibits radial symmetry. The RBFNN exhibits the ability to achieve universal approximation while avoiding the presence of local minima issues.⁶⁶

In addition, it has a simplified structure and converges faster than other neural networks. Instead of employing inputs and their respective weights, the training procedure of the RBFNN relies on the standard Euclidean distance, which measures the distance between inputs and the network's center points. This approach effectively maps input neurons to a hidden layer that exhibits non-linear characteristics. An activation function is then applied to compute the value of each node within this hidden layer. Like an ML P, the network's output is computed as a weighted sum of the hidden neuron activations. The hidden neuron numbers are established through the training process. In each iteration, one RBF neuron is gradually integrated into the network. A single RBF neuron is incrementally incorporated into the network during each iteration. This process continues until either the error aim is achieved or the maximum allowable number of neurons⁶⁷ is reached. Fig. 6 shows an RBFNN model.

Fig. 6 Structure of an RBFNN model (copyright 2021, Elsevier).

6. Applications of artificial neural networks in water treatment membranes

Abbas and Al-Bastaki⁶⁸ proposed a ANN model for saline and brackish water RO that accounts for a variety of operating conditions. To evaluate the capability of interpolation, the authors utilized the best model to predict the permeate flux using untrained data, resulting in an R² value of 0.98. However, extrapolation revealed that the model's predictions were inaccurate.

Using a year's worth of operational data, Lee et al. have developed an ANN model to enhance the operation of a seawater RO desalination plant, using temperature control.⁶⁹ Due to the difference in variance between the outputs, the model demonstrated the capability to determine permeate TDS with greater accuracy than flow rate. The researchers utilized the model to examine the impact of temperature, facilitating the development of an optimization method. This technique involved implementing a linear increment in feed temperature control to get the minimum permeate TDS concentration. Righton created a neural network model for RO groundwater and wastewater treatment.⁷⁰ Furthermore, the ANN technique was employed to model two RO desalination techniques, resulting in accurate salt rejection and permeate flux predictions (Fig. 7).

Fig. 7 Application of ANNs in water treatment membranes.

The study conducted by Libotean et al.⁷¹ provides evidence that implementing an ANN model effectively predicts the operational efficiency of an RO plant. This prediction is made by considering the changes in salt passage and permeate flux. The study assessed the efficacy of three forecasting methods: the marching forecast, sequential forecast and the standard time-series correlation (STSC). A comparative analysis was conducted between the ML P model and SVR model. Using short-term memory intervals of up to 24 hours, the study demonstrated that it was possible to make reasonably accurate forecasts using the models.

Barello et al. conducted a modeling analysis of an RO desalination process to predict water permeability by considering membrane type, salinity, pressure, and contamination.⁷² To facilitate the training of the ANN, the researchers carried out an extensive study to evaluate the impact of various parameters, including the number of hidden layers, number of neurons, and choice of transfer function, on the selection of an ideal network configuration. The performance of a model utilizing a tan-sigmoid transfer function and the minimum number of hidden layers was improved. Examining the quantity of neurons in the layer exhibited oscillatory phenomena; nevertheless, it was observed that a concealed layer consisting of 20 neurons yields a minimal error. Furthermore, the ANN calculated water permeability was verified by comparing it with correlations reported in the scientific literature.

Farahbakhsh et al.⁷³ applied ANN modeling to a RO process to investigate the antifouling properties of a membrane prepared by merging polypyrrole (PPy)-coated multiwalled carbon nanotubes (MWCNTs). For both unprocessed and oxidized MWCNTs-PPy membranes, two neural networks were constructed. The model demonstrated that the oxidized MWCNTs-PPy membrane generated higher water flux with a constant trend. Additionally, the aggregation of molecules on the membrane's surface was predicted to reduce salt rejection and water flux. Table 7 below shows the different features of ANN models developed for membrane processes. MLP with Levenberg Marquardt as a training algorithm shows the highest value of R² which is 0.9964. The input parameters for this study were membrane type, salinity, time and operating pressure while the output was water permeability.

Table 7 Features of ANN models developed for membrane processes

Process	Method	Input	Output	Activation	Training algorithm	Result	Ref.
Groundwater desalination plant	RBFN	—	Permeate TDS, permeate flow rate	—	—	Error = 1.73	74
RO desalination plant	ML P	Flow rate, pressure, pH, temperature	Product conductivity, product flux	tan-sigmoid	Gradient-descent algorithm	R ^2 = 0.916	74
RO water desalination unit	ML P	Temperature, pressure, and salt concentration	Permeate rate	log-sigmoid	Levenberg Marquardt	R ^2 = 0.988	68
Separation of CaCl2 and NaCO3	ML P	Feed concentration, flow rates, pressure	Solute rejection percentage, permeate flux	log-sigmoid	Levenberg Marquardt	R ^2 = 0.9889	70
Seawater desalination plant	ML P	Feed TDS, temperature, flow rate, TMP, time,	Flow rate, permeate flux	log-sigmoid	Back propagation algorithm	R ^2 = 0.96	69
Brackish water desalination plant	ML P	Conductivity, flow rate, pressure, temperature, pH	Salt passage, permeate flux	tan-sigmoid	Levenberg Marquardt	Error = 0.9%	71
RO desalination process	ML P	Membrane type, salinity, time, operating pressure,	Water permeability	tan-sigmoid	Levenberg Marquardt	R ^2 = 0.9964	72
Small and large-scale brackish water desalination	RBFN ML P	Conductivity, temperature, pressure, pH	Permeate flow, TDS	tan-sigmoid/Gaussian kernel	Levenberg Marquard	R ^2 = 0.9873	75
Brackish water desalination plant	—	TMP, time, temperature, concentration	Water flux		Levenberg Marquard	R ^2 = 0.9947	73
Bench-scale FO unit	ML P	FS viscosity, osmotic pressure difference, DS viscosity, DS temperature, FS temperature	Membrane flux	tan-sigmoid	Levenberg Marquard	R ^2 = 0.9803	76

---

6.1. Machine learning applications in desalination

ML is useful in desalination systems. ML algorithms can aid in improving the effectiveness and performance of desalination processes by utilizing the potential of artificial intelligence and predictive modeling (Fig. 8). To create precise forecasts and improve system parameters, these algorithms can analyze vast data, including membrane characteristics, feedwater quality, performance indicators, and operating circumstances.⁷⁷

Fig. 8 Applications of ML in desalination systems.

6.2. Performance predictions using operational parameters

Table 8 provides an overview of the predominant methodologies employed in studies focusing on predicting desalination system performance. As evident from the table, most of these studies have opted for ANNs as their preferred tool.^78–80 This preference can be attributed to the inherent advantages of ANN models, such as their exceptional ability to forecast the performance of non-linear systems, coupled with their capacity for high-level generalization and accuracy.

Table 8 Summary of research findings on operational parameter use for predicting desalination plant performance

System type	Data driven method	Dataset size	Inputs	Outputs	Main remarks	Ref.
RO	ML PANN, RBFANN	—	Pressure, temperature, conductivity, pH	Permeate TDS, permeate flowrate	The data set is divided into two parts: 70% training and 30% testing	75
					When it is trained using LM algorithm, the performance was seen to be better
					RBF network is trained using OLS backpropagation model
RO	ML PANN	97–129	Temperature, influent concentration, influent flow, recovery percentage	Effluent TDS	The behavior of RO membranes was predicted to be using computational models	82
					LM algorithm is used to train the neural network
					The neurons in the first and second hidden layers ranged from 4 to 6 and 8 to 13, respectively
Multi-stage flash (MSF)	RBFANN	380	Salinity, boiling point temperature	Temperature elevation	The structure of the ANN is defined as a 2–12–1 configuration	83
					The data has been split into three partitions: 70% for training, 15% for validation, and 15% for testing
					The RBFANN model had superior prediction performance when compared to the ML PANN model, thermodynamic models, and empirical correlations
RO	ANN	9	TDS inlet, inlet concentration, time	Permeate flow	The dataset was divided into three parts for training, validation, and testing, with a split ratio (SR) of 33.3% for each	84
					Four neurons made up the first hidden layer, while three neurons made up the second
					The artificial neural network (ANN) utilizing a fitness approximation network achieved the lowest mean squared error (LMSE) and R^2
RO	ML PANN with GA	70	Inlet pressure, inlet flow rate, inlet temperature, inlet concentration	Response parameter of chlorophenol rejection	The data has been split into three partitions: the training set consisting of 68.57% of the data, the validation set consisting of 15.71% of the data, and the test set consisting of 15.71% of the data	85
					The model's performance was assessed by testing it with two distinct hidden layer configurations: one containing 2 neurons and the other containing 8 neurons
RO	ANN	63	Feed pressure, salt concentration, temperature	Water permeate rate	The first and second hidden layers consist of 3, 5, 10, and 15 neurons respectively	68
RO	ML PANN	1806	Raw water salinity, plant location, plant type, plant capacity	Capital cost of plant	The suggested model has the potential to be utilized to generate a rational approximation of the investment expenses associated with future RO plant developments	86
MD	ANN	252	Feed salt concentration, vacuum pressure, feed flow rate, feed salt concentration	Permeate flux	The data has been split into three sets as follows: the training set, accounting for 66% of the data; the validation set encompasses 17% of the data, and the test set likewise comprises 17% of the data	87
RO	SVR	3990	Flow rate, conductivity, pressure	Permeate conductivity, permeate flow rate, retentate conductivity, retentate flow rate	The data was divided into two sets, with 60% allocated for training purposes and 40% reserved for testing	81
					The construction of steady state and transient models for a RO facility was undertaken
					A proposed approach for time forecasting was introduced to demonstrate the temporal variation in conductivity during transient operation
MD	ANN	38	Feed inlet temperature, vacuum pressure, feed inlet temperature, feed salt concentration, feed flow rate	Permeate flux	The design of the ANN under consideration follows a 4–3–1 configuration	78
					The data has been divided into three sets based on the following split ratio: 70% for training, 15% for validation, and 15% for testing
					The parametric research conducted
					using the constructed ANN model revealed that the permeate flux was most significantly influenced by vacuum pressure and feed input temperature
RO	ANN	436	Dipole moment, molecular width, molecular length, molecular weight, salt rejection, pressure, surface membrane charge	Rejection rate	The data was split into three sets: the training set, representing 80% of the data; the validation and testing set, comprising 20% of the data; the BANN model demonstrated superior performance compared to both the single neural network and BA ML R techniques	78
RO	ANN, RF, SVR	—	Feed conductivity, feed temperature, electrical power	Feed flow rate, pressure, permeate conductivity, permeate flow rate, water productivity	Random forest (RF) and support vector regression (SVR) models demonstrate superior predictive capabilities in relation artificial neural networks (ANN), with statistical significance at a 5% level	88
RO	SVR, ANN	10 min steps for 3 months	Feed pressure, feed conductivity, feed flow rate, feed temperature, permeate pressure, permeate flow rate, permeate conductivity	Salt passage, permeate flux	The data has been divided into three sets with the following split ratios: 40% for training, 10% for validation, and 50% for testing	89
					The findings demonstrated strong predictive validity in projecting short-term memory time intervals ranging from 8 to 24 hours for permeate flux and salt passage, with a forecasting horizon of up to 24 hours
RO	ML PANN-PSO, DT, SVM	150	Temperature, feed pressure, conductivity	Permeate flow rate, permeate TDS	The architecture of the artificial neural network (ANN) is structured as a 4–3–1 configuration	71
					The hybrid ML PANN-PSO model demonstrated superior performance compared to the DT and SVM models
					The hybrid model demonstrated a reduction in uncertainty when applied to the simulated data

---

Although artificial neural networks (ANNs) are extensively utilized, various ML models have also been employed to predict the efficiency of different desalination systems. Pascual et al.,⁸¹ for instance, used support vector regression (SVR) to forecast the efficiency of an RO plant. Their findings concluded that SVR exhibits the potential to predict crucial parameters like conductivity and flow rate for both retentate and permeate streams, achieving impressively low average absolute relative errors ranging from 0.70% to 2.46%.

In Table 8 various ML models, such as support vector machines (SVM), artificial neural networks (ANN), and random forest (RF), were studied for different membrane processes, including RO, multi-stage flash (MSF), and membrane distillation (MD). Different output parameters were analyzed, and these models were used to optimize the operational parameters of the respective processes.

6.3. Performance prediction using design parameters

Table 9 summarizes the research that has attempted to forecast the desalination system performance using design factors as inputs to their models. Notably, only a limited subset of research investigations has ventured into integrating design parameters into the prediction process for desalination system performance.

Table 9 Summary of research findings on forecasting desalination system performance through design parameters

System type	Data-driven method	Data set size	Inputs	Outputs	Main remarks	Ref.
HDH (solar seawater greenhouse)	ANN	66	Roof transparency	Water production rate	The structure of the ANN is specified as 4–9–1	91
					One often employed approach for hyperparameter tuning is the trial-and-error method
					The dataset was divided into 3 subsets for training, validation, and testing purposes. The training set accounted for 70% of the data, while the validation set comprised 15% and the test set consisted of 5%
					The LM algorithm emerged as the most effective training method
			Length, width and height of front evaporator
RO	RBFANN	304	Model parameters (fractional pore area, potential parameter, average pore length)	Pure solvent flux, separation factor, total flux	The architecture of the artificial neural network (ANN) is structured as a 9–20–1 configuration	90
					The data was divided into two sets, with 80% allocated for training purposes and 20% reserved for testing
			Membrane properties (friction constant between solvent, solute and membrane, pore radius)
					The performance of RBFANN was superior to that of the previous methods (mathematical and mechanistic-based models)
			Operational parameters (temperature, pressure)
HDH (seawater greenhouse system)	SVR	66	Roof transparency, first evaporator height, greenhouse length and width	Water production	The dataset was divided into two subsets for the purpose of training and testing the model. The training set consisted of 70% of the data, while the remaining 30% was allocated for testing	92
					The constructed model was utilized to examine the impact of each input parameter on consumption of energy and water production
HDH (seawater greenhouse system)	ML PANN	30	Roof transparency	Power consumption	The dataset was divided into two subsets: the training set, which comprised 70% of the data, and the test set, which comprised the remaining 30%	93
			Height, width and length of first evaporator
					The RVFL-AEO model exhibited superior performance in comparison to the RVFL model, suggesting that the inclusion of AEO plays a significant role in achieving optimal RVFL parameters that ultimately improve the accuracy of the model
MD (VMD)	ANN	36	Membrane length, feed flow rate, feed inlet temperature	Specific heat energy consumption, permeate flux, heat consumption	The design of the artificial neural network (ANN) under consideration follows a certain configuration, denoted as 3–7–1	94
					The data was divided into three sets with the following split ratios: 70% for training, 10% for validation, and 20% for testing
					The increase in feed flow rate and feed inlet temperature resulted in an increase in permeate flux. Moreover, as the length of the membrane increased, there was a decrease in the permeate flux

---

Various ML models, such as artificial neural networks (ANN), support vector regression (SVR), RBFANN, and ML PANN, were studied for different membrane processes, including RO, HDH (solar seawater greenhouse), and membrane distillation (MD). Different output parameters were analyzed, and these models were used to optimize the design parameters of the respective processes.

Iranmanesh et al.⁹⁰ conducted a comparative analysis of model parameters and operational variables, highlighting the critical importance of precise hyperparameter tuning techniques in machine learning approaches for evaluating the performance of humidification–dehumidification (HDH) systems. Their results highlighted that combining the random vector functional link (RVFL) model with an artificial ecosystem-based optimization algorithm significantly improved model accuracy. Furthermore, another research effort⁸⁸ investigated the utilization of the SVR model in evaluating the efficiency of a seawater greenhouse system. This study unveiled the impressive predictive capacity of the SVR model in forecasting both freshwater production rates and energy consumption.

In the case of multi-effect distillation (MD) systems, one study focused on considering the membrane length in a vacuum membrane distillation (VMD) configuration, along with operational parameters like temperature and feed mass flow rate, as inputs for modeling an ANN model. This investigation unveiled the substantial influence of membrane length on both energy consumption and permeate flux, emphasizing the critical significance of including membrane length as an input parameter when employing data-driven (DD) methods for model development.

6.4. Process control

Table 10 summarizes research on the utilization of data-driven (DD) methods to enhance the effectiveness of different desalination methods. It is noteworthy that ANN models have garnered significant attention from researchers due to their effectiveness in predicting and controlling non-linear systems. Additionally, the results presented in ref. 95 demonstrate the successful application of LSTM deep learning models for the real-time control of RO desalination systems.

Table 10 Summary of research findings on the implication of data-driven approaches for control applications

System	Data-driven method	Data set size	Input	Output	Main remarks	Ref.
RO	ANN	1197	Temperature, pressure, conductivity	Flow, pressure	The ANN architecture employed for flow prediction is structured as follows: 3 input nodes, 38 hidden nodes in the 1st layer, 4 hidden nodes in the 2nd layer, and 1 output node. Additionally, another ANN architecture uses 3 input nodes, 69 hidden nodes in the 1st layer, 13 in the 2nd layer, and 1 in the output node	96
					The ANN architecture employed for pressure prediction is specified as 3 : 56 : 9 : 1, and 3 : 71 : 17 : 1
					ANNs were employed to control the variable functioning of an RO plant effectively
					ANNs effectively managed the unpredictable and diverse range of electrical power resources
					ANN models created set points for operating pressure and feed flow
RO	RNN (LSTM)	1871	Feed pressure	Permeate concentration, permeate flow rate	The LSTM model was employed as a robust predictive controller model	99
RO	DRL	8760	Electricity price, water demand, PV output	Pollution cost, cost of battery storage system, operating cost	The data was divided into two sets, with 90% allocated for training and 10% reserved for testing	97
					The study focuses on the optimal control objective of managing energy in a hybrid energy system, considering several targets and constraints
					The information entropy theory is employed to determine the weight factor that facilitates the assessment of the trade-off among several aims. According to the paper, a proficiently trained agent can offer an enhanced control policy, resulting in a potential cost reduction of up to 14.17% compared to alternative approaches
MD (solar PGMD)	ANN	540	Feed flow rate	Distillate flow rate	The artificial neural network (ANN) architecture is structured as a 1–5–1 configuration	98
					One often employed approach for hyper-parameter tuning involves trial and error
					The dataset was divided into 3 subsets for training and evaluation. The training set comprised 80% of the data, while the validation set accounted for 15%, and the remaining 5% was allocated to the test set
					A dynamic artificial neural network (ANN) control system was designed to optimize the distillate flow rate
					Implementing the adaptive control system resulted in a 17.2% increase in daily productivity
MSF	ANN	4500	Set point, blow down flow rate, mass flow rate of the heater vapor	Brine salinity, level of last stage, top brine temperature	This study's artificial neural network (ANN) architecture follows a 3–12–1 configuration. The hyper-parameter tuning approach employed was trial and error. 75% of the data is allocated to the training set, and the remaining 25% is designated for the test set. Comprising 25% of the data	100
					Three controllers utilizing an ANN model were evaluated to regulate the top brine temperature, salinity, and the level of the last stage
					The findings indicate that implementing a control mechanism in MSF desalination systems is feasible
RO	GA AND ANN	474	Flow rate, conductivity, time	Permeate conductivity, permeate flow	The data was partitioned into three distinct sets: the training set, comprising 60% of the data; the validation set, consisting of 20% of the data; and the test set, which likewise constituted 20% of the data	101
					The investigation focused on examining long-term forecasting and control strategies for the RO system for 5000 hours of operation
					By employing control measures, conductivity permeability can be declined in both the model and experimental prediction phases

---

Various ML models, such as ANN, RNN (LSTM), and DRL, were studied for different membrane processes, including RO, MD (solar PGMD) and MSF. Different output parameters were analyzed, and these models were used to optimize the process control of the respective processes.

However, it is evident from Table 10 that there is a notable gap in the literature regarding the use of DD control methods to enhance performance and reduce costs in renewable-based desalination systems, although solar and wind energy sources exhibit intermittent characteristics. For instance, Cabrera et al.⁹⁶ showcased the remarkable capabilities of ANN models in effectively regulating a wind-powered RO desalination facility with intermittent operations. In another study,⁹⁵ the implementation of a reinforcement learning model resulted in a substantial 14% cost reduction in a RO desalination system powered by renewable energy. Similarly, Gandhi et al.⁹⁷ reported successful outcomes using the sequential extreme learning method for improving performance and reducing costs in a solar still (SS) desalination system. Furthermore, Porrazzo et al.⁹⁸ achieved a notable approximately 17% increase in the productivity of daily freshwater of the pressure retarded osmosis (PRO) desalination system by employing adaptive control of the feed mass flow rate with an ANN model.

6.5. Maintenance

Table 11 comprehensively summarizes research investigations employing data-driven methods to explore wetting and fouling occurrences in membrane-based desalination systems. Fouling represents a significant challenge across various membrane desalination technologies, and the precise anticipation of fouling is pivotal in elevating system efficiency, curtailing cost, and ensuring sustainability.

Table 11 Summary of research findings on using data-driven techniques for maintenance

System type	Data driven model	Dataset size	Input	Output	Main remarks	Ref.
RO	CNN	13 708	Image	Flux decline, fouling growth	The present study explores various methodologies employed in conjunction with the convolutional neural network (CNN) model to detect and analyze fouling phenomena	102
					Regarding performance, CNN exhibited superior results compared to mathematical approaches
RO	ANN	6 years process data base	Water quality parameters (total ammonia and chlorine, turbidity), hydraulic parameters (pressure and flow rate)	New calculated pressure index	A total of 59 hydraulic and water quality measures were employed to assess and quantify the underlying factors contributing to membrane fouling within an RO system	103
					For the initial validation of the generated model, a substantial dataset was utilized
					The identification of the most accurate indicators of fouling was achieved by the utilization of ANN. The study demonstrated the potential application of the model in mitigating fouling rates
MD (VMD)	ANN	149	Feed side temperature, time, feed flow rate, pressure	Permeate flux	The architecture of the ANN is specified as a 5–10–1 configuration	104
					75% of the data used for training; 15% for the validation and 15% for testing
					The study focused on the impact of membrane fouling operational parameters
					The ANN model that was constructed was integrated with the GA optimization method to identify the optimal values for the operational parameters
					Membrane fouling was seen to occur more prominently under elevated feed temperatures and reduced vacuum pressures. Moreover, elevated feed solute concentrations were found to be associated with increased membrane fouling
MD	ANN	229	Raw water turbidity, permeate flow rate, operating time	Transmembrane pressure drops	The architecture of the ANN is specified as 3–5–1	105
					One often used approach for hyper-parameter tuning is the trail-error method
					A comparative analysis was conducted to evaluate the predictive capabilities of mathematical models and artificial neural network (ANN) models in the context of membrane fouling investigation
					The ANN demonstrated significant superiority in predicting the total market performance (TMP) compared to the use of blocking rules across all trial periods
ED	ANN	22	Salt concentration, current, crossflow velocity	Stack resistance	According to the report, this model has the capability to accurately forecast the fouling rate, even when provided with a restricted amount of experimental data	106
					The study demonstrated that neural differential equations exhibit strong extrapolation capabilities when applied to simulate colloidal fouling in electrodialysis
					The Sobol sensitivity analysis reveals that the crossflow velocity has a direct, linear effect of 41%, whereas the current and salt concentration have effects of 18.6% and 13.1% respectively

---

Various ML models, such as ANN and CNN, were studied for different membrane processes, including RO, MD (VDM), MD and ED. Different output parameters were analyzed, and these models were used to optimize the maintenance of the respective processes.

Fouling mechanisms are generally characterized by the undesirable accumulation of materials such as solid particles in the biological substances, ions, and feed stream both on the surface of the membrane and within its pores. This gradual accumulation leads to a decline in permeate flux over time, resulting in higher expenses for producing purified water.

Traditional mechanistic modeling methods have encountered challenges in accurately forecasting fouling mechanisms in various membrane-based desalination technologies. These challenges primarily stem from the dynamic nature of fouling phenomena, the intricacies inherent in mathematical modeling, and the necessity of making simplified assumptions during model development. As a result, researchers have increasingly shown interest in DD approaches as a viable alternative to providing more precise predictions of fouling mechanisms.

7. Conclusion

This study provides a detailed analysis of the utilization of data-driven techniques in membrane desalination systems. The studies that have undergone examination have been extensively classified according to the data-driven method type, the desalination system type, and the utilization of data-driven methods for analyzing different desalination systems. The study revealed that data-driven approaches have predominantly been utilized in the analysis of desalination systems across four distinct applications: performance prediction based on design parameters, performance prediction incorporating operating parameters, control and maintenance.

Based on the analysis, it can be concluded that within the field of ML techniques, ANN and RSM emerged as the most often utilized methods. Furthermore, the findings presented in several studies demonstrate the effectiveness of alternative ML techniques, including support vector machines (SVM), adaptive neuro-fuzzy inference systems (ANFIS), and random forest (RF), in accurately analyzing desalination systems. These results demonstrate the need for further research in exploring additional traditional ML methods for analyzing various desalination technologies.

Despite the present advancements in this domain, additional comprehensive research and investigations are required to enhance desalination efficacy and propel the commercialization of this technology. The choice of the desalination method should rely on the characteristics and quality of the produced water. Furthermore, it is imperative that the studies prioritize the examination of the financial aspect and the potential for maintaining the effectiveness of the treatment to ensure its viability for industrial implementation.

Conflicts of interest

There are no conflicts to declare.

Data availability

Data will be made available on reasonable request.

Acknowledgements

The authors would like to acknowledge the Lahore University of Management Sciences (LUMS) for the Faculty Startup Grant (STG-180).

References

1. M. A. Abdelkareem, M. El Haj Assad, E. T. Sayed and B. Soudan, Recent progress in the use of renewable energy sources to power water desalination plants, Desalination, 2018, 435, 97–113,  DOI:10.1016/j.desal.2017.11.018.
2. A. Alkaisi, R. Mossad and A. Sharifian-Barforoush, A Review of the Water Desalination Systems Integrated with Renewable Energy, Energy Procedia, 2017, 110, 268–274,  DOI:10.1016/j.egypro.2017.03.138.
3. W. L. Ang, A. W. Mohammad, N. Hilal and C. P. Leo, A review on the applicability of integrated/hybrid membrane processes in water treatment and desalination plants, Desalination, 2015, 363, 2–18,  DOI:10.1016/j.desal.2014.03.008.
4. L. F. Greenlee, D. F. Lawler, B. D. Freeman, B. Marrot and P. Moulin, Reverse osmosis desalination: water sources, technology, and today's challenges, Water Res., 2009, 43, 2317–2348,  DOI:10.1016/j.watres.2009.03.010.
5. A. Subramani, M. Badruzzaman, J. Oppenheimer and J. G. Jacangelo, Energy minimization strategies and renewable energy utilization for desalination: a review, Water Res., 2011, 45, 1907–1920,  DOI:10.1016/j.watres.2010.12.032.
6. G. Ignacz, A. K. Beke, V. Toth and G. Szekely, A hybrid modelling approach to compare chemical separation technologies in terms of energy consumption and carbon dioxide emissions, Nat. Energy, 2025, 10, 308–317,  DOI:10.1038/s41560-024-01668-7.
7. A. Subramani, N. Voutchkov and J. G. Jacangelo, Desalination energy minimization using thin film nanocomposite membranes, Desalination, 2014, 350, 35–43,  DOI:10.1016/j.desal.2014.07.011.
8. R. Semiat, Energy Issues in Desalination Processes, Environ. Sci. Technol., 2008, 42, 8193–8201,  DOI:10.1021/es801330u.
9. L. Gao, A. Rahardianto, H. Gu, P. D. Christofides and Y. Cohen, Energy-Optimal Control of RO Desalination, Ind. Eng. Chem. Res., 2014, 53, 7409–7420,  DOI:10.1021/ie402304d.
10. A. R. Bartman, A. Zhu, P. D. Christofides and Y. Cohen, Minimizing energy consumption in reverse osmosis membrane desalination using optimization-based control, Proc. Am. Control Conf., 2010, 3629–3635,  DOI:10.1109/ACC.2010.5530614.
11. B. Sun, L. Zhang, L. Yang, F. Zhang, D. Norse and Z. Zhu, Agricultural Non-Point Source Pollution in China: Causes and Mitigation Measures, Ambio, 2012, 41, 370–379,  DOI:10.1007/s13280-012-0249-6.
12. O. Lefebvre and R. Moletta, Treatment of organic pollution in industrial saline wastewater: a literature review, Water Res., 2006, 40, 3671–3682,  DOI:10.1016/j.watres.2006.08.027.
13. S. Alzahrani, A. W. Mohammad, N. Hilal, P. Abdullah and O. Jaafar, Comparative study of NF and RO membranes in the treatment of produced water-Part I: Assessing water quality, Desalination, 2013, 315, 18–26,  DOI:10.1016/j.desal.2012.12.004.
14. A. J. Kondash, E. Albright and A. Vengosh, Quantity of flowback and produced waters from unconventional oil and gas exploration, Sci. Total Environ., 2017, 574, 314–321,  DOI:10.1016/j.scitotenv.2016.09.069.
15. H. Chang, T. Li, B. Liu, R. D. Vidic, M. Elimelech and J. C. Crittenden, Potential and implemented membrane-based technologies for the treatment and reuse of flowback and produced water from shale gas and oil plays: A review, Desalination, 2019, 455, 34–57,  DOI:10.1016/j.desal.2019.01.001.
16. S. Munirasu, M. A. Haija and F. Banat, Use of membrane technology for oil field and refinery produced water treatment—A review, Process Saf. Environ. Prot., 2016, 100, 183–202,  DOI:10.1016/j.psep.2016.01.010.
17. R. Keyikoglu, O. Karatas, H. Rezania, M. Kobya, V. Vatanpour and A. Khataee, A review on treatment of membrane concentrates generated from landfill leachate treatment processes, Sep. Purif. Technol., 2021, 259, 118182,  DOI:10.1016/j.seppur.2020.118182.
18. W. Chen, A. Zhang, G. Jiang and Q. Li, Transformation and degradation mechanism of landfill leachates in a combined process of SAARB and ozonation, Waste Manage., 2019, 85, 283–294,  DOI:10.1016/j.wasman.2018.12.038.
19. G. Lofrano, S. Meriç, G. E. Zengin and D. Orhon, Chemical and biological treatment technologies for leather tannery chemicals and wastewaters: a review, Sci. Total Environ., 2013, 461–462, 265–281,  DOI:10.1016/j.scitotenv.2013.05.004.
20. R. Kishor, D. Purchase, G. D. Saratale, R. G. Saratale, L. F. R. Ferreira, M. Bilal, R. Chandra and R. N. Bharagava, Ecotoxicological and health concerns of persistent coloring pollutants of textile industry wastewater and treatment approaches for environmental safety, J. Environ. Chem. Eng., 2021, 9, 105012,  DOI:10.1016/j.jece.2020.105012.
21. J. Dasgupta, J. Sikder, S. Chakraborty, S. Curcio and E. Drioli, Remediation of textile effluents by membrane based treatment techniques: A state of the art review, J. Environ. Manage., 2015, 147, 55–72,  DOI:10.1016/j.jenvman.2014.08.008.
22. G. Q. Chen, S. Talebi, S. L. Gras, M. Weeks and S. E. Kentish, A review of salty waste stream management in the Australian dairy industry, J. Environ. Manage., 2018, 224, 406–413,  DOI:10.1016/j.jenvman.2018.07.056.
23. P. Dhanke, S. Wagh and A. Patil, Treatment of fish processing industry wastewater using hydrodynamic cavitational reactor with biodegradability improvement, Water Sci Technol. J. Int. Assoc. Water Pollut. Res., 2019, 80, 2310–2319,  DOI:10.2166/wst.2020.049.
24. D. Karadag, O. E. Köroğlu, B. Ozkaya and M. Cakmakci, A review on anaerobic biofilm reactors for the treatment of dairy industry wastewater, Process Biochem., 2015, 50, 262–271,  DOI:10.1016/j.procbio.2014.11.005.
25. R. Keyikoglu, O. Karatas, H. Rezania, M. Kobya, V. Vatanpour and A. Khataee, A review on treatment of membrane concentrates generated from landfill leachate treatment processes, Sep. Purif. Technol., 2021, 259, 118182,  DOI:10.1016/j.seppur.2020.118182.
26. S. Wu, S. Zou, G. Liang, G. Qian and Z. He, Enhancing recovery of magnesium as struvite from landfill leachate by pretreatment of calcium with simultaneous reduction of liquid volume via forward osmosis, Sci. Total Environ., 2018, 610–611, 137–146,  DOI:10.1016/j.scitotenv.2017.08.038.
27. R. Toczyłowska-Mamińska, Limits and perspectives of pulp and paper industry wastewater treatment – A review, Renewable Sustainable Energy Rev., 2017, 78, 764–772,  DOI:10.1016/j.rser.2017.05.021.
28. N. A. Ahmad, P. S. Goh, L. T. Yogarathinam, A. K. Zulhairun and A. F. Ismail, Current advances in membrane technologies for produced water desalination, Desalination, 2020, 493, 114643,  DOI:10.1016/j.desal.2020.114643.
29. S. Adham, A. Hussain, J. Minier-Matar, A. Janson and R. Sharma, Membrane applications and opportunities for water management in the oil & gas industry, Desalination, 2018, 440, 2–17,  DOI:10.1016/j.desal.2018.01.030.
30. R. Shahouni, M. Abbasi, M. Dibaj and M. Akrami, Utilising Artificial Intelligence to Predict Membrane Behaviour in Water Purification and Desalination, Water, 2024, 16, 2940,  DOI:10.3390/w16202940.
31. N. N. R. Ahmad, W. L. Ang, C. P. Leo, A. W. Mohammad and N. Hilal, Current advances in membrane technologies for saline wastewater treatment: A comprehensive review, Desalination, 2021, 517, 115170,  DOI:10.1016/j.desal.2021.115170.
32. K. C. Ho, Y. H. Teow, A. W. Mohammad, W. L. Ang and P. H. Lee, Development of graphene oxide (GO)/multi-walled carbon nanotubes (MWCNTs) nanocomposite conductive membranes for electrically enhanced fouling mitigation, J. Membr. Sci., 2018, 552, 189–201,  DOI:10.1016/j.memsci.2018.02.001.
33. L. Meng, M. Wu, H. Chen, Y. Xi, M. Huang and X. Luo, Rejection of antimony in dyeing and printing wastewater by forward osmosis, Sci. Total Environ., 2020, 745, 141015,  DOI:10.1016/j.scitotenv.2020.141015.
34. S. Wu, S. Zou, G. Liang, G. Qian and Z. He, Enhancing recovery of magnesium as struvite from landfill leachate by pretreatment of calcium with simultaneous reduction of liquid volume via forward osmosis, Sci. Total Environ., 2018, 610–611, 137–146,  DOI:10.1016/j.scitotenv.2017.08.038.
35. C. Wu and Q. Li, Characteristics of organic matter removed from highly saline mature landfill leachate by an emergency disk tube-reverse osmosis treatment system, Chemosphere, 2021, 263, 128347,  DOI:10.1016/j.chemosphere.2020.128347.
36. E. Sahinkaya, S. Tuncman, I. Koc, A. R. Guner, S. Ciftci, A. Aygun and S. Sengul, Performance of a pilot-scale reverse osmosis process for water recovery from biologically-treated textile wastewater, J. Environ. Manage., 2019, 249, 109382,  DOI:10.1016/j.jenvman.2019.109382.
37. N. M. Mokhtar, W. J. Lau, A. F. Ismail, S. Kartohardjono, S. O. Lai and H. C. Teoh, The potential of direct contact membrane distillation for industrial textile wastewater treatment using PVDF-Cloisite 15A nanocomposite membrane, Chem. Eng. Res. Des., 2016, 111, 284–293,  DOI:10.1016/j.cherd.2016.05.018.
38. A. El-Abbassi, M. Khayet, H. Kiai, A. Hafidi and M. C. García-Payo, Treatment of crude olive mill wastewaters by osmotic distillation and osmotic membrane distillation, Sep. Purif. Technol., 2013, 104, 327–332,  DOI:10.1016/j.seppur.2012.12.006.
39. A. Albahnasawi, E. Yüksel, E. Gürbulak and F. Duyum, Fate of aromatic amines through decolorization of real textile wastewater under anoxic-aerobic membrane bioreactor, J. Environ. Chem. Eng., 2020, 8, 104226,  DOI:10.1016/j.jece.2020.104226.
40. R. Umaiyakunjaram and P. Shanmugam, Study on submerged anaerobic membrane bioreactor (SAMBR) treating high suspended solids raw tannery wastewater for biogas production, Bioresour. Technol., 2016, 216, 785–792,  DOI:10.1016/j.biortech.2016.06.017.
41. J. Mulopo, Bleach plant effluent treatment in anaerobic membrane bioreactor (AMBR) using carbon nanotube/polysulfone nanocomposite membranes, J. Environ. Chem. Eng., 2017, 5, 4381–4387,  DOI:10.1016/j.jece.2017.08.027.
42. V. Piemonte, M. Prisciandaro, L. Mascis, L. Di Paola and D. Barba, Reverse osmosis membranes for treatment of produced water: a process analysis, Desalin. Water Treat., 2015, 55, 565–574,  DOI:10.1080/19443994.2015.1014862.
43. S. Mondal and S. R. Wickramasinghe, Produced water treatment by nanofiltration and reverse osmosis membranes, J. Membr. Sci., 2008, 322, 162–170,  DOI:10.1016/j.memsci.2008.05.039.
44. P. Xu and J. E. Drewes, Viability of nanofiltration and ultra-low pressure reverse osmosis membranes for multi-beneficial use of methane produced water, Sep. Purif. Technol., 2006, 52, 67–76,  DOI:10.1016/j.seppur.2006.03.019.
45. L. Ji, Y. Liu, Y. F. Zhang and X. D. Liu, Study on the Treatment of High Salinity Oil Waste Water by Reverse Osmosis Technology, Adv. Mater. Res., 2011, 418–420, 90–93,  DOI:10.4028/www.scientific.net/AMR.418-420.90.
46. B. Pei, J. Chen, P. Liu, T. He, X. Li and L. Zhang, Hyperbranched poly(amidoamine)/TMC reverse osmosis membrane for oily saline water treatment, Environ. Technol., 2019, 40, 2779–2788,  DOI:10.1080/09593330.2018.1452982.
47. P. S. Goh, A. F. Ismail, B. C. Ng and M. S. Abdullah, Recent Progresses of Forward Osmosis Membranes Formulation and Design for Wastewater Treatment, Water, 2019, 11, 2043,  DOI:10.3390/w11102043.
48. R. L. McGinnis, N. T. Hancock, M. S. Nowosielski-Slepowron and G. D. McGurgan, Pilot demonstration of the NH3/CO2 forward osmosis desalination process on high salinity brines, Desalination, 2013, 312, 67–74,  DOI:10.1016/j.desal.2012.11.032.
49. J. Kim, J. Kim, J. Kim and S. Hong, Osmotically enhanced dewatering-reverse osmosis (OED-RO) hybrid system: Implications for shale gas produced water treatment, J. Membr. Sci., 2018, 554, 282–290,  DOI:10.1016/j.memsci.2018.03.015.
50. W. J. Lee, P. S. Goh, W. J. Lau, C. S. Ong and A. F. Ismail, Antifouling zwitterion embedded forward osmosis thin film composite membrane for highly concentrated oily wastewater treatment, Sep. Purif. Technol., 2019, 214, 40–50,  DOI:10.1016/j.seppur.2018.07.009.
51. Y. Chun, S.-J. Kim, G. J. Millar, D. Mulcahy, I. S. Kim and L. Zou, Forward osmosis as a pre-treatment for treating coal seam gas associated water: Flux and fouling behaviour, Desalination, 2017, 403, 144–152,  DOI:10.1016/j.desal.2015.09.012.
52. P. Xu, T. Cath and J. E. Drewes, Novel and Emerging Technologies for Produced Water Treatment, in Hydraulic Fracturing Technical Workshop, 2011, pp. 1–6.
53. B. D. Coday and T. Y. Cath, Forward osmosis: Novel desalination of produced water and fracturing flowback, J. - Am. Water Works Assoc., 2014, 106, E55–E66,  DOI:10.5942/jawwa.2014.106.0016.
54. D. L. Shaffer, J. R. Werber, H. Jaramillo, S. Lin and M. Elimelech, Forward osmosis: Where are we now?, Desalination, 2015, 356, 271–284,  DOI:10.1016/j.desal.2014.10.031.
55. Q. Long, Y. Jia, J. Li, J. Yang, F. Liu, J. Zheng and B. Yu, Recent Advance on Draw Solutes Development in Forward Osmosis, Processes, 2018, 6, 165,  DOI:10.3390/pr6090165.
56. E. A. Bell, T. E. Poynor, K. B. Newhart, J. Regnery, B. D. Coday and T. Y. Cath, Produced water treatment using forward osmosis membranes: Evaluation of extended-time performance and fouling, J. Membr. Sci., 2017, 525, 77–88,  DOI:10.1016/j.memsci.2016.10.032.
57. B. D. Coday, N. Almaraz and T. Y. Cath, Forward osmosis desalination of oil and gas wastewater: Impacts of membrane selection and operating conditions on process performance, J. Membr. Sci., 2015, 488, 40–55,  DOI:10.1016/j.memsci.2015.03.059.
58. D. Qin, Z. Liu, D. Delai Sun, X. Song and H. Bai, A new nanocomposite forward osmosis membrane custom-designed for treating shale gas wastewater, Sci. Rep., 2015, 5, 14530,  DOI:10.1038/srep14530.
59. T. Liden, D. D. Carlton, S. Miyazaki, T. Otoyo and K. A. Schug, Forward osmosis remediation of high salinity Permian Basin produced water from unconventional oil and gas development, Sci. Total Environ., 2019, 653, 82–90,  DOI:10.1016/j.scitotenv.2018.10.325.
60. I. H. Sarker, A. S. M. Kayes, S. Badsha, H. Alqahtani, P. Watters and A. Ng, Cybersecurity data science: an overview from machine learning perspective, J. Big Data, 2020, 7, 1–29,  DOI:10.1186/s40537-020-00318-5.
61. M. Mohammed, M. Khan and E. Bashier, Machine Learning: Algorithms and Applications, 2016,  DOI:10.1201/9781315371658.
62. T. Bonny, M. Kashkash and F. Ahmed, An efficient deep reinforcement machine learning-based control reverse osmosis system for water desalination, Desalination, 2022, 522, 115443,  DOI:10.1016/j.desal.2021.115443.
63. G. Alam, I. Ihsanullah, M. Naushad and M. Sillanpää, Applications of artificial intelligence in water treatment for optimization and automation of adsorption processes: Recent advances and prospects, Chem. Eng. J., 2022, 427, 130011,  DOI:10.1016/j.cej.2021.130011.
64. M. J. Talukder, A. S. Alshami, A. Tayyebi, N. Ismail and X. Yu, Membrane Science Meets Machine Learning: Future and Potential Use in Assisting Membrane Material Design and Fabrication, Sep. Purif. Rev., 2024, 53, 216–229,  DOI:10.1080/15422119.2023.2212295.
65. M. Hamachi, M. Cabassud, A. Davin and M. Mietton Peuchot, Dynamic modelling of crossflow microfiltration of bentonite suspension using recurrent neural networks, Chem. Eng. Process., 1999, 38, 203–210,  DOI:10.1016/S0255-2701(99)00004-5.
66. Z. Zhao, Y. Lou, Y. Chen, H. Lin, R. Li and G. Yu, Prediction of interfacial interactions related with membrane fouling in a membrane bioreactor based on radial basis function artificial neural network (ANN), Bioresour. Technol., 2019, 282, 262–268,  DOI:10.1016/j.biortech.2019.03.044.
67. G. B. Sahoo and C. Ray, Predicting flux decline in crossflow membranes using artificial neural networks and genetic algorithms, J. Membr. Sci., 2006, 283, 147–157,  DOI:10.1016/j.memsci.2006.06.019.
68. A. Abbas and N. Al-Bastaki, Modeling of an RO water desalination unit using neural networks, Chem. Eng. J., 2005, 114, 139–143,  DOI:10.1016/j.cej.2005.07.016.
69. Y. G. Lee, Y. S. Lee, J. J. Jeon, S. Lee, D. R. Yang, I. S. Kim and J. H. Kim, Artificial neural network model for optimizing operation of a seawater reverse osmosis desalination plant, Desalination, 2009, 247, 180–189,  DOI:10.1016/j.desal.2008.12.023.
70. R. Righton, Development of an artificial neural network model for predicting the performance of a reverse osmosis (RO) unit, MEng (Master of Engineering) thesis, Curtin University, 2009, https://espace.curtin.edu.au/handle/20.500.11937/1684.
71. D. Libotean, J. Giralt, F. Giralt, R. Rallo, T. Wolfe and Y. Cohen, Neural network approach for modeling the performance of reverse osmosis membrane desalting, J. Membr. Sci., 2009, 326, 408–419,  DOI:10.1016/j.memsci.2008.10.028.
72. M. Barello, D. Manca, R. Patel and I. M. Mujtaba, Neural network based correlation for estimating water permeability constant in RO desalination process under fouling, Desalination, 2014, 345, 101–111,  DOI:10.1016/j.desal.2014.04.016.
73. J. Farahbakhsh, M. Delnavaz and V. Vatanpour, Simulation and characterization of novel reverse osmosis membrane prepared by blending polypyrrole coated multiwalled carbon nanotubes for brackish water desalination and antifouling properties using artificial neural networks, J. Membr. Sci., 2019, 581, 123–138,  DOI:10.1016/j.memsci.2019.03.050.
74. K. A. M. Al-Shayji, Modeling, Simulation, and Optimization of large-Scale Commercial Desalination Plants, PhD thesis, 1998, https://vtechworks.lib.vt.edu/handle/10919/30462.
75. A. M. Aish, H. A. Zaqoot and S. M. Abdeljawad, Artificial neural network approach for predicting reverse osmosis desalination plants performance in the Gaza Strip, Desalination, 2015, 367, 240–247,  DOI:10.1016/j.desal.2015.04.008.
76. J. Jawad, A. H. Hawari and S. J. Zaidi, Modeling and Sensitivity Analysis of the Forward Osmosis Process to Predict Membrane Flux Using a Novel Combination of Neural Network and Response Surface Methodology Techniques, Membranes, 2021, 11, 70,  DOI:10.3390/membranes11010070.
77. G. Ignacz, L. Bader, A. K. Beke, Y. Ghunaim, T. Shastry, H. Vovusha, M. R. Carbone, B. Ghanem and G. Szekely, Machine learning for the advancement of membrane science and technology: A critical review, J. Membr. Sci., 2025, 713, 123256,  DOI:10.1016/j.memsci.2024.123256.
78. W. Cao, Q. Liu, Y. Wang and I. M. Mujtaba, Modeling and simulation of VMD desalination process by ANN, Comput. Chem. Eng., 2016, 84, 96–103,  DOI:10.1016/j.compchemeng.2015.08.019.
79. R. Chauhan, P. Dumka and D. R. Mishra, Modelling conventional and solar earth still by using the LM algorithm-based artificial neural network, Int. J. Ambient Energy, 2022, 43, 1389–1396,  DOI:10.1080/01430750.2019.1707113.
80. M. Faegh, P. Behnam, M. B. Shafii and M. Khiadani, Development of artificial neural networks for performance prediction of a heat pump assisted humidification-dehumidification desalination system, Desalination, 2021, 508, 115052,  DOI:10.1016/j.desal.2021.115052.
81. X. Pascual, H. Gu, A. R. Bartman, A. Zhu, A. Rahardianto, J. Giralt, R. Rallo, P. D. Christofides and Y. Cohen, Data-driven models of steady state and transient operations of spiral-wound RO plant, Desalination, 2013, 316, 154–161,  DOI:10.1016/j.desal.2013.02.006.
82. E. S. Salami, M. Ehetshami, A. Karimi-Jashni, M. Salari, S. N. Sheibani and A. Ehteshami, A mathematical method and artificial neural network modeling to simulate osmosis membrane's performance, Model. Earth Syst. Environ., 2016, 2, 1–11,  DOI:10.1007/s40808-016-0261-0.
83. A. Aminian, Prediction of temperature elevation for seawater in multi-stage flash desalination plants using radial basis function neural network, Chem. Eng. J., 2010, 162, 552–556,  DOI:10.1016/j.cej.2010.05.060.
84. A. Salgado-Reyna, E. Soto-Regalado, R. Gómez-González, F. J. Cerino-Córdova, R. B. García-Reyes, M. T. Garza-González and M. M. Alcalá-Rodríguez, Artificial neural networks for modeling the reverse osmosis unit in a wastewater pilot treatment plant, Desalin. Water Treat., 2013, 1–11,  DOI:10.1080/19443994.2013.862023.
85. A. T. Mohammad, M. A. Al-Obaidi, E. M. Hameed, B. N. Basheer and I. M. Mujtaba, Modelling the chlorophenol removal from wastewater via reverse osmosis process using a multilayer artificial neural network with genetic algorithm, J. Water Process Eng., 2020, 33, 100993,  DOI:10.1016/j.jwpe.2019.100993.
86. M. I. Kizhisseri, M. M. Mohamed and M. A. Hamouda, Prediction of capital cost of ro based desalination plants using machine learning approach: 7th International Conference on Environment Pollution and Prevention, ICEPP 2019, E3S Web Conf. 158 (2020),  DOI:10.1051/e3sconf/202015806001.
87. M. Tavakolmoghadam and M. Safavi, An Optimized Neural Network Model of Desalination by Vacuum Membrane Distillation Using Genetic Algorithm, Procedia Eng., 2012, 42, 106–112,  DOI:10.1016/j.proeng.2012.07.400.
88. L. Khaouane, Y. Ammi and S. Hanini, Modeling the Retention of Organic Compounds by Nanofiltration and Reverse Osmosis Membranes Using Bootstrap Aggregated Neural Networks, Arabian J. Sci. Eng., 2017, 42, 1443–1453,  DOI:10.1007/s13369-016-2320-2.
89. P. Cabrera, J. A. Carta, J. González and G. Melián, Wind-driven SWRO desalination prototype with and without batteries: A performance simulation using machine learning models, Desalination, 2018, 435, 77–96,  DOI:10.1016/j.desal.2017.11.044.
90. F. Iranmanesh, A. Moradi and M. Rafizadeh, Implementation of radial basic function networks for the prediction of RO membrane performances by using a complex transport model, Desalin. Water Treat., 2016, 57, 20307–20317,  DOI:10.1080/19443994.2015.1111806.
91. T. Zarei and R. Behyad, Predicting the water production of a solar seawater greenhouse desalination unit using multi-layer perceptron model, Sol. Energy, 2019, 177, 595–603,  DOI:10.1016/j.solener.2018.11.059.
92. T. Zarei, R. Behyad and E. Abedini, Study on parameters effective on the performance of a humidification-dehumidification seawater greenhouse using support vector regression, Desalination, 2018, 435, 235–245,  DOI:10.1016/j.desal.2017.05.033.
93. F. A. Essa, M. Abd Elaziz and A. H. Elsheikh, Prediction of power consumption and water productivity of seawater greenhouse system using random vector functional link network integrated with artificial ecosystem-based optimization, Process Saf. Environ. Prot., 2020, 144, 322–329,  DOI:10.1016/j.psep.2020.07.044.
94. C. Yang, X. Peng, Y. Zhao, X. Wang, J. Fu, K. Liu, Y. Li and P. Li, Prediction model to analyze the performance of VMD desalination process, Comput. Chem. Eng., 2020, 132, 106619,  DOI:10.1016/j.compchemeng.2019.106619.
95. D. Karimanzira and T. Rauschenbach, Deep Learning Based Model Predictive Control for a Reverse Osmosis Desalination Plant, Z. Angew. Math. Phys., 2020, 08, 2713–2731,  DOI:10.4236/jamp.2020.812201.
96. P. Cabrera, J. A. Carta, J. González and G. Melián, Artificial neural networks applied to manage the variable operation of a simple seawater reverse osmosis plant, Desalination, 2017, 416, 140–156,  DOI:10.1016/j.desal.2017.04.032.
97. A. M. Gandhi, S. Shanmugan, S. Gorjian, C. I. Pruncu, S. Sivakumar, A. H. Elsheikh, F. A. Essa, Z. M. Omara and H. Panchal, Performance enhancement of stepped basin solar still based on OSELM with traversal tree for higher energy adaptive control, Desalination, 2021, 502, 114926,  DOI:10.1016/j.desal.2020.114926.
98. R. Porrazzo, A. Cipollina, M. Galluzzo and G. Micale, A neural network-based optimizing control system for a seawater-desalination solar-powered membrane distillation unit, Comput. Chem. Eng., 2013, 54, 79–96,  DOI:10.1016/j.compchemeng.2013.03.015.
99. G. Zhang, W. Hu, D. Cao, W. Liu, R. Huang, Q. Huang, Z. Chen and F. Blaabjerg, Data-driven optimal energy management for a wind-solar-diesel-battery-reverse osmosis hybrid energy system using a deep reinforcement learning approach, Energy Convers. Manage., 2021, 227, 113608,  DOI:10.1016/j.enconman.2020.113608.
100. S. S. Madaeni, M. Shiri and A. R. Kurdian, Modeling, optimization, and control of reverse osmosis water treatment in kazeroon power plant using neural network, Chem. Eng. Commun., 2015, 202, 6–14,  DOI:10.1080/00986445.2013.828606.
101. S. Tayyebi and M. Alishiri, The control of MSF desalination plants based on inverse model control by neural network, Desalination, 2014, 333, 92–100,  DOI:10.1016/j.desal.2013.11.022.
102. S. Park, S.-S. Baek, J. Pyo, Y. Pachepsky, J. Park and K. H. Cho, Deep neural networks for modeling fouling growth and flux decline during NF/RO membrane filtration, J. Membr. Sci., 2019, 587, 117164,  DOI:10.1016/j.memsci.2019.06.004.
103. E. A. Roehl, D. A. Ladner, R. C. Daamen, J. B. Cook, J. Safarik, D. W. Phipps and P. Xie, Modeling fouling in a large RO system with artificial neural networks, J. Membr. Sci., 2018, 552, 95–106,  DOI:10.1016/j.memsci.2018.01.064.
104. S. Mittal, A. Gupta, S. Srivastava and M. Jain, Artificial Neural Network based modeling of the vacuum membrane distillation process: Effects of operating parameters on membrane fouling, Chem. Eng. Process., 2021, 164, 108403,  DOI:10.1016/j.cep.2021.108403.
105. Q.-F. Liu and S.-H. Kim, Evaluation of membrane fouling models based on bench-scale experiments: A comparison between constant flowrate blocking laws and artificial neural network (ANNs) model, J. Membr. Sci., 2008, 310, 393–401,  DOI:10.1016/j.memsci.2007.11.020.
106. B. De Jaegher, E. Larumbe, W. De Schepper, A. Verliefde and I. Nopens, Colloidal fouling in electrodialysis: A neural differential equations model, Sep. Purif. Technol., 2020, 249, 116939,  DOI:10.1016/j.seppur.2020.116939.

---