Autopsy Results and Inorganic Fouling Prediction Modeling Using Artificial Neural Networks for Reverse Osmosis Membranes in a Desalination Plant

Abstract

Nowadays, reverse osmosis (RO) desalination has become a highly effective and economical solution to address water scarcity worldwide. The membranes used in this type of separation are influenced by both pre-treatment operations and feed water quality, leading to fouling, a complex phenomenon responsible for reducing treatment performance and shortening membrane lifespan. In this study, an autopsy of a RO membrane from the Boujdour plant was performed, and a fouling prediction tool based on transmembrane pressure (TMP) was developed using MATLAB/Simulink (R2015a) with an artificial neural network (ANN) model. The impact of membrane fouling on treatment performance was also examined through one year of monitoring. A detailed analysis of the fouled membrane was conducted using SEM/EDS techniques to characterize the fouling on the membrane’s surface and cross-section. The results revealed significant fractures on the membrane surface, with fouling predominantly consisting of organic deposits (characterized by a high oxygen concentration of 39.69%) and inorganic fouling, including Si (7.99%), Al (2.79%), Mg (1.56%), Fe (1.27%), and smaller quantities of K (0.87%), S (0.36%), and Ca (0.12%). The ANN model for predicting transmembrane pressure was successfully developed, achieving a high R² value of 92.077% and a low mean square error (MSE) of 0.005657. This predictive model demonstrates the ability to forecast future TMP cycles based on historical data. The research provides a detailed understanding of the types of fouling affecting RO membranes and contributes to the development of preventive strategies to mitigate membrane fouling.

1. Introduction

Freshwater scarcity worldwide is one of the most significant threats not only to communities but also to the environment and biodiversity. Exacerbated by climate change and population growth, this issue has worsened alarmingly in recent decades [1]. Although seawater, which accounts for 97.2% of the Earth’s water, is abundant, it is not consumable in its natural form. Desalination and wastewater reuse are emerging as promising solutions to address growing water demands. Desalination is the process of extracting freshwater from brackish or saline water, such as seawater. While seawater desalination is not an environmentally friendly technique, it is becoming increasingly widespread, with desalination plants proliferating across the globe [2]. Reverse osmosis is used to remove dissolved solids, including ions and small- to large-molecular-weight solutes, from water. Today, reverse osmosis has become the leading technology for water desalination [3]. However, membrane fouling remains the primary challenge in reverse osmosis systems, as it is the main factor behind reduced treatment performance and the shortened lifespan of RO membranes [4]. Substances responsible for membrane fouling can generally be classified as inorganic compounds, organic compounds, colloidal particles, and biological substances. Inorganic fouling results when the concentration of soluble salts exceeds their solubility limit. Organic fouling occurs naturally in the form of dissolved components or colloids (such as humic acids, fulvic acids, proteins, etc.) accumulating on the membrane surface. Colloidal fouling is caused by the accumulation of organic or inorganic matter, which forms a layer of cake. Biological clogging results from microorganisms adhering to the surface of the membrane, leading to the formation of a biofilm composed of microbial cells and extracellular polymeric substances [5].

Membrane fouling occurs due to various mechanisms, including concentration polarization, adsorption, pore blockage, and particle deposition on the membrane surface [6]. To mitigate fouling, membrane cleaning is performed and is recommended when the permeate flow decreases by around 10%, the transmembrane pressure increases by 15%, or the permeate conductivity rises by 10%. However, frequent cleaning can compromise membrane robustness, resulting in increased downtime, a higher risk of membrane degradation, and ultimately, a greater operational load on the desalination process [7]. The cleaning method depends on the nature of the fouling and can be either physical, chemical, or a combination of both. Physical cleaning is typically used for reversible fouling, while chemical cleaning is necessary for irreversible fouling [8]. Effectively managing fouling requires a thorough understanding of its underlying mechanisms. The membrane autopsy technique provides detailed insights into the fouling process, helping to enhance treatment quality and efficiency [9]. Karmal et al. [10] identified mineral fouling on reverse osmosis membranes, characterized by deposits of calcium carbonate (CaCO₃) in the form of calcite crystals. Similarly, Al-Abri et al. [11] reported the presence of both inorganic and organic particles on reverse osmosis membrane surfaces, with inorganic scalants such as Na, Cl, Mg, Al, and traces of Ca being detected. Fortunato et al. [12] demonstrated that the inorganic substances found on membrane surfaces primarily consist of iron, aluminum, and magnesium silicate. Bahar Ozbey et al. [13] investigated membranes used in industrial park wastewater treatment and observed fouling caused by CaCO₃ along with significant amounts of molybdenum (Mo). They proposed an appropriate chemical cleaning protocol to address this issue. Myoung Jun et al. [14] conducted an autopsy study of reverse osmosis membranes used in Sydney’s urban wastewater recycling systems for domestic and commercial wastewater. Their findings confirmed that protein compounds were the dominant foulants, while calcium, sodium, magnesium, and iron were the main inorganic elements present. To better understand and control reverse osmosis membrane fouling, modeling plays a crucial role [15]. It provides insights into the fouling mechanisms, enhances knowledge, and optimizes plant operations. In addition to approaches based on direct experimental data, some studies have shown that reverse osmosis membrane fouling can also be predicted using membrane performance normalization parameters. This method can monitor changes in performance over time and estimate when the membrane will reach a certain level of critical fouling, which is useful for programming cleaning or membrane replacement processes [16]. The integration of these approaches enhances the available diagnostic tools, complementing artificial intelligence-based methods such as neural networks, which offer accurate prediction from complex operational data. Recently, the application of ANNs in fouling modeling has gained significant attention due to their efficiency and ability to handle multiple parameters simultaneously [17]. Numerous studies focus on predicting membrane fouling using ANNs. Zhitao Zhao et al. [18] developed a radial basis function ANN to forecast interfacial interactions on a randomly rough membrane surface over a short separation distance range (0.158–10 nm), demonstrating the model’s high computational accuracy. Yun Teng et al. [19] explored the integration of mechanistic and data-driven models to predict reverse osmosis membrane fouling based on real operational data. Their model, which predicted TMP using an adsorption model combined with a data-driven approach, effectively described membrane fouling. Younjong Park et al. [20] conducted regression-based fouling predictions for antimicrobials in municipal wastewater treatment using multilayer perceptron (MLP). Their model, trained on 525 days of experimental data from a major pilot antimicrobial reactor, achieved a Training R² of 0.8622 and an RMSE of 3.897 kPa. Yasser Algoufily et al. [21] proposed a fouling monitoring and prediction model for membrane bioreactors using an ANN. Their model estimated total membrane resistance based on TMP and accounted for both deterministic and stochastic fouling mechanisms.

This study examines the results of membrane autopsy and fouling prediction through transmembrane pressure at the Boujdour reverse osmosis desalination plant, aiming to improve fouling control in the treatment process. In the first part, the impact of membrane fouling on treatment performance was evaluated over a one-year monitoring period. Three parameters, TMP, permeate conductivity, and permeate flow, were analyzed to assess the effect of fouling on separation quality. However, this monitoring alone was insufficient to identify the specific materials causing fouling. Consequently, a membrane autopsy was conducted. Analytical techniques, including scanning electron microscopy (SEM) and energy-dispersive spectroscopy (EDS), were employed to characterize the clogging layer on the membrane’s top surface after five years of operation. A cross-sectional analysis was also performed to understand the distribution patterns of the fouling layer across the membrane surface. In the second part, a TMP prediction model was developed using the ANN method. According to the literature, studies on RO membrane fouling prediction using machine learning methods have generally relied on output parameters as permeate flux, fouling layer thickness, silt density index (SDI), or total dissolved solids (TDS) concentration in the permeate. However, although transmembrane pressure (TMP) is considered a key fouling indicator, it has been studied in membrane separation techniques other than reverse osmosis [22]. Thus, this research aims to predict membrane fouling in the Boujdour RO plant using TMP as an output parameter. The input parameters considered include temperature, turbidity, pH, feed conductivity, and feed flow. The study also assesses the ability of artificial intelligence to anticipate membrane fouling. The prediction framework was implemented in MATLAB/Simulink (R2015a), utilizing normalized operational data from the Boujdour plant. The optimal model configuration was identified through multiple simulations, achieving an improved R² value and a reduced mean squared error (MSE). The findings from the membrane autopsy and the TMP prediction model provide valuable insights for future preventive measures, enabling a better understanding and control of reverse osmosis membrane fouling.

2. Material and Methods

2.1. Site of the Study

Boujdour City is situated in southern Morocco, on the Atlantic coast. It is part of the central Saharan provinces and is marked by an arid climate influenced by marine humidity along the coastline and a scarcity of both surface and groundwater. Between 1977 and 1995, the desalination plant utilized mechanical compression distillation to produce drinking water, with a capacity of 250 m³/day from seawater. Since 1995, the plant has adopted a reverse osmosis system for drinking water production. Initially, it operated with two reverse osmosis skids, each with a capacity of 30 L/s. In 2005, it was enlarged by another skid, providing a flow rate of 15 L/s. In 2016, the plant further increased its production capacity with a new reverse osmosis extension, offering a flow rate of 80 L/s [23]. Figure 1 illustrates the plant’s desalination process.

2.2. Description of the Treatment Process

The Boujdour desalination plant can be schematically divided into 4 stations.

2.2.1. Seawater Catchment

Borehole water abstraction uses submersible pumps to naturally filter the water through sediment layers in the coastal zone, which is then transferred to a reservoir for storage. In contrast to the mechanism of pumping directly from the beach, water pumped from boreholes provides the water free of organic germs and sand.

2.2.2. Pre-Treatment

The pre-treatment process at the Boujdour plant consists of four stages:

✓

Sulfuric Acid Injection

Sulfuric acid (H₂SO₄) is one of the strongest simple acids. The purpose of this injection is to prevent carbonate precipitation on the membranes and to supply CO₂ for post-treatment by lowering the pH. The released CO₂ reacts with lime to achieve a total alkalinity (TAC) value of 8 °F, according to Equation (1), as follows:

2CO₂ + Ca (OH)₂→Ca (HCO₃)₂

(1)

✓

Sand Filtration

This system removes most of the suspended particles, oil, and grease present in seawater. The filters consist of three layers: gravel, sand, and anthracite placed on a support structure. To maintain the performance of the sand filters, they are backwashed daily in three stages: compressed air, an upward water flow, and rinsing with a gradually decreasing water flow. Filtration occurs under a pressure of 4.1 bar.

✓

Sequestrant Agent Injection

A sequestrant is a substance that reacts with dissolved ions to prevent their precipitation. It is used primarily to prevent calcium precipitation, particularly calcium sulfates, on the reverse osmosis membranes. Additionally, it increases the size of suspended particles, facilitating their removal at the microfiltration stage.

✓

Microfiltration

After the water has been treated with the required reagents and passed through sand filtration, it undergoes microfiltration. This step protects the reverse osmosis membranes by capturing all microparticles larger than 5 μm. Microfilter cartridges are replaced when an excessive pressure drop (>0.6 bar) is detected in the microfilter system.

2.2.3. Reverse Osmosis Membrane System

Pretreated water enters the system at a pressure of 63 bar and is separated at the outlet into two types of water: filtered water (permeate), which has an acidic pH (pH = 6), and reject water (concentrate or retentate), which contains the compounds retained by the membrane. The pressure tubes used have a porosity of 0.1 nm and contain seven spirally wound polyamide membranes. This type of membrane is generally characterized by good chemical, thermal, and mechanical stability, as well as low cost. However, it has low permeability and is sensitive to chlorine and fouling. Each SKID consists of 17 pressure tubes, and each tube contains seven polyamide membranes, resulting in a total of 119 membranes per SKID.

Table 1. Characteristics of feed and treated water at Boujdour desalination station.

Composition	Feed Water	Treated Water	Moroccan Standards
pH	7.64	7.56	6.5–8.5
Turbidity (NTU)	1.43	0.21	1
Conductivity (μS·cm−1)	48,500	835.33	1300
Alkalinity TAC (meq·L−1)	2.95	1.26	-
Total Hardness TH (meq·L−1)	141.15	1.38	6
Boron (mg·L−1)	5.21	1.36	-
Sulphate (mg·L−1)	2960	25.17	200
Chloride (mg·L−1)	19,500	246.37	300
Fluoride (mg·L−1)	1.08	0.03	0.7
Calcium (mg·L−1)	538.99	6.97	<500
Magnesium (mg·L−1)	1408.12	14.67	100
Nitrate (mg·L−1)	1.86	0.16	<50
Ammonium (mg·L−1)	0.05	0.01	0.05
Silicate (mg·L−1)	8.92	0.03	-
Barium (mg·L−1)	0.11	0.006	0.7
Copper Cu2+ (mg·L−1)	0.08	0.00	<1
Total iron (mg·L−1)	0.50	0.1	<0.3
Zinc Zn2+ (mg·L−1)	0.08	0.01	<5
Manganese (mg·L−1)	0.11	0.02	<0.1
Dry residue (g·L−1)	37.10	0.47	1

provides details on the capacity of the reverse osmosis unit skids.

2.2.4. Post Treatment

The permeate from reverse osmosis is stored in the treated water tank, where it is remineralised with lime to raise the pH and achieve the desired calcium-carbonate balance (the pH must be between (8–8.5)) and disinfected by chlorination to protect the water in the distribution circuit, ensuring residual oxidation of microorganisms during transport and storage.

2.3. Membrane Chemical Cleaning Process at Boujdour Plant

After the reverse osmosis plant has been in operation for some time, the membranes may become clogged with various soluble and suspended substances present in the feed water. Membrane cleaning is typically required in the following cases:

•

A 10–15% decrease in normalized permeate flow;

•

A 10–15% increase in normalized permeate conductivity;

•

A 10–15% increase in the normalized pressure drop between feed and concentrate.

For the most complex clogging phenomena, the descaling process indicated in Figure 2 remains the most effective solution at Boujdour desalination plant.

The chemical cleaning solutions used are as follows:

•

Acid phase: A solution with a pH of 2 to 3 is typically prepared using 2% citric acid (C₆H₈O₇), which corresponds to 7.7 kg of citric acid dissolved in 379 L of water;

•

Basic phase: A basic solution with a pH of 10 to 11 is used, generally composed of 2.0% sodium tripolyphosphate (STTP) and 0.8% Na-EDTA.

Figure 3 illustrates the adopted system for the chemical cleaning of the reverse osmosis skid.

2.4. Feed and Treated Water Characteristics

The supply of water is sourced from coastal wells rather than being pumped directly from the ocean to ensure it is free from sand and other contaminants. Table 1 presents the average quality parameters for both feed water and finished water, with reference to Moroccan drinking water standards [24].

2.5. Preparation of Membrane Samples

The most effective approach to determining the source of membrane fouling is conducting a membrane autopsy [25]. For this study, an element of the fouled membrane, which had been in service for nearly 5 years at the Boujdour plant, was destructively analyzed. The membrane under examination is a natural spiral-wound reverse osmosis membrane (FILMTEC SW30XHR-440i) made of an aromatic polyamide composite, known for its superior salt rejection performance (manufactured by Dow Chemical, Midland, MI, USA). A sample was taken directly from the plant by cutting the clogged membrane and separating it from the spiral module. This was achieved using a motorized circular saw to cut through the layers. Subsequently, small square membrane coupons, measuring 7 cm × 7 cm, were cut and stored in boxes for future analysis.

2.6. Analysis of the Fouled Membrane

To analyze the composition of clogged materials on the membrane surface and examine its morphology, a membrane autopsy was conducted using scanning electron microscopy (SEM) coupled with energy-dispersive X-ray spectroscopy (EDS), specifically the Thermo Scientific Quattro SEM-EDS system (Thermo Fisher Scientific, Waltham, MA, USA). This analytical method is among the most widely employed techniques for identifying membrane fouling [26,27]. Several SEM images were captured at different magnifications across various parts of the membrane sample to provide insights into the type of fouling. The analyses were carried out at the Centre National for Scientific and Technical Research (CNRST).

2.7. Data Collection

The analysis was based on actual operating data collected from the Boujdour reverse osmosis desalination plant, specifically the new facility. The data used to understand the impact of fouling membrane on plant performance concern daily data for the year 2020, while the data used to make the ANN prediction cover the years 2019 and 2020. The target parameter for the analysis was the transmembrane pressure, which is currently considered a key indicator of membrane fouling. The input parameters included temperature, turbidity, pH, feed conductivity, and feed flow. The data were normalized according to the normalization method outlined in Excel [23]. The artificial neural network model was developed using MATLAB/Simulink software (R2015a).

2.8. Artificial Neural Network

Artificial intelligence (AI) is currently making remarkable progress in predicting the behavior of membrane separation in reverse osmosis units. Among AI methods, neural networks are a powerful tool for modeling complex non-linear systems [28]. Clogging remains a challenging phenomenon to address due to variations in feed water quality and the type of pre-treatment. Artificial neural networks (ANNs) have increasingly been applied to develop regression models and predict such phenomena [29]. In this study, a multilayer neural network was constructed based on the backpropagation algorithm, with five input parameters, to predict the TMP. The model’s performance was optimized by varying the number of neurons, hidden layers, and activation functions to identify the best-performing configuration, using MATLAB software (R2015a).

3. Results and Discussion

In this research, the results are presented in three parts. The first part examines the effect of membrane fouling on the performance of the Boujdour reverse osmosis plant. The second part discusses the membrane autopsy using the SEM/EDS method. The third part presents the results of TMP prediction modeling using the ANN method.

3.1. Impact of Membrane Fouling on the Performance of Boujdour Reverse Osmosis Plant

To explain the influence of the fouling membrane on operating parameters in particular, i.e., permeate conductivity, permeate flow, and TMP, the monitoring of these variables was observed during one year of service in accuracy in the year 2020.

Figure 4 illustrates fluctuations in permeate flow, transmembrane pressure (TMP), and permeate conductivity during 2020. The permeate flow starts at approximately 380 m³/h in January 2020. A slight drop occurs, followed by moderate fluctuations, but the flow remains relatively stable between 350 and 400 m³/h until early March. During this period, a noticeable drop is observed, with the permeate flow rapidly declining to around 300 m³/h. Subsequently, the flow stabilizes at approximately 250–300 m³/h for the remainder of 2020. Minor fluctuations occur, but no significant upward or downward trend is observed after this stabilization. The sharp drop in March–April might be attributed to a system change, such as membrane fouling. After this decline, the stabilized flow rate suggests that conditions had normalized or that adjustments had been implemented to maintain the flow [30]. The TMP is calculated as the difference between the average feed pressure and the permeate pressure [31]. The graph illustrates a general upward trend in TMP over the year. The pressure starts at approximately 0.15 bar in January and increases to around 0.25 bar by December, with peaks reaching up to 0.35 bar. The graph reveals several fluctuations, with alternating periods of rising and falling pressure. The most notable fluctuations occur in May, July, September, and November, showing significant increases followed by slight decreases. Marked increases in TMP, particularly toward the end of the year, may indicate increased membrane fouling. This could result from the accumulation of organic, inorganic, or microbiological deposits on the membrane surface. Permeate conductivity is an important indicator of treatment performance. The graph shows small variations, with periodic increases and decreases. There is a slight upward trend in permeate conductivity during 2020, with values rising from approximately 1100 μS/cm to 1350 μS/cm. An increase can be seen in late April and early May, coinciding with the increase in TMP and the decrease in permeate flow. The change in permeate conductivity can also be explained by the clogging of RO membranes [32]. In summary, the graph suggests that the reverse osmosis membrane undergoes periodic fouling, leading to an increase in TMP and permeate conductivity and a decrease in permeate flow, with significant clogging observed around April and May.

3.2. Membrane Autopsy

Analyses were performed at the National Centre for Scientific and Technical Research (CNRST) using scanning electron microscopy and energy-dispersive spectroscopy (SEM/EDS). To better understand the structure and morphology of the elements, a visual inspection was conducted, and two SEM/EDS methods were applied: the first focused on the membrane’s cross-section, while the second examined its top surface.

3.2.1. Visual Inspection of Membranes

Visual inspection provides an initial understanding of the fouling affecting the membrane [33]. Examination of the clogged membrane sample, as shown in the images in Figure 5, revealed no physical deformation caused by pressure, with no visible cracks or significant deterioration [34]. The membrane displayed a rugged texture and a lattice structure, which are typical features of reverse osmosis membranes that rely on a dense fine layer to block salts and impurities. Deposits are irregularly distributed across the membrane surface, with a light brown coloration covering the surface. This coloration is likely due to the high sand content in the seawater and organic matter, the exact composition of which is yet to be determined. The fouling layer adhered to the membrane surface is neither viscous nor gelatinous, and the deposited material can be easily detached from the membrane. The uniform layer of deposits observed on the surface may indicate the presence of biofilms, mineral deposits, or organic material.

3.2.2. EDS/MEB Cross-Section Results

SEM analysis was applied to the extracted membrane sample to obtain a clear inspection of the clogged material’s form [35]. SEM images of the cross-section at different resolutions (500×, 1000×, and 800×) are shown in Figure 6. These images reveal globular and tubular structures, some with smooth surfaces and others with slightly rough surfaces. The globular structures are likely the membrane pores. The annotated SEM image includes measurements of 17.27 µm, 24.28 µm, 22.54 µm, and 92.54 µm. The 17.27 µm measurement corresponds to the size of the membrane pores, the 24.28 µm measurement corresponds to the thickness of the clogged layer on the membrane surface, and the final 92.54 µm measurement corresponds to the internal thickness of the membrane. For cross-sectional EDS analysis, the sample was oriented perpendicularly to the incoming electron beam. This method allows for the detection of the evolution of clogged material along the thickness of the membrane. The figure shows the EDS results for the cross-section. The EDS graph in Figure 5 shows the variation in the chemical elements present, indicated by the y-axis, which corresponds to the intensity of the X-ray signal detected, in relation to the x-axis, which represents the distance in microns across the sample analyzed, as marked by the yellow arrow in the preceding SEM results. The higher the number of net counts, the greater the concentration of an element in that specific region. Based on the results, a significant number of elements were detected. The colored lines represent the different elements detected in the sample, each assigned a different color: Carbon (C), Oxygen (O), Sodium (Na), Magnesium (Mg), Aluminum (Al), Silicon (Si), Sulfur (S), Chlorine (Cl), Potassium (K), and Iron (Fe). High peaks indicate a higher concentration of an element in that specific region of the sample, and the variability in element levels at different distances demonstrates the heterogeneity of the chemical composition of the sample surface.

From the previous diagram, the following compositions are observed:

•

Carbon (C): High peaks are seen around 0–30 microns, 50 microns, and 70–90 microns, indicating a significant presence of carbon in these regions. These carbon peaks correspond to the globular structures or organic fibers observed in the SEM image, suggesting a carbon-rich composition, likely organic materials or composites in the membrane;

•

Oxygen (O): A notable concentration of oxygen is observed, especially toward the end of the measurement range, around 80–110 microns. The presence of oxygen is associated with areas of roughness or crack interfaces, potentially due to surface oxidation. Areas with high concentrations of oxygen and other elements like iron may indicate oxidation or corrosion, identifiable as spots or regions with a different texture;

•

Chlorine (Cl): Also shows noticeable concentration toward the end of the range, from 80 microns upward;

•

Sulfur (S): Observed mainly around 80–110 microns. Sulfur and chlorine peaks suggest the presence of mineral compounds on the rough surface, which may be correlated with cracks and fissures observed in the SEM image;

•

Other elements (Na, Al, Si, Mg, Fe): These show relatively low levels, but slight variations indicate their presence in the sample. These elements show increases toward the end, suggesting a region rich in these elements. Traces of these elements may represent impurities that could be dispersed in the areas observed.

The variation in element concentrations indicates that the sample is chemically heterogeneous. Some regions are rich in carbon, while others are rich in oxygen, sulfur, or chlorine. High concentrations of certain elements can help identify the materials present. For example, a high concentration of carbon may indicate organic compounds, as the membrane is polyamide in nature, while high concentrations of silicon and oxygen may suggest the presence of silicates.

3.2.3. EDS/SEM Top Surface Analysis Results

SEM Top Surface Result

Figure 7 shows SEM images of the clogged top surface of the membrane sample at various magnification scales (1500×, 250×, 2000×, 8000×) across multiple points. The images clearly reveal that the surface is entirely covered with particles of varying shapes and sizes. The surface displays crystalline or granular structures, with the crystals appearing predominantly cubic or rhombic shapes that are often characteristic of specific crystals or precipitates. These crystals are dispersed on a darker matrix, where the contrast indicates that the crystals are denser or of a different composition compared to the underlying substrate. In the center of image (d), an elongated, oval, and symmetrical structure is observed, which appears to be a microorganism or a diatom (a type of microalgae). This structure stands out due to its contrasting morphology compared to the cubic shapes of the surrounding crystals. The images also highlight a heavily fouled membrane layer, consisting of particles encrusted within a crystallized matrix. These particles will be identified through EDS analysis [36]. Additionally, spherical shapes are visible, likely indicative of inorganic salts such as calcium carbonate (CaCO₃) [37].

EDS Top Surface Result

With the aim of studying the structure of the various substances deposited on membrane top surface layers, three selected points on the sample (noted Base 7) are shown in Figure 8; they are analyzed by the EDS technique and the results of the chemical composition are shown in the same figure. Base (7)_point 1, Base (7)_point 2, and Base (7)_point 3 are shown with tables of proportion concentration ions (weight and Atomic) detected in three points in fouled membrane. The aim of choosing these 3 points is to know the composition of each element positioned by the points chosen, as each point corresponds to a different geometric form of the elements making up the clogged substances on the membrane surface.

The major elements detected by EDX are as follows, in descending order:

Point 1: Cl > O > Na > Si > Fe > Al > Mg > K > S.

Point 2: Cl > Na > O > Si > Al > Fe > K > Mg > S.

Point 3: O > Si > Cl > Al > Na > Fe > Mg > K > S > Ca.

It is clear that Cl, O, and Na are the predominant deposited elements. Si, Fe, and Al were detected in secondary quantities, while K, Mg, S, and Ca constitute the minority of deposited elements. The EDS analyses show results consistent with previous studies [38,39], indicating that the deposited substances contained significant quantities of Na, Cl, and O, with low proportions of Mg and K. Based on the EDS results from the three points, the global atomic proportions of detected elements, on average, are as follows:

O (39.69%) > Na (22.83%) > Cl (22.48%) > Si (7.99%) > Al (2.79%) > Mg (1.56%) > Fe (1.27%) > K (0.87%) > S (0.36%) > Ca (0.12%).

From the EDS results, the following observations can be made:

•

The elevated levels of Cl and Na can be attributed to their significant presence in seawater [40];

•

The high concentration of O is explained by the presence of organic deposits. This organic matter originates partly from the composition of the feed water and partly from the use of a sequestering agent. While the sequestering agent inhibits particle agglomeration and attachment to the membranes, it can also act as a nutrient for bacterial cultures on reverse osmosis membranes [41]. Additionally, in the SEM image Base (7) _Point 1, the fouled matter appears to have a bacterial form, correlating with the high concentration of O at this point;

•

Biofouling can be eliminated by using modified membranes such as natural antimicrobials [42] or by grafting or depositing on the surface a multilayer covered with nanoparticles [43,44];

•

The presence of Al and Si indicates aluminum silicates, which are common pollutants identified during reverse osmosis membrane autopsy [45];

•

The Fe concentration may result from two sources: the chemical composition of materials used in the system (e.g., pipes, pumps, etc.) and the use of an iron-aluminum coagulant;

•

The low concentrations of Mg, K, and Ca are explained by their relatively low presence in the feed water [46];

•

The small amount of S is attributed to the injection of sulfuric acid during the pre-treatment process. This acid injection helps precipitate carbonates from the membranes and provides CO₂ for post-treatment while lowering the pH.

3.3. ANN Fouling Prediction Results

In this section, an artificial neural network (ANN) model was developed to predict the transmembrane pressure (TMP) using MATLAB software. The chosen network was a multilayer perceptron with a backpropagation learning algorithm. The data were normalized and divided into three categories: 70% for training, 15% for validation, and 15% for testing.

Table 2. Input and output parameters and their interval for ANN prediction.

Number	Parameter	Interval	Units
Inputs:
1	Temperature	19.1–27	°C
2	pH	7.2–7.6
3	Turbidity	0.17–0.88	NTU
4	Feed conductivity	41,500–47,800	µs/cm
5	Feed flow	273–302	m3/h
Output:	TMP	0.1–2.14	Bar

summarizes the input and output parameters, along with the range of values considered in this study to predict the TMP. The inputs and targets were normalized to the interval [0, 1] by the following Formula (2):

$$X^{\prime }={\displaystyle \frac{Xi-Xmin}{Xmax-Xmin}}$$

(2)

Xi is real data, Xmin is the minimum, and Xmax is the maximum value of the input and output variables [47].

More than 200 simulations were performed to determine the optimal number of hidden layers, neurons, and the type of activation function. A selection of these simulations is summarized in

Table 3. ANN training trials with different numbers of nodes, layers, and activation functions.

Nbr of Simulations	Nbr of Layers	Activation Function	Nbr of Neurons	MSE	R %
					Training	Validation	Test	All
1	1	Tang	3	0.020791	0.76	0.62	0.822	0.75168
2	1	Log	15	0.018674	0.87	0.84	0.799	0.846
3	2	Tang-Tang	5-10	0.0308	0.9692	0.356	0.9666	0.89772
4	2	Log-Log	30-40	0.010326	0.89674	0.72586	0.70374	0.85496
.	.	.	.	.	.	.	.	.
.	.	.	.	.	.	.	.	.
197	3	Tang-Log-Tang	20-15-15	0.07863	0.92687	0.70317	0.61521	0.83224
198	3	Log-Log	30-15	0.0074753	0.97195	0.88211	0.58363	0.88758
199	3	Tang-Tang-Tang	5-10-5	0.005657	0.92077	0.92403	0.7127	0.88553
200	3	Log-Tang-Log	30-20-10	0.011585	0.88506	0.938	0.87907	0.88851

The bold form of model 199 represents a distinction from the optimal model.

to identify the best ANN model for predicting the TMP.

3.3.1. ANN Architecture Model

Based on the iterations performed, the successful model corresponds to trial 199 (Table 3). It consists of three hidden layers: the first layer contains 5 neurons, the second 10 neurons, and the third 5 neurons. The activation function used was a tangent function in all three layers, with the backpropagation (BP) algorithm employed for learning and the Levenberg–Marquardt function used as the learning algorithm. Figure 9 illustrates the diagram of the identified ANN model [48].

3.3.2. ANN Performance Model

The developed model corresponds to a significant training R² of 92.077%, indicating a strong correlation between the targets and outputs for the training data. This means the model fits well with the data used for training. The validation phase has an R² of 92.403%, showing a slightly better correlation than that of the training phase, which indicates that the model generalizes well to the validation data. The test phase shows an R² of 71.27%, which is lower, indicating a moderate correlation between the targets and predictions. This may suggest some degree of overfitting or that the test set is more challenging or different from the training/validation sets. The overall R² is 88.553%, indicating a strong global correlation, although slightly lower than that of the training and validation sets, which demonstrates good model performance with a minimum MSE of 0.005657. Figure 10 illustrates the regression plot results of the ANN model at the three stages of modeling.

Figure 11 illustrates the evolution of the MSE over the training, validation, and test sets as a function of epochs. It can be observed that the training phase error decreases steadily over the epochs, which is typical of a model that adapts well to the training data. The validation phase error initially decreases, then begins to slightly increase after epoch 21. This suggests that the model starts to overfit after this epoch. The test set error is higher than the other two curves, indicating some difference between the training/validation sets and the test set (potentially due to a different distribution or a more challenging sample). The best validation performance, with an MSE of 0.005657, was achieved at epoch 21. This means that the model reached its optimal generalization point at this stage. This epoch is marked by a green vertical line with a circle at the intersection of the validation curve. The model is optimal at epoch 21 for minimizing the validation error, representing the best trade-off between a good fit on the training data and good generalization to external data. After this epoch, the model risks overfitting, which could degrade its ability to generalize.

According to the ANN model results, we can confirm that the developed model has a superior performance with an R² of 92.077% and an MSE of 0.005657. The model was compared with two ANN models from the literature for TMP prediction. Chengxin et al. [49] found an ANN model with performance (R² = 0.906, MSE = 0.064). Tianjie et al. [50] developed another TMP prediction model with an R² of 0.8622. These results show that our ANN model outperforms the others, with an R² of 92.077% and an MSE of 0.005657. However, it is important to note that the ANN model was trained using data from Boujdour desalination plant, with a specific configuration of membranes and modules in order to better understand the operating mechanisms of the plant. Consequently, any change in supplier, membrane type, or configuration may require the model to be readapted or retrained to ensure accurate predictions. Also, this model cannot be generalized to all reverse osmosis plants, as its performance is highly dependent on the configuration of the membranes, the operating conditions, and the local characteristics of treated water.

4. Conclusions

In this research, a structural membrane autopsy analysis was conducted to understand the nature of fouling in the Boujdour desalination plant. Additionally, an efficient ANN model for predicting membrane fouling through TMP was developed using MATLAB/Simulink. The effect of fouling on treatment performance was studied by monitoring three operating parameters: TMP, permeate flow rate, and permeate conductivity. An autopsy of one fouled membrane after five years of service was performed using visual inspection and electron microscopic techniques (SEM/EDS). The optimal ANN model developed includes three hidden layers (5Tang-10Tang-5Tang) using the BP algorithm. The results of this study can be summarized as follows:

•

Fouling significantly impacts treatment performance. A marked reduction in permeate flow, along with a considerable increase in TMP and a slight rise in permeate conductivity, were observed, particularly during the months of March, April, and May;

•

Surface analysis of the fouled sample showed that all membrane layers were light brown in color and covered with various forms of crystal spheres and other dirt types;

•

SEM/EDS cross-sectional examination of the fouled membrane indicated a surface foulant layer thickness of 24.28 µm. The main detected elements were C, O, Cl, and S, with smaller quantities of Na, Al, Si, Mg, and Fe;

•

SEM/EDS analysis of the top surface of the fouled membrane revealed organic fouling, indicated by a high concentration of O, and inorganic fouling, including Al, Si, Fe, and other inorganic substances such as Mg, K, and Ca in limited proportions, due to the chemical composition used in pre-treatment and the nature of the feedwater (seawater);

•

From the SEM/EDS results of both the top surface and cross-section, similar organic and inorganic fouling was detected, with common elements including O, Al, Si, Mg, Fe, S, and Na;

•

It should be noted that this study did not focus on biological fouling but rather on inorganic fouling, due to the limited experimental time available and the absence of specific microbiological characterization techniques. This aspect, which is essential for an overall understanding of fouling mechanisms, will be the subject of a future research project;

•

This study demonstrated the robust capability of the ANN model for predicting membrane fouling in the Boujdour reverse osmosis desalination plant. The ANN showed excellent performance, with an R² of 92.077% and an MSE of 0.005657, based on operating data from the plant;

•

The research presents a new strategy for efficiently optimizing the prediction of membrane fouling, which is beneficial for better understanding and controlling membrane fouling in the Boujdour desalination plant;

•

To improve treatment at the Boujdour plant and overcome the problem of membrane fouling, the pre-treatment process should be focused on the elimination of organic and inorganic matter, particularly Si, Al, and Fe, before the reverse osmosis process;

•

The use of powerful chemical agents in membrane cleaning may cause membrane surface degradation as well as environmental impacts and high costs; however, it is recommended to study the effectiveness of appropriate chemical agents under careful control and explore more sustainable alternatives such as modified membranes.

•

In view of the increased cost of advanced technologies such as ultrafiltration and nanofiltration and their higher energy consumption, it is preferable to optimize the current pre-treatment at Boujdour plant, which is based on sand filtration, microfiltration, and addition of antiscalents and sulfuric acid. This aspect will be examined in greater detail in future studies.