Modeling and optimization of nonylphenol removal from contaminated water media using a magnetic recoverable composite by artificial neural networks

In the past few decades, rapid development of industrialization has triggered environmental pollution by various contaminants. Among all hazardous organic pollutants, nonylphenol (NP) is intrinsically a xenophobic substance from an organic compound family known as alkylphenols. NP is used in the manufacture of various materials such as laundry and dish detergents, emulsifiers, solubilizers, and antioxidants (USEPA 1996). However, the major reason for the presence of NP in the water and soil resources is the decomposition and degradation of nonylphenolethoxylates (Birkett & Lester 2002). The typical concentrations of NP in river water, landfill leachates and drinking water have been reported in the ranges 0.7 ng/L to 1.5 μg/L (Bester et al. 2001), 10–170 μg/L (Öman & Hynning 1993) and 15–85 μg/L (Petrovic et al. 2003), respectively. Despite the extremely low concentration of NP in aquatic environments, it can result in massive toxicity in marine creatures.

Up to now, a wide range of remediation techniques such as photochemical oxidation (Fonseca et al. 2004; Ide et al. 2011), electrochemical decomposition (Kuramitz et al. 2002), and biodegradation (Armenante et al. 1999) have been applied to remove NP from the environment but their financial and operational difficulties have made researchers hesitant to employ them in a large scale. Therefore, the application of a promising, innovative, and economical approach seems to be a vital necessity. Adsorption is known as one of the most desirable processes to control the excessive concentrations of such pollutants. Hence, different conventional adsorbents (e.g. activated carbon (Jafari et al. 2016), carbon nanotube (Shirmardi et al. 2013), sediments (Liping et al. 2014) and spent tea (Babaei et al. 2015)) have so far been used for the adsorption of various contaminants. Of these applied adsorbents, activated carbon has been more welcomed by environmentalists compared with other adsorbents for its cost-effectiveness feature, a high porosity in structure, and a large capacity to adsorb a wide spectrum of contaminants (Asgari et al. 2013; Kakavandi et al. 2016a). Although these adsorbents have a high adsorption capacity, specifically in low concentrations of contaminants, their difficult application in process engineering due to their small particle size, dispersion in the medium, and separation and filtration problems has obliged researchers to synthesize the magnetic adsorbents.

In general, findings of previous studies have shown that iron particles, either in micro or nano sizes, have an undeniable role in treatment of contaminated aqueous media (Liu et al. 2015; Xing et al. 2016). In addition to their high tendency towards reacting with a wide range of contaminants, their magnetic features enable them to be separated from aqueous solution using a strong magnetic field (Esfahani et al. 2014a). On the other hand, powder activated carbon (PAC) triggers several problems in site remediation due to its high dispersivity and tiny particle size. Therefore, combining both PAC and iron magnetic nanoparticles will not only provide better kinetics for the adsorption of contaminants, but also prevent the aforementioned operational problems and secondary pollution (Mohseni-Bandpi et al. 2015; Kakavandi et al. 2016b).

The conventional methods of examining batch experiments include studying the impacts of a factor by keeping the other effective experimental agents constant. Obviously, this approach suffers from a lack of possibility to find the interaction effects of input variables. It also increases the costs of operation by the need for a large number of experiments. Such restrictions can be decreased by means of applying an optimization approach that covers not only the interaction effects of input variables but their individual impacts as well. Among all optimization techniques, artificial neural networks (ANNs) constitute one of the most well-known and authenticated methods, which have been used for optimizing a wide range of remediation processes such as reduction, adsorption, Fenton oxidation (Kakavandi et al. 2016c), and photocatalytic degradation (Khataee et al. 2010). The ANN is intrinsically a progressive mathematical method which draws the output according to the experimental conditions. In addition, it was widely applied to model the rate of removal in the adsorption of contaminants in a solid–liquid system (Balci et al. 2011). A brilliant study in this regard would be that of Khataee and co-workers (Hassani et al. 2014) in which they predicted the effects of each experimental variable on the adsorption efficiency of a cationic dye on Turkish lignite by means of neural network modeling.

Although the ANN approach was applied to optimize the removal of various contaminants from an aqueous solution, its efficiency for the modeling of NP removal using Fe₃O₄/AC nanoparticles has not been discussed fully. Therefore, the aims of this study are the optimization of NP removal using Fe₃O₄/AC as well as finding the effects of each input variable. Furthermore, the best kinetic and isotherm models for NP adsorption are introduced. The adsorption thermodynamics and reusability of the Fe₃O₄/AC nanoparticles are also studied. Finally, the potential of the composite for removal of NP in real water is evaluated.

Stock solution of NP was prepared by dissolving a suitable amount of NP in a specific volume of methanol. NP was obtained from Sigma Aldrich Co. while concentrated hydrochloric acid (12 N HCl), sodium hydroxide (NaOH), acid nitric (HNO₃, 65%), PAC and ferric nitrate (Fe(NO₃)₃.9H₂O) were purchased from Merck, Darmstadt, Germany. All reagents were analytical grade.

The structure and surface characteristics of AC and Fe₃O₄/AC nanoparticles and size of Fe₃O₄ particles were evaluated with a scanning electron microscope (SEM, Philips XL30) and transmission electron microscope (TEM, Philips XL208). The X-ray diffraction (XRD) pattern of the samples was prepared using a Quantachrome Nova2000 instrument in 2θ of 10–70°. Additionally, the elemental analysis of Fe₃O₄/AC composite was determined using a SEM equipped with energy dispersive X-ray analyzer (EDX) (Philips, XL-30). The textural properties of the samples (e.g. specific surface area and pore volume) were investigated using the Brunauer, Emmett and Teller (BET, Quantachrome, NOVA 2000) method using N₂ adsorption–desorption isotherms at 77.3 K. The magnetic properties of the composite were also evaluated by a vibrating sample magnetometer (VSM) (7400, Lakeshare, USA) at room temperature.

To design a typical adsorption system, it is critical to determine the adsorption parameters. The adsorption equilibrium is always explained by the isotherm equations, whose parameters indicate the surface characteristics and affinity of the adsorbent toward the adsorbate. In order to determine the adsorption equilibrium of NP onto the Fe₃O₄/AC composite, four isotherm models – Langmuir, Freundlich, Liu, and Temkin – were studied. For further details of these models, please see Supplementary material Table S1 (available with the online version of this paper). Additionally, the study of kinetic adsorption of NP was performed by six different models: pseudo-first-order equation of Lagergren, pseudo-second-order equation of Ho, Elovich, fractional power function, Avrami fractional order, and intra-particle diffusion. For further details of these models, please see Supplementary material Table S1.

To evaluate the possibility of regeneration and reuse of the adsorbent, methanol, as a desorbing solution, was used to extract the adsorbed NP on Fe₃O₄/AC. The reusability of the adsorbents was determined using five adsorption-regeneration cycles. A sample of 0.1 g of Fe₃O₄/AC nanoparticles was shaken with 100 mL solution of 1 mg/L NP for 4 h at 25 ± 1 °C and pH 3.0. The solid product was magnetically collected, washed and then dried at 80.0 °C. Afterwards, samples of 0.05 g of Fe₃O₄/AC nanoparticles loaded with NP were shaken at 200 rpm for 24 h with 10 mL of desorbing solutions at 25 ± 1 °C. Then the desorption percentage was determined using the ratio of the weight of desorbed NP and the weight of adsorbed NP. After desorption, regenerated adsorbents were dried in an oven at 80.0 °C for 100 min and were used for the next adsorption-regeneration cycle.

As shown in XRD analysis (Figure 2(f)), a broad diffraction peak at 2θ = 25° can correspond to the characteristic reflection of carbon of amorphous nature, and this peak was narrow and weaker for Fe₃O₄/AC composite when compared with the virgin AC. For Fe₃O₄ nanoparticles, the results revealed that the XRD pattern of Fe₃O₄ was in a good agreement with that of pure magnetite. As shown in Figure 2(f), six sharp peaks at 2θ values of 30.07° (220), 35.44° (311), 43.15° (400), 54.6° (422), 56.99° (511), and 62.6° (440) in the patterns of both Fe₃O₄ and Fe₃O₄/AC were attributed to the characteristic peaks of the cubic phase Fe₃O₄ according to the JCPDS No. 19-0629. These results illustrated that the Fe₃O₄ nanoparticles have successfully been coated on the AC surface.

Textural properties (specific surface area, volume and average diameter of the pores) of AC, Fe₃O₄ and Fe₃O₄/AC were analyzed using BET technique and results are presented in Table 1. The results of the BET analysis indicated that the highest surface area of adsorbent was 671.2 m²/g, which was lower as compared to virgin AC (733 m²/g). This can be explained by the formation of Fe₃O₄ inside the AC pores and/or loss of AC pores by Fe₃O₄ nanoparticles, as reported in the literature (Kakavandi et al. 2016c). The average size and volume of pores of the composite were obtained to be 3.5 nm and 4.87 mL/g, respectively. This average size, based on the IUPAC classification (micropores (d < 2 nm), mesopores (2 < d < 50 nm) and macropores (d > 50 nm)), could be classified into the mesopores group (Jafari et al. 2016). Maximum saturation of magnetization (6.94 emu/g) was found to be for Fe₃O₄/AC nanoparticles according to the VSM analysis. As illustrated in Figure 2(e), the composite showed an excellent magnetic response to a magnetic field, suggesting that it could be separated easily and rapidly due to this high magnetic sensitivity. Therefore, Fe₃O₄/AC nanoparticles can be potentially employed as a magnetic sorbent to remove pollutants from the aqueous media, avoiding a secondary pollution, without filtration and centrifuging.

The high adsorption efficiency at pH of 3 can be attributed to the interaction between the dipole of the phenol function in NP and the positively charged Fe₃O₄/AC surfaces (Babaei et al. 2011). Moreover, at low pH values, which are full of hydrogen ions in solution, and hydroxyl groups form, NP did not dissociate. Under the driving of π–π stacking interaction, the hydrogen bond could be formed easily leading to an increase in the adsorption efficiency (Pan et al. 2013). However, at alkaline conditions, the surface of the adsorbent is negative and thus an electrostatic repulsion occurs between NP molecules for the adsorption sites, which tends to decrease the adsorption capacity. In addition, it is possible that under this condition the OH⁻ ions compete with NP molecules for active sites on the Fe₃O₄/AC. According to Table 1, the pH_pzc value of Fe₃O₄/AC was 6.8, indicating that the surface charge of the adsorbent is positive when pH_pzc > pH and is negative at pH_pzc < pH. Therefore, at acidic conditions (pH_pzc > pH) the Fe₃O₄/AC surface charge is positive and the electrostatic force between the adsorbent surface and NP molecules increases. At pH values >6.8 (pH_pzc < pH), however, the surface charge of Fe₃O₄/AC is negative and the adsorption potential decreases significantly due to the electrostatic repulsion between NP molecules and the adsorbent surface, as illustrated in Figure 3(a).

The contact time between the adsorbate and the adsorbent is the other important parameter that affects the performance of any adsorption process. The effect of contact time on the NP adsorption capacity for initial various concentrations is illustrated in Figure 3(b). The results showed that the rate of NP adsorption onto Fe₃O₄/AC nanoparticles was rapid initially and then decreased gradually until the equilibrium was reached, beyond which there was no further adsorption. This can be explained by the fact that active sites are initially vacant and are then filled by adsorbate with increasing contact time, finally leading to a saturated adsorbent surface (Babaei et al. 2015). Figure 3(b) shows that the adsorption capacity of NP was almost constant for the contact times after 30 minutes at all the studied concentrations. In fact, the adsorption process reaches the equilibrium point at 30 min. Therefore, this was chosen as the equilibrium time of NP adsorption onto Fe₃O₄/AC nanoparticles.

The mass of adsorbents is another likely feature that affects the removal of any kind of contaminants. In this regard, various amounts of Fe₃O₄/AC nanoparticles as an adsorbent (0.02–1.5 g/L) were used for the adsorption of 5 mg/L NP at optimum pH of 3 which was obtained in previous experiments with 30 min contact time. Based on Figure 3(c), a direct relationship between adsorbent dosage and adsorption efficiency of NP is observed. Specifically, by increasing the adsorbent dosage from 0.02 to 0.2 g/L, adsorption efficiency of NP rose significantly from 59.93 to 98.51%. Afterwards, the graph represents a steady state pattern that confirms the optimum adsorbent dosage for NP adsorption is 0.2 g/L, which was applied in the following experiments. From these results, it can be concluded that the interaction between adsorbent and adsorbate confirms a relationship whereby sorption sites are increased by enhancing the mass of adsorbents (Farasati et al. 2013). This finding is in agreement with the results of previous studies reported in the literature (Babaei et al. 2015).

Another important experimental factor examined with regard to the adsorption of NP was the initial NP concentration. The adsorption of NP at different ion concentrations (1, 2.5 and 5 mg/L) was carried out in an optimum pH of 3 and 0.2 g/L of Fe₃O₄/AC nanoparticles for a period of 30 min at room temperature. The obtained results are illustrated in Figure 3(d). Accordingly, higher NP adsorption was observed in lower NP concentrations. In other words, increasing NP concentrations caused a significant decrease in the removal efficiency of Fe₃O₄/AC nanoparticles. As can be seen in Figure 3(d), the adsorption efficiency of NP dropped when the initial NP concentrations increased from 1 to 5 mg/L. The probable reason behind this phenomenon can be related to the filling of active sites onto the functional groups of the surfaces of adsorbents, which were intensified by increasing the initial concentrations of the adsorbate (Soltani et al. 2014).

As shown in Table 2, the values of rate constants (k_f, k_s and k_Av) decrease by increasing the initial NP concentration. This can be due to the high competition of adsorbate molecules for the adsorption surface sites at high concentrations, which leads to lower sorption rates. This observation has also been reported by other researchers studying the adsorption of contaminants (Depci et al. 2012; Azari et al. 2015).

The intra-particle diffusion model can be used to verify the influence of mass transfer resistance on the binding of adsorbate to the adsorbent. NP is probably transported from aqueous solution to the adsorbents by intra-particle diffusion. To investigate this issue, the results for kinetic adsorption were fitted with the intra-particle diffusion model. The lower values of R² for the intra-particle diffusion model show that this model is not applicable for NP adsorption on Fe₃O₄/AC nanoparticles. Furthermore, as observed in Table 2, the values of C of the intra-particle diffusion model are not equal to zero at any of the studied concentrations, indicating that intra-particle diffusion is not the only controlling step for the NP adsorption process. Hence, other mechanisms such as complexes or ion-exchange may also control the rate of adsorption (Kakavandi et al. 2016a).

Isotherm investigation is one of the most integral types of adsorption studies that evaluate the reaction between both the adsorbate and the adsorbent. Therefore, we studied the isotherm behavior of NP adsorption onto the surfaces of Fe₃O₄/AC composite using four models. Like the kinetic studies, isotherm parameters were fitted via non-linear regression analysis using the MATLAB software package. The non-linear graphs of the isotherm of NP adsorption at different temperatures, drawn based on q_e vs. C_e, are given in Figure 4(d)–4(f). Additionally, Table 3 shows the obtained isotherm parameters of NP adsorption. According to Table 3, of all the applied models, Liu was more capable of fitting adsorption data of NP removal with its high coefficient of determination (R² > 0.99) and the low root-mean-squared error (RMSE) and sum of squared error (SSE) values in all temperatures. On the other hand, in the same experimental conditions, Temkin was found to be the least suitable isotherm model in fitting NP adsorption data. Also, the adsorption capacities of the Liu model represent a direct relationship with temperature: adsorption capacity decreased significantly by increasing temperature, illustrating that the process of NP adsorption is exothermic in nature. The Liu isotherm model is intrinsically a mixture of both the Freundlich and Langmuir models; therefore, the monolayer hypothesis of the Langmuir model is eliminated and the infinite adsorption assumption specifically rooted from the Freundlich model is also overruled. This model predicts that all the reactive sites on the surfaces of the adsorbent have no similar energy. Hence, the adsorbent may have active sites preferred by the adsorbate molecules for occupation. However, saturation of the active sites should occur in contrast to what occurs in the Freundlich isotherm model. Considering various functional groups on the Fe₃O₄/AC nanoparticles, our results show that the active sites of the Fe₃O₄/AC will not naturally be able to share the same amounts of energy. Based on the Liu model assumptions, all reactive sites on the surface of the adsorbent have no similar energy. Hence, the adsorbent would have several reactive sites that tend to be occupied with the adsorbate molecules (Jafari et al. 2016).

As seen in Table 3, the values of n of the Freundlich model in the range of 1–10 is an appropriate index to prove the reliability of the adsorbent. Based on Table 3, n varies from 1.04 to 2.04 in all temperatures, showing the applied composite is a promising adsorbent for NP removal. The K_F of Freundlich also decreased from 162.3 to 68.56 L/g, which indicates that the adsorption process of NP is exothermic (Yang et al. 2011).

The magnitude of activation energy gives information about the type of adsorption. The physisorption process usually has energies in the range of 5–40 kJ/mol, while higher activation energies (40–800 kJ/mol) suggest chemisorption (Babaei et al. 2015). The activation energy was calculated to be 6.8 kJ/mol, indicating that the adsorption of NP onto the Fe₃O₄/AC was a physisorption process (Jafari et al. 2016).

According to the thermodynamic results, the obtained ΔG^o was negative in all temperatures indicating the spontaneous nature of the adsorption process (see Table 4). The negative value of ΔH^o (−38.1 KJ/mol) also showed that the process of NP adsorption onto Fe₃O₄/AC is exothermic. Enthalpy change value between 2.1 and 20.9 kJ/mol is frequently considered to indicate physical adsorption processes, whereas for chemical adsorption it lies in the range of 80–200 kJ/mol. In this work, the ΔH° value was found to be −38.1 kJ/mol, suggesting that the transportation of NP ions from the aqueous solution to the Fe₃O₄/AC composite surface occurred physically, which is consistent with the results obtained from the activation energy of the adsorption (Babaei et al. 2015). Additionally, from the negative value of ΔS^o, it can be postulated that a drop occurs in the randomness of the solid/surface interface at the internal structure of the NP adsorption onto the applied composite.

where X_min and X_max are minimum and maximum actual experimental values of data sets, respectively (Kıranşan et al. 2015).

During the training phase, each node received the input signals from input data or nodes of previous lines and biases, aggregated them by using the weights, and passed the result after suitable transformation as the output signal through a transfer function. Then the weights were adjusted by back-propagating the error calculated from the difference between the predicted and the experimental outputs (Esfahani et al. 2014b).

To find the optimal number of nodes in the hidden layer and the best possible weights and biases, the numbers of nodes were varied from 2 to 20. The optimum number of hidden nodes was determined by using mean square error (MSE) of the training and validation sets. According to the obtained results (Figure 5(b)), MSE decreased as the number of neurons was increased up to 12 and then increased. Therefore, input, hidden and output layers, respectively, with 4, 12 and 1 neurons were selected as the optimum ANN topography. The set of connection weights and biases that cause the optimum ANN topography are listed in Table 5.

Validity of the formulated ANN model was tested by comparing the model-predicted values of MB adsorption onto the Fe₃O₄/AC nanoparticles with those experimentally obtained (Figure 5(c)). Accordingly, the high correlation coefficient (R² = 0.986) proves the reliability of ANN to predict the adsorptive removal of MB.

The relative importance of each variable on the adsorption efficiency of NP has been reported in Table 6. All of the investigated variables strongly influence the NP adsorption onto the Fe₃O₄/AC composite and cannot be neglected, but the pH of aqueous solution plays the most pivotal role on the adsorption process.

In order to investigate the removal potential of NP using the Fe₃O₄/AC nanoparticles in real environments, the adsorption performance was evaluated in three samples, river water, tap water and wastewater effluent, under the optimized conditions. The percentage of NP adsorption by the Fe₃O₄/AC in the above samples containing 1 mg/L concentrations of NP are shown in Figure 6(b). Under these conditions, the adsorption efficiencies of 70.7, 73.5 and 67.3% were observed for the river water, the tap water and the wastewater effluent, respectively. According to Figure 6(b), there was no significant difference in the removal percentages of NP using Fe₃O₄/AC nanoparticles for the investigated samples. This means that other substances which are present in the river water or wastewater effluent do not have a noticeable influence on the adsorption percentage of NP onto the Fe₃O₄/AC nanoparticles. These results illustrate that Fe₃O₄/AC nanoparticles have a good potential for the removal of NP in real samples. Therefore, the removal of NP from full scale using Fe₃O₄/AC nanoparticles as a suitable adsorbent can be a cost–effective and useful technique.

Batch experiments were carried out to evaluate the performance of Fe₃O₄/AC grafted iron oxide nanoparticles on adsorptive removal of NP from solutions. Findings revealed fast NP removal at pH of 3 and 30 min contact time. A sharp increase in NP adsorption efficiency was observed by enhancing adsorbent dosage, whereas decreasing initial NP concentrations increased NP adsorption efficiency. In addition, the adsorption process of NP removal followed the Liu isotherm model and Avrami fractional order kinetic model. The negative values of thermodynamic parameters (i.e., ΔG^o and ΔH^o) showed that the contact of Fe₃O₄/AC nanoparticles with NP is spontaneous and exothermic. Furthermore, according to the results of ANN studies, pH of solution was the most effective input variable affecting NP adsorption efficiency, and initial NP concentration, contact time and adsorbent dosage were placed at the next steps, respectively. All in all, it can be found from this research that the as-synthesized composite is an efficient adsorbent that not only can remediate aqueous media from NP but also could be separated easily by means of a magnetic field.

The authors of this research gratefully acknowledge Ahvaz Jundishapur University of Medical Sciences for its academic and financial supports.