Artificial Neural Networks and Genetic Algorithms: An Efficient Modeling and Optimization Methodology for Degradation of Atenolol Using Activated Persulfate with Ultrasound

Atenolol (ATN) is a slowly biodegradable antagonist β -blocker drug and remains in the environment for a long period of time. Tis drug has a harmful efect on the environment and human and animal bodies. In this study, using activated persulfate with ultrasound for the degradation of ATN was investigated. Te efect of independent variables including pH, ATN concentration, persulfate dose, contact time, and ultrasonic power has been studied at 5 levels. Central composite design (CCD) was used for designing the experiments in Design Expert 11.0 software. Te ATN concentration was measured using high-performance liquid chromatography (HPLC). Genetic algorithms (GA) and artifcial neural network (ANN) were used for optimization and prediction, respectively. Te results indicated that at the optimal conditions for the ex-periment (pH of 6.79, reaction time of 19 min, initial ATN concentration of 15.92 mg/L, US power of 109.56 W, and PS dose of 1317.88 mg/L), the highest ATN degradation efciency was 98.9%. Te ATN degradation could be represented by the pseudo-zero-order kinetics. Also, the data of ATN were well ftted with the ANN model ( R 2 � 0.98). Te results showed that the best pH range to eliminate ATN is the near neutral range and the GA was found to be an efective tool to optimize the experimental conditions for the removal of ATN. Te ultrasonic/persulfate process as a useful technique has a high potential to remove ATN from aqueous solutions.


Introduction
In recent decades, a group of emerging pollutants, pharmaceutical and personal care products (PPCPs), and medical compounds, due to worldwide use and complex sustainable structures, are one of the major environmental problems, which have attracted increasing attention [1,2]. Atenolol (ATN) is a representative PPCP compound and one kind of the most prescribed antagonist β-blocker drugs that has been used for nearly 40 years, commonly used to treat hypertension, arrhythmia, glaucoma, and angina, and against coronary heart diseases [1,3,4].
Tis beta-blocker drug is resistant to natural attenuation and also biodegrades very slowly. Terefore, it enters surface and ground waters and fnally remains in the environment for a long period of time. Tis drug has also harmful consequences in animals and aquatic organisms such as the efect on heart rates and behavioral disorders in some invertebrates and vertebrates [2,5].
ATN has also been shown to inhibit the growth of embryonic cells and can also afect the level of testosterone in humans by afecting the endocrine glands [1,6].
Studies have shown that the toxicity of ATN to aquatic organisms and humans is the most serious among other bblockers, and temperature afects the rate of degradation of ATN, indicating higher stability at 4°C. Furthermore, oral doses of ATN are excreted mostly (about 90%) unaltered through urine because it is not fully metabolized by the human body [7][8][9].
ATN extensively exists in a variety of aqueous environments, such as hospital wastewater, groundwater, and surface water [2,6]. According to previous studies, conventional wastewater treatment plants (WWTPs) cannot remove β-blockers efectively, so beta-blockers in WWTP are often easily discharged into surrounding rivers and lakes. For this reason, more efcient and advanced water treatment processes are needed to remove these compounds [5].
Advanced oxidation processes (AOPs) are generally defned as oxidation methods and in the last decade have received special attention due to being economical and having high efciency. AOPs are primarily based on producing high oxidizing radicals such as OH•, O•, and SO 4 • that can remove a large range of organic pollutants [16,17].
Key AOPs known as produce active radical processes include the use of strong oxidizing agents such as hydrogen peroxide (H 2 O 2 ), ozone (O 3 ), Fenton, photo-Fenton, photocatalysts (iron ions, electrodes, and metal oxides), radiation (UV light, sunlight, and ultrasound), and electrochemical oxidation that alone or in combination are efective in eliminating volatile organic matters and odor control [18,19].
Among the various AOPs, ultrasound (US) is more popular for wastewater treatment due to the lack of production of intermediate pollutants, the activation of persulfate ion, and production of free radicals [20].
US treatment, which produces radical hydroxyl and oxygen due to microbubbles through the cavitation phenomenon in the solution, is an efcient method for destroying organic pollutants [20,21].
Persulfate (PS) has desirable properties including easy storage and transportation, high stability, rapid reaction, and cost-efectiveness, and it is one of the strongest oxidants compared to others.
Te production of sulfate radicals and subsequent hydroxyl radicals by persulfate is shown in the following equation [22]: Advanced design of experiments (DOE) techniques help us better understand and optimize the response. Tere are two main types of experimental designs, including central composite designs and Box-Behnken designs. A central composite design is the most commonly used design of experiments method. Central composite designs are a factorial or fractional factorial design with center points, augmented with a group of axial points (also called star points) that let you estimate curvature and nonlinear relationships. CCD is a more useful methodology for modeling various technological processes in the case of small number of preformed experiments compared with "one variable at a time" approach [23].
Today, one of the essential stages in the advanced oxidation processes is to model and optimize the parameters those are efcient in the process [24]. Nowadays, one of the most signifcant prediction and optimization methods is the use of artifcial intelligent models and evolutionary algorithms, such as artifcial neural networks (ANNs) and genetic algorithm (GA), which have been applied in many studies [25].
ANNs are very efcient tools for nonlinear statistical data modeling because they use learning algorithms that can independently make predictions as they receive new input [26]. GA is a method for optimization and solving problems based on a natural selection process that imitates biological evolution [27]. Tis method could be applied to solve problems that are not well suited for standard optimization algorithms, including problems in which the objective function is stochastic, nondiferentiable, highly nonlinear, or discontinuous [25].
Considering the adverse efects of atenolol and its impact on human health, the aim of this study was to remove ATN using a persulfate/ultrasonic process as an efective, simple, and environmentally friendly method. In addition, the efect of fve quantitative efective variables including persulfate concentration, ultrasonic power, initial concentration of ATN, pH, and reaction time on the ATN removal efciencies was evaluated. Also, the experimental data were modeled and optimized through ANN and GA, respectively. In addition, to the best of the authors' knowledge, this type of analysis for ATN has been reported in no previous literature. − ), and humic acid (C 9 H 9 NO 6 ) were supplied from Sigma-Aldrich (USA). Acetonitrile (HPLC grade; C 2 H 3 N ≥ 99%) and ethanol (HPLC grade; C 2 H 5 OH ≥ 99/9%) were purchased from Merck Co. (Germany). Hydrochloric acid (HCl) and sodium hydroxide (NaOH) were used to adjust the pH. All of the materials were of analytical grade and used as received without further purifcation. Te deionized water was used to dilute the aqueous solutions.

Analytical Methods.
Te concentration of ATN in samples was determined by using high-performance liquid chromatography (HPLC), Waters Company (USA), with a UV-2075 detector at 231 nm and C18 column. Te mobile phase was acetonitrile/potassium dihydrogen phosphate with 70/30 (v/v) ratio at the fow rate of 1 mL/min. 20 µL of the sample was injected while the column temperature was set at 25°C. Te removal efciency of ATN was calculated according to the following equation: where C 0 is the initial ATN concentrations in the solution (mg/L) and C e is equilibrium ATN concentrations after degradation (mg/L). An ultrasonic homogenizer (Bandelin-UW 3200, Germany) was used to produce ultrasonic waves for the degradation of the experiments. Te specifcations of the ultrasonic device used in this study are given in Table 1.

Design of Experiments.
In this research, the CCD was applied for the design of the experiments, using Design Expert 11.0 software, considering 5 levels for each variable. Finally, a matrix with 32 runs was suggested by the software for this study. Te suggestion of the software for the number of experiments was based on 16 cube points, 6 center points in cube, and 10 axial points. Table 2 shows the variables and levels considered.

Kinetics Analysis.
Te study of the degradation kinetics of atenolol was performed under the optimum condition. Te efect of scavenger pollutions, chloride, nitrate, and humic acid on ATN degradation was also investigated. Te analysis of humic acid content was performed using UV adsorption at 254 nm. Chloride and nitrate ions were also quantifed [28].

Modeling and Optimization Details
2.6.1. Artifcial Neural Network. Te ANN was used to determine nonlinear relations between variables. MATLAB software R2017a version was used with the neural network toolbox (nnool) for the ANN modeling. A three-layer feedforward backpropagation neural network was developed by applying the tangent sigmoid (tansig) and linear (purlin) transfer functions in hidden and output layers, respectively, with the Levenberg-Marquardt training algorithm. Te ANN model comprised fve input parameters which include initial concentration, pH, PS, US, and time. Te output layer was composed of one neuron representing removal efciency. Te 32 obtained samples were randomly divided into three categories including training (70%), validation (15%), and test (15%) subsets.
Te response of a neural network as well as the prevention of overftting was signifcantly afected by the number of neurons in hidden layers, and thus, the number of neurons in the hidden layer was obtained by trial and error within the interface of the following equations: where i, o, n, and n h are the number of input neurons, output neurons, datasets, and the hidden layer neurons, respectively [29]. Te accuracy of the optimized ANN model was calculated using the determination coefcient (R 2 ) and also the mean square errors (MSE) (equations (5) and (6)) [26]: where y i and y i-pred are the experimental and predicted values obtained from ANN, N is the number of experiences, and y i-av is the average of the experimental values.

Genetic Algorithm Optimization.
Te GA was implemented in the MATLAB software R2017a version. GA was used to determine the optimal conditions for ATN removal using the ANN model developed in this study as ftness function as follows: ftness function � Max (removal rate from ANN simulation).  Other settings intended for the genetic algorithm are number of generation 250, the rank scaling function, the selection function of Stochastic uniform, the number of elite 2, crossover fraction equal to 0.8, and the mutation function of constraint dependent and combination of scattered function [30].

Modeling with ANN.
To determine the best neural network topology with the mentioned specifcations, several networks with diferent numbers of hidden neurons were developed, and fnally, the 5 : 5 :1 topology showed the best performance. Figure 1(a) shows the structure of a three-layer neural network that has fve parameters in the input layer, fve neurons in the hidden layer, and one parameter in the output layer. Figure 1(b) represents the training, validation, and testing errors against the number of epochs (iterations) for the ANN models. Usually, the error decreases after some epochs of training and the training algorithm terminates if the validation error increases after six sequence iterations, or the limitations of the maximum error/epoch are exceeded. As demonstrated in Figure 1(b), the training algorithms for ANN are terminated in 2 epochs (because of the validation error), while the training errors are small enough. Te test and validation error plots demonstrate similar characteristics. Tis is an indication of appropriately divided data in this study. Te MSE values of the validation step are equal to 0.001044.
In Figure 1(c), the regression plots reveal the network outputs regarding targets for training, validation, and test groups. For a perfect ft, the data should fall along a 45degree line, where the network outputs are equal to the targets. It is observed that the output tracks the targets very well for training (0.982), validation (0.999), and testing (0.968). Also, the coefcient of determination for the whole data was obtained 0.982. In this study, the network response is satisfactory and simulation could be used for entering new inputs. What is clear from the ANN results is that no overftting has occurred in the developed network [31]. In Figure 1(d), a comparison diagram of the experimental and predicted results is presented.

Te Efect of Infuencing Variables on ATN Degradation.
Using the designed neural network, the diagrams ( Figure 2) were drawn in MATLAB software. Te parameters shown inside each graph were considered variable, and the other parameters for each graph were fxed and equal to its optimal value (obtained from the genetic algorithm). Figure 2(a) indicates that the initial concentration of ATN has a nonlinear correlation with the removal efcacy. Te removal efciency of ATN improved with an increase in the initial ATN concentration from 1 to 15 mg/L which shows the optimal concentration; then after that, increasing the concentration reduces the removal efciency. In my opinion, when the initial concentration decreases in the solution, the number of ATN molecules among the water molecules is low and hidden between them, so ultrasonic waves have no signifcant efect on the ATN molecules and they reduce the removal efciency.

Efect of Initial Concentration of ATN.
On the other side, Aravindakumar et al. believed that increasing the initial concentration of ATN and also increasing the pH due to the recombination of OH• radicals and the existence of byproducts of ATN oxidation that compete with ATN molecules in the consumption of OH• cause a lack of OH• to react with atenolol and these factors reduce the removal efciency [32].
Dehdashti et al. who studied on atenolol removal using multiwalled carbon nanotubes concluded that increasing the initial concentration of ATN has decreased the removal efciency, and the reason is that increasing the initial concentrations of atenolol causes the rapid saturation of limited adsorption sites on an adsorbent [33].
Persulfate also had an almost nonlinear relationship. However, at a higher persulfate concentration, higher removal efcacy was obtained. It may be because OH • radicals' formation is almost steady on the higher dose of PS and that they become the restricting reactants. Moreover, the presence of further quantity of SO 4 -gives higher degrees of oxidizing species. However, on the other hand, large amounts of PS have a negative efect on ATN degradation efciency and show an inhibition efect because SO − 4 radicals react with themselves and produce S 2 O 8 as shown in the following equation [34]: A recent study on ATN removal by heat-activated persulfate found the same results. It says that the ATN degradation rates well followed pseudo-frst-order kinetics under diferent PS dosages [35].

Efect of pH.
Te pH of solution is a signifcant factor that afects the type and amount of produced free radicals, and the rate of decomposition of target compounds in the advanced oxidation process is based on the persulfate [22]. Te efect of diferent ranges of pH solution and ATN concentration on the degradation of ATN and other parameters which have been kept at optimal condition was studied in this research. Figure 2(b) shows nonlinear association between pH and the removal rate. As seen, with the rising of pH from 3 to 7 and also increasing of ATN concentration, the removal efciency of ATN was increased to more than 90%. However, after this point, increasing pH and ATN concentration reduced the removal efciency. Te reaction of activated sulfate with ultrasonic waves was investigated in 3 ranges of acidic (pH < 7), neutral (pH = 7), and alkaline (pH > 7). In each range, some reactions occur and radicals are produced that afect the mechanism of PS oxidation.
Salehi et al. conducted a study on advanced oxidation which showed that in acidic pH, the oxidation efciency increased within the UV/K 2 S 2 O 8 process so that with decrease in the pH, persulfate is decomposed into sulfate free radicals and can be further catalyzed by acid (equations (8) and (9)  Journal of Environmental and Public Health In a study conducted by Harati and Rezaei Kalantary, when the pH is in the neutral range (pH � 7), hydroxyl radicals are produced which perform higher oxidation due to their longer half-life than sulfate radicals (as shown in equations (10)- (12)). On the other hand, in these conditions, there are other adverse reactions that lead to the consumption of hydroxyl radicals in solution. In equations (13) and (14), sulfate radicals can enter the reactions selectively [37].
Te study of Gao et (15), increasing the ultrasonic waves converts the sulfate anion into a sulfate radical which is responsible for the degradation of organic matter in an acidic environment, thus increasing the degradation of atenolol [17,20].
Ultrasonic waves can also produce hydroxyl radicals according to equations (16)- (18): OH . + OH . ⟶ H 2 O 2 (17) Razmi et al. studied the optimization of phenol removal from wastewater by activation of persulfate and ultrasonic waves in the presence of biochar catalyst modifed by lanthanum chloride. Ultrasonic waves cause the formation and breakage of gas bubbles and produce hydroxyl radicals and hydrogen that can react with phenol and lead to its breakdown [38].
Te study of Liu et al. on the removal of ketoprofen from water by sono-activated persulfate oxidation shows that greater ultrasonic power input can also produce more reactive radicals in the US/PS system. It was the same as fndings of this study [21].

Efect of Contact Time.
Another factor infuencing the speed of reactions is contact time. Figure 2(d) shows the time versus the pH in ATN degradation. Te efect of free radicals on ATN is reduced by the formation of compounds resulting from the breakdown of ATN. On the other hand, at lower contact time due to the lack of maximum formation of free radicals by ultrasonic waves and also the participation and use of the radical sulfate made in the process at the initial time, the efciency is less than the optimal state [39]. Movahedian Attar et al. reported similar results to our study that the optimum conditions were obtained at 17.5-18.5 mg/L and 60-65 min for 4-chlorophenol concentration and contact time, respectively, with the removal rate of 85.74%. However, by increasing contact time and 4chlorophenol concentration, the removal efciency has been declined. It could be because when the contact time increases more than the optimal time, it prevents hydroxyl radical accumulation because self-scavenging reaction may occur by nonradicals with persulfate. It says that the ATN degradation rates well followed pseudo-frst-order kinetics under diferent PS dosage anions [22].
Dehdashti et al. in the study on using multiwalled carbon nanotubes for ATN removal showed that with extension of contact time from 5 to 60 minutes, an increasing trend was observed in the adsorption process. Ten, a downward trend occurred until 90 minutes. Te reason could be the fast flling of adsorption active sites by the adsorbate at the initial times [40].

Degradation Kinetics Modeling of ATN.
Te study of kinetics is necessary to evaluate the degradation rate and behavior of ATN molecules as a function of reaction time. In Table 3, the equations of pseudo-zero-, pseudo-frst-, and pseudo-second-order kinetic models used in this study are presented. Kinetic plots ftted to the results of ATN removal efciency by ultrasonic/persulfate are shown in Figure 3. As the constant of the degradation rate (k) and the correlation coefcients (R 2 ) are shown in Table 4, degradation of ATN is described better by pseudo-zero-order (R 2 > 0.9197) than other kinetic models. Tian et al. studied about synergic effectiveness of US/PS on diesel hydrocarbons (DHC) degradation from soil. Tey have also concluded that the pseudozero-order is more suitable for DHC degradation [41]. However, Li et al. investigated the treatment of soil contaminated with total petroleum hydrocarbons (TPH) using activated persulfate oxidation, ultrasound, and heat. Tey revealed that the pseudo-second-order reaction better evaluates the degradation of the TPH process by activated PS [42].  Figure 4(a) reveals the best and average ftness values in each generation. Figure 4(b) shows the best ftness values in the fnal generation. Hence, the optimal value of the input parameters, which are pH, time, US, PS, and ATN, was equal to 6.79, 19.0 min, 109.56 W, 1317.88 mg/L, and 15.92 mg/L, respectively. It leads to the highest removal efciency of 98.92%. Table 5 compares the results of this study with some similar studies on removal efciency. As can be seen in Table 5, there are various methods for removing pollutants from aqueous solutions. Some of the benefts of ultrasonication are safety, high efciency, energy conservation, and cleanliness of the process [32]. Low ability to decompose organic matter at room temperature, cheapness, and high radical stability produced in diferent situations [17]. are the advantages of persulfate. Terefore, using ultrasonic/persulfate combination is more efective than other methods for ATN removal.
3.5. Sensitivity Analysis. Sensitivity analysis was conducted by the Pearson correlation method to evaluate which input independent parameter had the greatest efect on the removal efciency. Tis analysis was performed separately for each of the input variables in a way that one input parameter was changed at equal intervals of a given point while the other input variables were kept constant [45].      Journal of Environmental and Public Health Figure 5 indicates the comparison between the impacts of input parameters along with the coefcient efect of each parameter. According to Figure 5, the initial concentration of ATN with an impact factor of 0.76 is the most infuential parameter on ATN removal and time has the least impact. Te efect of the parameters is as follows: ATN > pH > PS > US > t. 3.6. Efect of HA, Nitrate, and Chloride on the Degradation of ATN. Nitrate (NO 3 − ), chloride (Cl − ), and natural dissolved organic matter (NDOM) such as HA are widespread in ground or surface waters [46]. Te presence of these materials is expected to interfere with the degradation of ATN by PS [35]. As shown in Figure 6, the degradation rate of ATN became slow with the presence of the scavengers. HA had the greatest efect among the scavengers, due to the fact that with increasing the concentration of HA from 0 to 8 mg/L, the degradation of atenolol decreased from 99% to 45.58%. As Lei et al. reported, inhibition of ATN degradation is related to competing between the ATN molecules and HA that during the reactions, HA binds to the OH − and SO 4 2− free radicals and reduces the consumption of radicals by ATN [41]. Cl − and NO 3 − are another materials and their inhibitory efects on diferent AOPs have been reported. Teir mechanism of action in the restriction is the same as HA, but their impact is less than it [35,46]. Liao et al. fgured out that the inhibitory efects could be attributed to the consumption of free radicals (OH and SO 4 − ) by Cl − and formation of less reactive inorganic radicals [47]:

Conclusion
Te results demonstrated that in the optimum condition of the ultrasonic/persulfate process, ATN removal of 98.92% was achieved. Degradation of ATN is described better by pseudo-zero-order (R 2 > 0.9197). ATN can be eliminated efectively at near neutral pH. Te scavenging experiments confrmed that hydroxyl and sulfate radicals play a fundamental function in the degradation process. Te outcomes confrmed that the background substances could have a negative impact on the elimination performance as HA > Cl − > NO 3 − . Te ANN model predicted the degradation efciency with high accuracy (R 2 : 98.2%).

Data Availability
Te data generated and analyzed during this study are available from the corresponding author upon request.

Ethical Approval
Te ethics code of this research is IR.MUI.RESEARCH.REC. 1398.649.

Conflicts of Interest
Te authors declare that they have no conficts of interest.

Authors' Contributions
Zeynab Moradmand performed investigation and data curation and wrote the original draft. Bahare Dehdashti contributed to the concept, conducted the study, and wrote the original draft. Farzaneh Mohammadi was responsible for investigation, modeling, and revision of the manuscript. Nasrin Zahedi performed investigation and revised the manuscript. Maryam Razaghi was responsible for conducting the study and revising the manuscript. Mohammad Mehdi Amin contributed to conceptualization, supervision, methodology, writing, reviewing, and editing.