Artificial neural network approach for lead removal from aqueous solution by agricultural waste deri

Abstract

Keywords: Adsorption; Artificial Neural Network; Biochar; Lead Ions (Pb2+); Modeling

Introduction

Lead is one of the oldest known heavy metals and is still used in various industries like battery production, metal plating, mining, pigments manufacturing, fuels, ceramic, and glass production due to its high density, low melting temperature, and corrosion resistance. On the other hand, due to its high solubility in aquatic environments, lead is a well-known major pollutant [1]. Lead is a toxic element to living organisms, and not necessary for physiological function. It can cause many diseases, and deaths even in low concentrations in water. Carcinogenic, and neurological damage, mental, and behavioral changes are associated with lead toxicity. It has negative effects on immune system, and reproductive system. Stomach, lung, and gallbladder cancers as well as increased incidence of all cancers were considered to be are associated with lead exposure [2]. According to the World Health Organization (WHO), the maximum permissible limit (MPL) of lead in wastewater treatment effluents is 10 µg/L. On the other hand, the permissible limit of Pb(II) which is given by the United States Environmental Protection Agency (USEPA) is 6 µg/L [3]. In the removal of lead from wastewater, some traditional methods such as precipitation, coagulation, ion exchange, reverse osmosis, electro-dialysis, chemical reduction, oxidation can be used [4]. With these traditional methods, heavy metals may not be completely removed. Apart from that, these techniques have some disadvantages; expensive equipment, and tracking systems requirement, excessive chemical, and energy needs, formation of toxic sludge, and other wastes. As an alternative to these methods, adsorption is a technique that has been successful in heavy metal removal with low cost, and high efficiency in many studies. In recent years, natural substances have been increasingly used as adsorbents in this process because of their low cost, and easy modification. Even though the availability of various commercial adsorbents, their widespread use is often restricted by the high cost, and limited availability. Therefore, researchers have been trying to develop alternative low-cost but efficient adsorbents using agricultural, and industrial wastes especially [5, 6]. Biochar, a carbon-rich byproduct synthesized through the carbonization of biomass in an oxygen-limited environment, has received more attention from researchers in recent years, and has been recognized as a multifunctional, and low-cost material for environmental applications due to its high absorptive ability. Many organic wastes such as agricultural, industrial, urban, and domestic solid waste can be used as raw materials to produce biochar. Carbonaceous materials are the best-known adsorbents that can be used to remove both organic, and inorganic contaminants from the aquatic environment. Therefore, biochar has also attracted worldwide attention as a useful, low-cost, environmentally friendly adsorbent for various pollutant reclamation due

to its large surface area, high adsorption capacity, microporosity, and ion exchange capacity [7]. Adsorption is a process that contains non-linear relations involved in complex chemical processes due to physicochemical characteristics, and the particular structure of adsorbent. The effect of adsorption parameters on the adsorption process is estimated by modeling studies, parameter values can be estimated to ensure high efficiency of heavy metal removal from existing wastewater [8]. Artificial neural network (ANN) is widely used for prediction in many industrial areas for the last 10 years and is an artificial intelligence method that attracts attention in wastewater treatment [9-12]. ANN approach is not dependent only on data, it can evaluate biased data, and even learn complex relationships, generalize, and answer questions it has never encountered before with an acceptable error [13]. There are modeling studies to remove lead ions from an aqueous solutions by using different adsorbents in the literature. Yetilmezsoy, and Demirel used ANN approach for modeling the % Pb2+ removal by Antep pistachio (Pistacia Vera L.) shells with R2 0.936 [14]. Khan et al. used ANN for the adsorption of Pb2+ ions by using rice husk carbon (RHC). Mean square error (MSE), and R2 values were obtained as 6.005, and 0.989, respectively with Levenberg– Marquardt algorithm (LMA) [15]. Sadrzadeh et al. used ANN approach to predict separation percent (SP) of lead ions from wastewater by using electrodialysis process. The modeling results showed that there is an excellent agreement between the experimental data, and the predicted values [16]. Ma et al. used ANN for Pb2+ removal from water by using Scenedesmus obliquus which achieved the highest Pb2+ removal efficiency at pH 6.0, and algal biomass 2 g/L. Testing model results showed that the correlation between the predicted data, and the experimental data was 0.997, with an accuracy rate of 95.4% [17]. Kardam et al. used ANN approach for the abatement of lead ions from an aqueous solution by using nanocellulose fibers (NCFs) derived from rice straw. Comparison between the network results, and experimental data gave a high degree of correlation coefficient (R2 = 0.995) indicating that the model is able to predict the sorption efficiency successfully [18]. Afroozah et al. used ANN, Fuzzy Inference System (FIS), and Adaptive Neuro-Fuzzy Inference System (ANFIS) for the prediction of Pb2+ removal from an aqueous solution using the magnetic nano adsorbent (Graphene/Nylon 6). ANFIS showed better performance than the other models [19]. Amiri et al. used ANN, and MNLR (Multivariate Non-Linear Regression) approach for Pb2+ removal from industrial wastewaters by ostrich bone char. ANN model showed good predicting performance with MSE=0.0002, and R2=0.98. Additionally, this study indicated characterisation of flexibility, and extensibility of ANN model [20]. In this study, ANN modeling of Pb2+ removal from an aqueous solutions has been carried out by using two

Sigma J Eng Nat Sci, Vol. 41, No. 1, pp. 64–73, February, 2023

different biochars produced from hazelnut shell, and walnut shell as an adsorbent. This work is a continuation of the work done by Kaya et al. (2020) [21]. % lead removal was predicted through ANN, and the results compared with experimental values. To create the optimum network architecture, ANN parameter selections were made, and predictive ability of the network was evaluated by R2, and MSE value.

Materials And Method

Preparation of Biochars Derived from Hazelnut and Walnut Shells In this study, hazelnut, and walnut shells which are widely available lignocellulosic agricultural wastes were carbonized in an electrically heated fixed bed pyrolysis reactor. Pyrolysis temperature was set to 500°C, and 700°C for hazelnut, and walnut shells, respectively. During the process, the inert nitrogen gas flow rate was set to 100 mL/ min, and the heating rate to 20°C/min. During the pyrolysis process, chemical bonds of biomasses were thermally degraded in an oxygen-free environment at high temperatures, and the end of 1 h, biochars with high surface area were obtained [21]. Adsorption Studies To determine the adsorption properties of biochars produced by pyrolysis of hazelnut, and walnut shells for the removal of lead ions (Pb2+) from an aqueous solutions, batch adsorption experiments were carried out. Commercially available stock lead nitrate solution (Pb(NO3)2) with a concentration of 1,000 mg/L for lead metal were diluted with distilled water to the desired concentration to prepare the test solutions to be used in the adsorption experiments. Parameters affecting the adsorption process such as pH (2.5-5), adsorbent dosage (1-3 g/L), initial heavy metal concentration (15-45 mg/L), contact time (up to 300 min), temperature (25-45°C), and mixing speed (200-600 rpm) were studied in a batch system. The samples were collected at predetermined time intervals, and adsorbent separated from the samples by filtering. The filtrate was analyzed by ICP-OES (Perkin Elmer, Optima 5300 DV) to determine the residual lead ion concentration. Based on the acquired values, the adsorption yield was calculated by using Equation (1). Lead Removal (%) =

Model theory The first artificial neuron was produced in 1943 by neuropsychiatrist Warren McCulloch, and scientist Walter Pits. However, due to the limited possibilities of the period, there was not much progress in this area. After that, Minsky, and Papert published a book in 1969, eliminating ethical concerns in the field of ANN, and paving the way to this new technology. Hence, the first visible developments in this area was date back to the 1990s [22]. ANN performs operations with cells called neurons and is an improved network modeled on the way the human brain works. ANN can provide linear, and nonlinear modeling between input, and output variables without any prior knowledge. Although there are many studies by many researchers about ANN there is no certainty that what are the key factors affecting the ANN’s performance [23]. ANNs are data-based systems created by connecting artificial nerve cells in layers. ANN aims to use the skills of the human brain, such as learning, and making very fast decisions under different conditions, in solving complex problems with the help of simplified models. Basically, all ANNs have a similar structure as shown in Figure 1 [24, 25]. where x (xi , i = 1, 2, …, n) are inputs to the neuron, and Y is output. Each input is multiplied by its weight wi, a bias b is associated with each neuron, and their sum goes through a transfer function f. As a result, the relationship between input, and output can be described as follows from Equation (2) [26]: Y = f (∑ i =1 wi xi + b) n

In this structure, some neurons are connected to the outer space to receive inputs, and to transmit outputs. All the remaining neurons are in hidden layers, meaning they only have connections within the network. While the input, and output layers consist of a single layer, there may be more than one hidden layer between these two layers. These hidden layers contain a large number of neurons, and these neurons are completely connected with other neurons in the network. Network architecture, learning

where Co, and Ce are the initial, and final lead concentration in solution phase (mg/L), respectively. In order to ensure the reproducibility of the results, all the adsorption experiments were performed in triplicate, and the average values were used in data analysis. Relative standard deviations were found to be within ± 1% [21].

Sigma J Eng Nat Sci, Vol. 41, No. 1, pp. 64–73, February, 2023

function, and training function are the three main component of ANN. There are no rules or mathematical algorithms for determining the hidden layer, and the number of hidden neurons in this layer. The decision is made by applying the method of trial, and error. Independent variables affecting the system is chosen as an input, and the variable that is tried to be predicted is selected as the output variable [27].

depends on what the network is asked to learn. Commonly used activation functions are sigmoid, and hyperbolic tangent functions. For example, if the network desired to learn the behavior of a model, the sigmoid function can be used or if it desired to learn the deviation, the hyperbolic tangent function can be used [35].

Data division ANN architecture contains data division step, and data set divided into 3 subsets: a training set, a testing set, and a validation set [28]. The data set used for training is called “training set”. Throughout the training, the same training set weights are given to the network over, and over to bring it to the most appropriate level. The word epoch is expressed as showing the entire training set to the network. After completion of training step, validation set how much the network has learned in each iteration, and they use the error values of the validation set. If the desired success is achieved with the data in the training set, the network is tested with data. The data set in which the test is performed is called the “test set” [29]. Training function Learning in a neural network means finding optimum values of all weights in ANN that will enable the network to produce correct outputs between input, and output data for a particular problem. The operation done to reach optimal weights is called “training the network”. Accordingly, in order for a network to be trainable, the weight values must be dynamically changeable within a certain rule. Typically, a training algorithm is used for the learning process, and it determines how the weights are arranged according to a learning sample. Levenberg-Marquardt (LM) method, Bayesian Regulation (BR) method, and Scaled Conjugate Gradient (SCG) algorithm are type of learning algorithms. The Levenberg-Marquardt (LM) algorithm is an approach to the Newton method and is the combination of the speed of the Newton method, and the robustness of the stepdown method [30-32]. Activation function The activation function establishes the link between input, and output. It creates output information by processing the information coming from the summation function. Like the summation function, this function has various functions depending on the ANN model to be applied [33]. Different activation functions are as follows: linear function, sigmoid function, hyperbolic tangent function, sine function, digit function. Depending on the activation function used, the output value is usually [-1,1] or [0,1]. This is usually a nonlinear function. The use of non-linear activation functions has enabled ANNs to be applied to complex, and very different problems [34]. The choice of activation function depends largely on the data of the ANN, and

Performance evaluation criteria Mean square error (MSE) as a performance function, and coefficient of correlation (R2) are used to evaluate models predicted results. They are shown in Equation 3, and 4 how MSE, and R2 are calculated. The best model performance is the structure in which R2 value approaches 1, and MSE approaches 0. 1 n ∑ ( Xt − Xt′)2 n i =1

where, Xt, and Xtʹ are the actual, and predicted values, and n is the total number of observations [36].

Results And Discussion

In this study, experimental results for the Pb2+ removal from an aqueous solution by biochars that produced by pyrolyzing of hazelnut, and walnut shell were modeled with ANN approach. Levenberg–Marquardt feed-forward algorithm was used. The ANN were developed in the Neural Fitting toolbox, Matlab R2017a. ANN structure of the adsorption process was given in Figure 2. pH, temperature, initial Pb2+ concentration, adsorbent amount, and mixing speed were used as input for the

Sigma J Eng Nat Sci, Vol. 41, No. 1, pp. 64–73, February, 2023

system to estimate the removal efficiency of lead ions. One hidden layer was used, and different numbers of neuron combinations were in this layer between 1-25. Optimum value was selected according to maximize of R2, and reduction of MSE model performance criteria. Levenberg-Marquardt (LM) training function which is known that it is a fast algorithm was used in training the network. This algorithm frequently used in feed-forward networks [37]. Also, ANN models were developed and trained with two different training algorithms including BR and SCG, and have been evaluated on the basis of their predictive capability.

Tanjant sigmoid (tansig) function was preferred as an activation function in this study. The most common problem encountered during ANN training is the overfitting problem. Once the network has been trained with a sufficient number of cycles, the error can be reduced to reasonable levels but after the network is fed with test data, very high error values are encountered. This indicates that the network memorizes the training data, and weights for the values appropriate to it that it has updated but has not reached the ability to generalize. Therefore, data division ratio is important to solve the overfitting problem. In this study when data divided into 70:15:15,

Figure 3. ANN toolbox for biochar derived from walnut shell.

Sigma J Eng Nat Sci, Vol. 41, No. 1, pp. 64–73, February, 2023

Figure 4. ANN regression plots for biochar derived from walnut shell.

Sigma J Eng Nat Sci, Vol. 41, No. 1, pp. 64–73, February, 2023

Figure 5. ANN toolbox for biochar derived from hazelnut shell.

and 60:20:20, R2 value was 1 but MSE values were too high. That’s why this data division ratio was changed to 50:25:25. The number of experimental measurement results used for training, validation, and testing were 6:4:4. ANN Modeling of Pb Removal by Using Biochar Derived from Walnut Shell 5 inputs (pH, temperature, initial Pb2+ concentration, adsorbent amount, and mixing speed) were defined in the system, and number of neurons in hidden layer was found to be 11 by trial, and error as seen from Figure 3. No 2+

significant change occurred on the method’s performance after 5 epochs. Best validation performance was MSE 0.011 for trainig, 0.0063 for validation, and 0.0095 for testing steps; R2 was 0.97 for all of the system (Figure 4). ANN Modeling of Pb2+ Removal by Using Biochar Derived from Hazelnut Shell 5 inputs (pH, temperature, initial Pb2+ concentration, adsorbent amount, and mixing speed) were defined in the system, and number of neurons in hidden layer was found to be 6 by trial, and error as seen from Figure 5. Maksimum epoch number was found to be 12. Best validation

Sigma J Eng Nat Sci, Vol. 41, No. 1, pp. 64–73, February, 2023

Figure 6. ANN regression plots for biochar derived from hazelnut shell.

performance was MSE 0.051 for trainig, 0.0015 for validation, and 0.0024 for testing steps; R2 was 0.98 for all of the system (Figure 6). ANN models were developed and trained with BR and SCG training algorithms, additionally. From this computation, LM has been recognized as the best training algorithm for Pb2+ removal efficiency predictions as seen from Figure 7.

As a result, LM algorithm provided better generalisation performance compared to the other algorithms due to its high R2 value and showed closer results to the experimentally measured. Also it has been observed that an increase in the number of maximum epoch required for training step. When precision of predictive ability was measured for each training algorithm, their performances

Sigma J Eng Nat Sci, Vol. 41, No. 1, pp. 64–73, February, 2023

References

1. Arbabi M, Hemati S, Amiri M. Removal of lead ions from industrial wastewater: A review of Removal methods. Int J Epidemiol Res 2015;2:105–109.
2. Flora G, Gupta D, Tiwari A. Toxicity of lead: A review with recent updates. Interdiscip Toxicol 2012;5:47–
3. Vidu R, Matei E, Predescu AM, Alhalaili B, Pantilimon C, Tarcea C, et al. Removal of heavy metals from wastewaters: A challenge from cur- rent treatment methods to nanotechnology applica- Figure 7. Comparative study results of LM, BR and SCG tions. Toxics 2020;8: 101. [CrossRef] methods. [4] Gunatilake SK. Methods of removing heavy metals from industrial wastewater. J Multidiscip Eng Sci Studi 2015;1:12–18.
4. Burakov AE, Galunin EV, Burakova IV, Kucherova AE, Agarwal S, Tkachev AG, et al. Adsorption of were determined as LM > BR > SCG for Pb2+ removal heavy metals on conventional and nanostructured prediction. materials for wastewater treatment purposes: A review. Ecotoxicol Environ Saf 2018;148:702–712. [CrossRef] CONCLUSIONS
5. Ozkan G, Akin BA. The prediction of chemical The authors would like to thank Hitit University for its oxygen demand (COD) or suspended solid (SS) support. removal using statistical methods and the artifical neural network in the sugar industrial wastewaters. CONFLICT OF INTEREST J Eng Appl Sci 2013;8:978–983.
6. Deniz ST, Ozkan P, Ozkan G. The accuracy of The author declared no potential conflicts of interest the prediction models for surface roughness and with respect to the research, authorship, and/or publication micro hardness of denture teeth. Dent Mater J of this article. 2019;38:1012–1018. [CrossRef] [13 Mjalli FS, Al-Asheh S, Alfadala HE. Use of arti- ETHICS ficial neural network black-box modeling for the prediction of wastewater treatment plants per- There are no ethical issues with the publication of this formance. J Environ Manage 2007;83:329–338. manuscript. [CrossRef] Sigma J Eng Nat Sci, Vol. 41, No. 1, pp. 1–10, February, 2023 73
7. Yetilmezsoy K, Demirel S. Artificial neural network containing reclaimed asphalt pavement using feed- (ANN) approach for modeling of Pb(II) adsorption forward and generalized regression neural net- from aqueous solution by Antep pistachio (Pistacia works. J Rehabil Civ Eng 2018;6:132–147. Vera L.) shells. J Hazard Mater 2008;153:1288–1300. [26] Jiang J, Trundle P, Ren J. Medical image analysis with [CrossRef] artificial neural networks. Comput Med Imaging
8. Khan T, Mustafa MRU, Isa MH, Manan TSBA, Ho Graph 2010;34:617–631. [CrossRef] YC, Lim JW, et al. Artificial neural network (ANN) [27] Bailer-Jones CAL, Gupta R, Singh HP. An for modelling adsorption of lead (Pb (II)) from Introduction to Artificial Neural Networks. In: aqueous solution. Water Air Soil Pollut 2017;228:1– Gupta R, Singh HP, Bailer-Jones CAL editors.
9. Sadrzadeh M, Mohammadi T, Ivakpour J, Kasiri Narosa Publishing House; 2001. pp. 1–18. N. Neural network modeling of Pb2+ removal from [28] Ranasinghe RATM, Jaksa MB, Kuo YL, Nejad FP. wastewater using electrodialysis. Chem Eng Process Application of artificial neural networks for predict- Process Intensif 2009;48:1371–1381. [CrossRef] ing the impact of rolling dynamic compaction using
10. Ma X, Guan Y, Mao R, Zheng S, Wei Q. Modeling dynamic cone penetrometer test results. J Rock of lead removal by living Scenedesmus obliquus Mech Geotech Eng 2017;9:340–349. [CrossRef] using backpropagation (BP) neural network algo- [29] Borovicka T, Jirina Jr M, Kordik P, Jirina M. Selecting rithm. Environ Technol Innov 2021;22:101410. Representative Data Sets. In Karahoca A, editor. [CrossRef]
11. Kardam A, Raj KR, Arora JK, Srivastava S. Advances in data mining knowledge discovery and Artificial neural network modeling for bio- applications. London: IntechOpen; 2012. pp. 43–70. [CrossRef] sorption of Pb (II) ions on nanocellulose [30] Barzilai J. On Neural-Network Training Algorithms. fibers. BioNanoSci 2012;2:153–160. [CrossRef] In: Lawless WF, Mittu R, Sofge DA, editors. Human-
12. Afroozeh M, Sohrabi MR, Davallo M, Mirnezami machine shared contexts. Massachusetts: Academic SY, Motiee F, Khosravi M. Application of artificial Press; 2020. pp. 307–313. [CrossRef] neural network, fuzzy inference system and adap- [31] Erdem F. Artificial neural network (ANN) approach tive neuro–fuzzy inference system to predict the to remazol yellow (RR) removal with S. cerevi- removal of Pb(II) ions from the aqueous solution siae. Uludağ Univ J Fac Eng 2019;24:289–297. by using magnetic graphene/nylon 6. Chem Sci J [CrossRef] 2018;9:1000185. [CrossRef] [32] Tiwaric CS, Naresh R, Jha R. Comparative study
13. Amiri MJ, Abedi-Koupai J, Eslamian S, Mousavi SF, of backpropagation algorithms in neural network Arshadi M. Modelling Pb(II) adsorption based on based identifıcation of power system. Int J Comput synthetic and industrial wastewaters by ostrich bone Sci Inf Technol 2013;5:93–107. [CrossRef] char using artificial neural network and multivari- [33] Karlik B, Vehbi A. Performance analysis of various ate non-linear regression. Int J Hydrol Sci Technol activation functions in generalized mlp architec- 2013;3:221–240. [CrossRef] tures of neural networks. Int J Artif Intell Expert
14. Kaya N, Arslan F, Yildiz Uzun Z. Production and Syst 2011;1:111–122. characterization of carbon-based adsorbents from [34] Urban S. Neural Network Architectures and waste lignocellulosic biomass: their effectiveness Activation Functions: A Gaussian Process Approach. in heavy metal removal. Fuller Nanotub Carbon 1st ed. München: Universitätsbibliothek der TU Nanostructures 2020;28:769–780. [CrossRef] München; 2018.
15. Gurney K. An Introduction to Neural Networks. 1st [35] Hajduk Z. Hardware implementation of hyperbolic ed. Florida: CRC Press; 1997. [CrossRef] tangent and sigmoid activation functions. Bull Pol
16. Zhang G, Patuwo BE, Hu MY. Forecasting with Acad Sci Tech Sci 2018;66:563–577. artificial neural networks: The state of the art. Int J [36] Seyedmohammadi J, Esmaeelnejad L, Ramezanpour Forecast 1998;14:35–62. [CrossRef] H. Determination of a suitable model for prediction
17. Kim SK, Huh JH. Artificial neural network block- of soil cation exchange capacity. Model Earth Syst chain techniques for healthcare system: Focusing on Environ 2016;2:156. [CrossRef] the personal health records. Electronics 2020;9:763. [37] Daliakopoulos IN, Coulibaly P, Tsanis IK. [CrossRef] Groundwater level forecasting using artificial neural
18. Mansourian A, Ghanizadeh AR, Golchin B. networks. J Hydrol 2005;309:229–240. [CrossRef] Modeling of resilient modulus of asphalt concrete

Share and Cite