Contents lists available at ScienceDirect

Chemometrics and Intelligent Laboratory Systems

journal homepage: www.elsevier.com/locate/chemolab

Back propagation artificial neural network and central composite designmodeling of operational parameter impact for sunset yellow and azur (II)adsorption onto MWCNT and MWCNT-Pd-NPs: Isotherm and kinetic study

Rezvan Karimi, Fakhri Yousefi⁎, Mehrorang Ghaedi⁎, Kheibar Dashtian

Department of Chemistry, Yasouj University, Yasouj 75914-353, Iran

A R T I C L E I N F O

Keywords:Azur (II)Sunset yellowPalladium nanoparticlesArtificial neural networkCentral composite design

A B S T R A C T

Impact operational parameters such as initial dyes concentration, adsorbent mass, pH and sonication time onthe efficiency of ultrasound-assisted adsorption of sunset yellow (SY) and azur (II) (AZ) onto MWCNT andMWCNT-Pd-NPs was studied by central composite design (CCD) and back propagation artificial neural network(ANN) models and resulted for tow adsorbent were compared to each other. Optimize conditions from CCD inpresence of MWCNT as adsorbent was found to be 6.01 and 9.98 mg L−1 AZ and SY concentration, 0.02 ofadsorbent mass, pH 7.0 and 3.8 min sonication time, respectively, while in the presence of MWCNT-Pd-NPsoptimize value were obtained at 10.0 and 8.0 mg L−1 AZ and SY concentration, 0.018 of adsorbent mass and pH5.0 and 3.5 min sonication time, respectively. In these conditions and at desirability value of 0.99, the maximumadsorption efficiency for AZ onto MWCNT and MWCNT-Pd-NPs are 82.97 and 94.95, respectively. Also, themaximum adsorption efficiency for SY onto MWCNT and MWCNT-Pd-NPs are 73.13 and 85.55, respectively.This values show that MWCNT-Pd-NPs in comparison with MWCNT at the lower contact time and adsorbentmass was able to remove greater amounts of under study dyes. Based on the ANN model the absolute averagedeviations (AADs) of AZ and SY dyes adsorption by MWCNT from experimental data are 0.27% and 0.43%, andthe R2 values are 0.911 and 0.954, respectively. Also, the AADs of AZ and SY dyes adsorption by MWCNTs-Pd-NPs are 0.44% and 0.65%, and the R2 values are 0.925% and 0.927%, respectively. This values show that ANNfor MWCNT-Pd-NPs in comparison with MWCNT was a more powerful tool for building up to construct anempirical model to predict under study dyes adsorption behavior. In addition, the obtained results have goodagreement with experimental data for MWCNT-Pd-NPs in comparison with MWCNT. Finally, the study ofequilibrium isotherm and kinetic models of adsorption process are shown that the Langmuir isotherm modeland pseudo second order kinetic model were the best models for both adsorbent.

1. Introduction

Water pollutions consist of highly toxic organic dyes is an importantenvironmental issue, and receives major worldwide concerns [1,2].Simultaneous dyes discharged with industrial wastewaters even at lowconcentrations are highly visible, infrangible, non-biodegradable andundesirable [3]. Toxic organic dyes due to considering the havedisadvantages such as carcinogenic and mutagenic create the seriousthreats to human health and organisms [4,5]. Therefore, the dyes mustbe removed from the wastewaters before entered to the human healthand other organisms body to solve the environmental issues. Theremoval approach generally includes the chemical, physical, andbiological methods, which amongst the adsorption as physicochemicalmethod is widely used because of its environmentally friendly, high

removal capacity, low cost and simplicity of operation [6–8]. In thismethod fined and design of novel adsorbent with high capacity andhigh rate is cases, which in previous report various highly efficientadsorbents, such as carbon based material, metallic and oxo-metallicnanoparticles, organic and inorganic polymers have been extensivelyapplied for removal of dyes from aquaes media [6,9–12]. Among thecarbon based material, multiwalled carbon nanotubes (MWCNTs) is acommercial support with medium surface area and pore size forremoval of dyes from aquaes media. Hence, the improved of thisdisadvantage can be performed with deposited of other materials suchas nonmetallic and oxo-metallic materials on its surface [13,14]. Thus,the immobilized of cheap, eco-friendly and efficient nanomaterial is asubject of intensive research. Therefore, we decide to deposited of Pdnanoparticles (Pd-NPs) onto MWCNT due to the stability of its

http://dx.doi.org/10.1016/j.chemolab.2016.10.012Received 25 June 2016; Received in revised form 27 September 2016; Accepted 25 October 2016

⁎ Corresponding authors: Tel/fax: +98-74-33223048E-mail addresses: fyousefi@yu.ac.ir (F. Yousefi), m_ghaedi@mail.yu.ac.ir (M. Ghaedi).

Chemometrics and Intelligent Laboratory Systems 159 (2016) 127–137

0169-7439/ © 2016 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).Available online 27 October 2016

crossmark

chemical structure, biocompatibility, strong oxidizing power and non-toxicity of this material [15–17]. This new material as MWCNT-Pd-NPs was characterized with SEM and XRD and subsequently wasapplied for removal of sunset yellow and azur II dyes under ultrasonicirradiation.

Performing the adsorption process at non-optimized conditions canbe lead to use of more reagent and spend much time with the lowestearnings [18]. Therefore, modeling and optimization of operationalparameter on process is very important. The optimization and model-ing approach generally includes the one at a time, response surfacemethodology (RSM) and the artificial neural network (ANN) methods[19,20]. One at a time method due to having a number of disadvantagesuch as not consider the interaction effects between operationalparameters and not valid optimum condition, recently less used, while,RSM and ANN models were improved these problem due to theirprediction accuracy on complex nonlinear systems [21,22]. Centralcomposite design (CCD) under response surface methodology is acombined of mathematical and statistical techniques for empiricalmodeling of many process [23,24]. In CCD method, the aim is tooptimize a dependent variables or responses (in this work removalpercentage (R%)) that verified by independent parameters (operationalparameters), as well as, comfortable conditions with minimum processknowledge, thereby saving time and experimental cost and materials[19,25,26]. Also, the ANN based on the common multi-layeredperceptron (MLP) consist of input, output and hidden layers of neuronsis a powerful tool for modeling of data obtained from CCD by havingmore advantages such as no assumptions concerning the nature of thephenomenological mechanisms, understanding the mathematical back-ground of problem underlying the process. Besides, it has the ability tolearn linear and nonlinear relationships between variables directlyfrom a set of examples than to conventional modeling techniques[27,28]. In other word, the ANN is a data driven technique andcomputes the relationship between input and output variables in aneffort to understand [19]. Therefore, in this work we focused onmodeling and optimization of adsorption of SY and AZ dyes ontoMWCNT and MWCNT-Pd-NPs with CCD and ANN method. Finally,the kinetic an isotherm of the process was studied and fitted withconventional method.

2. Experimental

2.1. Materials and apparatus

Chemical reagents including MWCNT, sodiumtetrachloropalladate(II) (Na2PdCl4), ethanol, stretch, ascorbic acid,HCl and sodium hydroxide (NaOH) without further purification werepurchased from (Merck company, Dermasdat, Germany). The pH wasadjusted and measured using pH/Ion meter model 686 (Metrohm,Switzerland, Swiss). SY and AZ concentration was determined usingJasco UV–vis spectrophotometer model V-530 (Jasco, Japan). Theprepared samples were characterized by SEM (KYKY-EM3200, China)and XRD (PW 1800, Philips, Germany). An ultrasonic bath withheating system (Tecno-GAZ SPA Ultra Sonic System, Parma, Italy) at40 kHz of frequency and 130 W of power was used for the ultrasound-assisted adsorption procedure.

2.2. Synthesis of MWCNT-Pd-NPs

The Pd nanoparticles deposited on MWCNT based on a one-stepreduction process in an aqueous solution was carried out as follow:300 µL aliquot of 0.02 M Na2PdCl4 was added into 50 mL of anaqueous solution containing 0.2 wt % of the soluble starch andsonicated at 80 °C for 1 h. Then, 5 mL of a 0.4 M ascorbic acid asreducing agent was added to the above solution under constant stirringand at the preset temperatures insofar as turned brown color ofsolution that indicated the initial formation of the Pd nanoparticles

was observed. The mixture was maintained at this insofar as color ofthe reaction solution was changed to dark brown. Appropriate aliquotsof the Pd nanoparticles solution was mixed with MWCNT in anErlenmeyer under sonication for up to 12 h. Finally, the product asMWCNT-supported Pd nanoparticles were filtered and extensivelywashed with ethanol and double distilled water and dried at 70 °Cfor 10 h.

2.3. Ultrasonic assisted adsorption experiment

The simultaneous ultrasonic assisted adsorption of SY and AZ ontoMWCNT and MWCNT-Pd-NPs was carried out as follow: 25 mL ofcertain concentration of SY and AZ in binary solution at different pHwas mixed completely with 0.02 g of the MWCNT and MWCNT-Pd-NPs and it subsequent exposure under ultrasound irradiation for 5 minto best dispersion of adsorbents into the solutions which raising masstransfer of liquid to solid. Finally, 3 mL of samples were drawn out andimmediately centrifuged and SY and AZ concentration in effluentsolution was obtained from calibration curve from absorbance spec-trum of these dyes in concentration range of 2–10 mg L−1,then dyesremoval percentage (R%) and adsorption capacity (Qe) was calculatedaccording to our previous report [29–32]. Typically, the absorbancespectra of these dyes in single and binary solution (in concentration of8 mg L−1 for each dye) is shown in Fig. 1, as seen that in binary solutionspectrum of this dyes are not overlap. Therefore, the removal percen-tage of these dyes was calculated in wavelength of 417 and 675 nm forSY and AZ, respectively.

2.4. Central composite design (CCD)

To find the optimum conditions, investigation of effect of parameterand their interaction for simultaneous adsorption of the SY and AZonto MWCNT and MWCNT-Pd-NPs, the response surface methodology

Fig. 1. Chemical structure of azur (II) (a) and sunset yellow (b) and absorption spectraof SY and AZ dyes in single and binary solutions (initial dye concentration for SY and ERwas 8.0 mg L−1) (c).

R. Karimi et al. Chemometrics and Intelligent Laboratory Systems 159 (2016) 127–137

128

has to be studied. Rotatable experimental design was carried out as aCCD consisting of 50 experiments. For five parameters (n=5) and twolevels (low (-) and high (+)), the total number of experiments was 50determined by the expression: 2 n (25=32: factor points)+2 n (2*5=10:axial points)+8 (center points: eight replications), as shown inTable 1for two adsorbent. The parameters in this experiment werethe dyes concentration (X1andX2), pH (X3), adsorbents mass (X4) andsonication time (X5). A full second-order polynomial model obtainedfor five factors can be permitted the response to be modelled, which canquadratic equation following their analysis:

∑ ∑ ∑iXiY = βo+ β + βiiXi + βijXiXji i i j=1

5

=1

52

> (1)

This model consists of five main effects, ten two-factor effects andfive curvature effects, where Y is the predicted response (removalpercentage); βo is the intercept and other terms was expressedaccording in last research. xi's are the independent variables that are

known for each experimental run. The parameter βo is the modelconstant; βi's are the linear coefficients; βii's are the quadraticcoefficients and βij's are the interaction coefficients [31].

To check the applicability of the proposed model for tow adsorbent,the determination of dyes concentration for all runs in CCD matrix ineffluent binary solution was carried out by spectrophotometry UV–visspectrophotometer. Analysis of variance (ANOVA) used to realize thediagnostic checking tests for adequacy of proposed model, the coeffi-cient of determination R2, adjust R2and predicated R2was investigated[33–35].

2.5. Artificial neural network

The ANN is the non-linear mathematical methods which get vastattention because of its simplicity, flexibility, and accessibility todifferent training algorithms plus its large modeling capability [36–38]. The ANN is obtained based on the activity procedure of human

Table 1The design matrix and the responses.

MWCNT MWCNT- Pd-NPs

X1 X2 X3 X4 X5 R%AZ R%SY X1 X2 X3 X4 X5 R%AZ R%SY

1 10 6 5 0.020 4 73.89 63.85 6 6 7 0.010 4 74.48 65.022 6 10 5 0.020 4 80.36 73.35 6 10 5 0.020 4 95.18 72.063 10 10 7 0.020 2 75.08 62.52 8 8 6 0.015 5 88.8 78.004 12 8 6 0.015 3 81.97 65.05 10 6 7 0.020 2 96.69 74.055 8 8 6 0.015 3 75.81 62.88 8 8 6 0.015 1 85.1 71.886 6 6 5 0.020 2 81.77 66.69 10 6 7 0.020 4 94.36 75.697 6 10 7 0.020 2 75.60 67.81 10 10 5 0.010 2 77.3 78.368 8 8 6 0.005 3 76.12 63.98 6 6 5 0.010 4 76.12 65.049 6 6 5 0.020 4 78.37 65.70 8 8 6 0.015 3 89.57 80.4710 6 10 7 0.010 4 78.36 60.55 10 6 7 0.010 2 94.75 70.9611 8 4 6 0.015 3 79.01 66.96 10 6 5 0.010 4 93.40 75.9612 6 10 5 0.010 2 74.20 62.57 10 10 5 0.020 2 81.49 79.1413 10 10 7 0.010 2 74.32 55.01 6 6 5 0.020 2 79.67 64.0114 8 8 6 0.015 3 75.98 64.63 10 10 7 0.010 4 82.36 76.4515 10 6 7 0.020 4 78.16 63.98 8 4 6 0.015 3 87.98 72.9816 6 6 5 0.010 4 75.55 64.95 4 8 6 0.015 3 82.59 67.1017 6 10 5 0.010 4 75.58 65.68 8 8 6 0.015 3 89.39 80.9718 6 10 7 0.010 2 69.85 58.90 6 10 7 0.020 4 91.30 71.6919 8 8 8 0.015 3 79.01 68.49 6 10 5 0.010 4 90.52 69.7820 8 8 6 0.015 1 67.10 60.99 6 6 5 0.010 2 76.48 64.9521 10 6 5 0.010 2 82.16 70.97 8 8 6 0.015 3 89.57 81.6322 10 6 5 0.010 4 76.26 65.01 10 6 5 0.020 2 93.84 77.0423 10 6 7 0.010 2 79.35 68.02 8 8 6 0.015 3 89.50 80.0924 10 10 5 0.010 4 72.95 65.70 6 6 5 0.020 4 80.01 69.9825 8 8 6 0.015 3 75.90 64.39 10 6 5 0.010 2 90.99 73.3926 8 8 6 0.025 3 79.43 71.54 8 8 6 0.015 3 89.43 80.5427 10 6 7 0.010 4 80.63 61.10 6 10 7 0.020 2 90.01 67.0128 6 6 7 0.010 2 71.51 70.19 10 10 7 0.010 2 79.40 74.7829 10 10 7 0.020 4 78.80 63.39 6 10 5 0.020 2 89.10 66.9630 6 6 7 0.020 2 75.30 73.14 8 8 6 0.015 3 89.58 79.6931 8 8 6 0.015 3 76.01 64.41 10 6 5 0.020 4 96.37 79.3432 10 10 5 0.020 4 72.59 70.91 8 8 6 0.025 3 83.98 71.0733 10 10 7 0.010 4 79.16 53.78 6 10 7 0.010 4 86.96 66.9634 10 10 5 0.010 2 75.55 64.70 10 6 7 0.010 4 91.83 71.0935 8 8 6 0.015 3 76.01 64.47 8 8 6 0.005 3 77.15 64.6736 6 6 7 0.020 4 79.26 73.38 10 10 7 0.020 4 86.33 83.5637 4 8 6 0.015 3 81.93 70.01 6 10 5 0.010 2 84.63 66.7438 8 8 6 0.015 3 75.97 64.00 8 8 4 0.015 3 91.46 78.8639 8 8 6 0.015 3 75.98 65.01 8 8 6 0.015 3 89.6 81.3640 6 6 5 0.010 2 77.78 68.70 8 8 6 0.015 3 89.49 80.3641 8 8 6 0.015 5 68.79 57.97 6 6 7 0.010 2 79.96 64.4642 6 10 5 0.020 2 79.98 67.85 10 10 5 0.020 4 90.31 86.0443 6 6 7 0.010 4 76.45 66.19 10 10 5 0.010 4 85.60 80.4344 10 10 5 0.020 2 76.12 66.50 6 10 7 0.010 2 86.12 63.6745 10 6 5 0.020 2 80.91 66.03 8 8 8 0.015 3 90.98 75.0346 10 6 7 0.020 2 78.00 68.19 8 12 6 0.015 3 88.00 79.9947 8 8 6 0.015 3 76.01 64.41 6 6 7 0.020 4 77.19 67.9948 8 8 4 0.015 3 79.10 71.01 10 10 7 0.020 2 82.69 80.0149 8 12 6 0.015 3 75.90 61.09 6 6 7 0.020 2 82.21 64.0350 6 10 7 0.020 4 82.97 73.69 12 8 6 0.015 3 92.50 87.01

R. Karimi et al. Chemometrics and Intelligent Laboratory Systems 159 (2016) 127–137

129

brain and has been applied to model of many scientific disciplines[39,40]. Good details and various applications of the ANN werepresented in prior publications [41,42].

The structure of ANN forms the input layer (independent vari-ables), the output layer (dependent variables) and one or more neuronlayers called hidden layers (Fig. 2). The number of network layers, thenumber of neurons in each layer, the nature of learning algorithms andneurons transfer functions are defined the construction of the ANN.

One class of feed forward neural network that has been usedcommonly for the approximate function is multilayer perceptron(MLP) [43]. The MLP neural networks are contained of one inputlayer, at least one hidden layer and an output layer. The MLP learns thedata pattern using an algorithm known as “training”. These algorithms

modify weights of the neurons according to the error between thevalues of real output and target output where provide non-linearregression between inputs and outputs variables. One of the examplesof training algorithms is the back propagation algorithm that widelyused to train of ANN in different applications. The transfer functionsthat used in the hidden layer; and output layer are “Tansig” and“Purelin”, respectively. The back propagation algorithm of Levenberg–Marquardt (trainlm), based on our experiments, was applied todetermine of most favorable network structure. The trial and errorapproach is the most basic method of training a neural network. Whilenumber of hidden layers is to be selected depending on the complexityof the problem, but generally one hidden layer is sufficient for modelingof most of the problems. In this method firstly the number of hiddenlayers considered one and the next step, in the trial and error approach,the number of neurons in the hidden layer varied one by one to attainthe desire objective functions outputs. In this work, the ANN underMATLAB toolbox was used as an experimental model to study thecondition of ultrasonic assisted adsorption of AZ and SY dyes. Theinput parameters to the ANN model are pH, dyes concentration,sonication time, adsorbents mass and the target is AZ and SY dyesremoval. Finally, the outputs obtained from the ANN modeling werecompared with the experimental data, and advantages were discussed.

In statistics, the mean squared error (MSE) of an estimatormeasures the average of the squares of the errors or deviations thatis, the difference between the estimator and what is estimated. MSE is arisk function, corresponding to the expected value of the squared errorloss or quadratic loss. The difference occurs because of randomness or

Fig. 2. Topology of conventional P S− − 1 MLP.

Scheme 1. Flow diagram for methodology of the work.

R. Karimi et al. Chemometrics and Intelligent Laboratory Systems 159 (2016) 127–137

130

because the estimator doesn't account for information that couldproduce a more accurate estimate [44]. The average absolute deviationof a data set is the average of the absolute deviations from a centralpoint. It is a summary statistic of statistical dispersion or variability. Inthis general form, the central point can be the mean, median, mode, orthe result of another measure of central tendency. Furthermore, thedeviation averaging operation may refer to the mean or the median[45]. The coefficient of determination (R2) is a key output of regressionanalysis. It is interpreted as the proportion of the variance in thedependent variable that is predictable from the independent variable.R2 is the square of the correlation (r) between predicted y scores andactually scores; thus, it ranges from 0 to 1 [46]. The mathematicaldefinition of the errors criteria including mean square error (MSE),absolute average deviation percentage (AAD%) and coefficient ofdetermination (R2) were given as bellow:

∑MSE ρ ρ= ( − )i

N

i ical

=1

exp 2

(2)

⎛⎝⎜⎜

⎞⎠⎟⎟∑AAD

Nρ ρ

ρ(%) = 1 −

i

Ni i

cal

i=1

exp

exp(3)

RΣ ρ ρ Σ ρ ρ

Σ ρ ρ( − ) − ( − )

( − )iN

i iN

i ical

iN

i

2 =1exp 2

=1exp 2

=1exp 2 (4)

It should be stated that to achieve generalization of the model andavoid overtraining, the net is started with two neurons in the hiddenlayer and gradually increased the number of neurons until nosignificant improvement in performance of the net was observed.Scheme 1 presents a diagram for the ANN model used in this work.

3. Results and discussion

3.1. Characterization of adsorbents

The XRD pattern (Fig. 3a) for MWCNT shows two peak at the 2θ of25.95° and 46° which indicate the reflection of carbon structure anddiffractions of graphite index of the MWCNT. After deposited of Pd-NPs onto MWCNT the other appearance diffraction peaks at 2θ 40.11°,46.50°, 68.04° and 81.78° corresponded to the (111), (200), (220) and(311) planes (Fig. 3a), respectively, are characteristic of face centercubic (fcc) crystalline of Pd-NPs. The average size of Pd-NPs wascalculated by Debye-Scherrer formula from the (111) reflection peak ofthe Pd face center cubic structure around 2θ 40.04°. The calculatedaverage particle size in this research is approximately 20 nm.

The morphological features of the prepared samples were studiedby SEM analysis (Fig. 3b and c), that SEM image of pure MWCNT(Fig. 3b) before the surface modification it appears to be smooth,homogeneous, tidy and approximately uniform size distribution. Afterthe deposited of with Pd-NPs onto MWCNT, SEM image reveal thatsmooth surface of MWCNTs has become rough, increase in the size andbundled and shows the spherical nanoparticles corresponding to Pdnanoparticles deposited on MWCNT (Fig. 3c).

3.2. CCD analysis

In first step, Lack of fit testes and model summary statistics forlinear, 2FI, quadratic and cubic models for adsorption of SY and AZonto MWCNT and MWCNT-Pd-NPs was performed and according tothe insignificant lack of fit and maximizing the R-squared, adjusted R-squared and the predicted R-squared of quadratic model for toresponses, this model was selected for other step analysis in towresponses and adsorbent. Then, ANOVA (Table 2) for quadratic modelwas applied to test the statistical significant factors. ANOVA foradsorption of SY and AZ dyes onto MWCNT shows good model F-values of 119.6 and 5571.81, respectively. Also, for adsorption of SY

and AZ dyes onto MWCNT-Pd-NPs shows good model F-values of121.6 and 6934.14, respectively. Low P-values ( < 0.0001) for moreterms in four model that confirm well applicability of constructedmodels for correlating the adsorption of SY and AZ onto MWCNT andMWCNT-Pd-NPs. Also, the Lack of Fit F-value for adsorption SY andAZ onto MWCNT were obtained to be 0.98 and 1.02, respectively. Also,values of 0.1218 and 0.1184 was obtained for adsorption SY and AZonto MWCNT-Pd-NPs, respectively. These P-values explanatory notsignificant relative to the pure error.Therefor, not-significant lack of fitis good and models is a fit for all of responses (see Table 2). Finally,based on these results (P-values lower than 0.05 for variable), anempirical relationship between the response and independent variableswas attained and expressed by the following second-order polynomialequation:

R in presence of MWCNT X X

X XX X X X X X X X

X X XX X X X X X X X X

X X X XX X X X

% = +75.95 + 0.028 × − 0.75 ×

− 0.035 × + 0.85 ×+ 0.38 × − 0.80 × + 0.85 × − 1.29 × − 0.94 ×

+ 0.46 × + 0.49× + 0.88 × + 1.80 × − 0.27 × + 1.50 ×

+ 0.38 × + 0.78 ×+ 0.46 × − 2.00 ×

AZ 1 2

3 4

5 1 2 1 1 4 1 5

2 3 2

4 3 4 3 5 4 5 1 1

2 2 3 3

4 4 5 5 (5)

R in presence of MWCNT Pd NPs X

X XX X X X X X X X

X X XX X X X X X X X X

X X X XX X X X

% − − = +64.22 − 1.49 ×

– − 1.37 × − 0.86 ×+ 2.00 × − 0.32 × − 1.44 × − 0.71 × − 0.68 ×

− 1.68 × + 1.67× + 1.53 × + 1.24 × − 0.28 × + 0.80 ×

+ 0.78 × + 1.34 ×+ 0.84 × − 1.23 ×

SY 1

2 3

4 5 1 3 1 4 1 5

2 3 2

4 2 5 3 4 3 5 1 1

2 2 3 3

4 4 5 5

(6)

R in presence of MWCNT X X

X XX X X X X X X X X X

XX X X X X X X X X

X X XX X X X X X X

% = 89.52 + 2.44 × − 0.13 ×

+ 1.74 × + 0.86 ×− 5.45 × + 0.081 × − 0.092 × + 0.62 × − 0.42 ×

+ 0.35× + 1.52 × − 0.19 × − 1.28 × + 0.12 ×

− 0.50 × − 0.38× + 0.42 × − 2.24 × − 0.64 ×

AZ 1 3

4 5

1 2 1 3 1 4 1 5 2 3

2

4 2 5 3 4 3 5 4 5

1 1 2

2 3 3 4 4 5 5

(7)

R in presence of MWCNT Pd NPs X

X XX X X X X X X X

X X XX X X X X X X X X

X X X X

% − − = 80.81 + 5.14 ×

+ 1.87 × − 0.99 ×+ 1.58 × + 1.49 × + 0.68 × − 0.45 × + 0.51 ×

+ 0.41 × + 0.34× + 0.65 × − 1.11 × − 1.25 × − 1.14 ×

− 3.41 × − 1.64 ×

SY 1

2 3

4 5 1 2 1 3 1 4

2 5 3

4 5 6 1 1 2 2 3 3

4 4 5 5

(8)

Removal percentage (R%) have been predicted by Eqs. (5–8) andindicated good agreements between the experimental and predictedvalues of adsorption efficiency (Fig. 4a and Table 3). The experimentalresults and the predicted values obtained from the model (Eqs. 5–8)was shown that predicted values matched the experimental valuesreasonably well with R2=0.9628 and 0.9992 for SY and AZ adsorptiononto MWCNT and R2=0.9702 and 0.9993 for SY and AZ adsorptiononto MWCNT-Pd-NPs (see Table 3). This implies that higher than 96%of the variations for percent dyes removal are explained by theindependent variables and this also means that the model does notexplain only about 4% of variation. Adjusted R2as corrects R2-value forthe sample size and the number of terms in the model by the degrees of

R. Karimi et al. Chemometrics and Intelligent Laboratory Systems 159 (2016) 127–137

131

freedom on its computations is also more suitable for comparingmodels with different numbers of independent variables. If there aremany terms in a model and not very large sample size, Adj-R2 may bevisibly smaller than R squared (R2) [47,48]. Adj-R2 values of 0.9799and 0.9996 for SY and AZ adsorption onto MWCNT was close to thecorresponding R squared (R2) value (see Table 3). Also Adj-R2 values of0.9801 and 0.9996 for SY and AZ adsorption onto MWCNT-Pd-NPswas close to the corresponding R squared (R2) value (see Table 3). Inaddition, the values higher than 0.988 for determination coefficient R2

for each response in presece of MWCNT and MWCNT-Pd-NPs (seeTable 3) is shown suitability of full quadratic model for predicting thereal behaviorof adsorption process. Additional confirmation on theapplicability of the model for predicting the performance of adsorptionefficiency are counted as the low and acceptable standard deviationvalues, low PRESS value, low coefficient of variation, low standarderror, high “Adeq precision” value (very higher than 4.0) and highmean value of ER % (see Table 3).

The Box Cox plot (Fig. 4) as a power tool to help analyzer determinethe most appropriate power transformation to apply to response dataand indicated that studies have been performed whether in coefficientinterval or not [49,50]. In these plots the current transformation 1.0indicates no transformation. According to the best values for R% of SYand AZ onto MWCNT and MWCNT-Pd-NPs have been performed 95%of coefficient interval and not necessary to transformation.

3.3. Optimize condition

In adsorption process find the best condition that the responses aremaximized is important cases. Therefore, desirability function ap-proach under CCD method was used for this goal by choice of valuefrom 0.0 (unacceptable) to 1.0 (acceptable).Optimize conditions fromCCD when the MWCNT was used for removal of SY and AZ dyes can be

found at 6.01 and 9.98 mg L−1 AZ and SY concentration, 0.02 ofadsorbent mass and 7.0 for pH value and 3.8 min sonication time,respectively, while in the presence of MWCNT-Pd-NPs optimizedvalues were obtained at 10.0 and 8.0 mg L−1 AZ and SY concentration,0.018 of adsorbent mass and 5.0 for pH value and 3.5 min sonicationtime, respectively. In these condition and at desirability value of 0.99,maximum adsorption efficiency of 82.97 and 94.95, and 73.13 and85.55 for AZ and SY onto MWCNT and MWCNT-Pd-NPs wereobtained, respectively. This values show that MWCNT-Pd-NPs incomparison with MWCNT at the lower contact time and adsorbentmass was able to remove greater amounts of under study dyes.

3.4. Effect of variables as response surface plots

Response surface plots as a function of two variables at a time andholding all other variables at center levels are more powerful tool forunderstanding main and the interaction effects of these two variables[51,52]. Therefore, the response surface curves (Fig. 5) were plotted tounderstand the interaction of the variables and to determine theoptimum level of each variable for maximum response.

Chemical structure of AZ and SY dyes show that their ionic chargein addition to MWCNT and MWCNT-Pd-NPs nature are greatlyaffected by pH. The dyes adsorption occurs by combination of variousmechanism like ion exchange, hydrogen bonding, electrostatic, andsoft-soft and dipole-ion interaction [18]. Effects of initial solution pHon under study dyes adsorption over MWCNT and MWCNT-Pd-NPssurface were investigated in the range of 4–8. In pH lower than 4.0 dueto simultaneous probable protonation of O, S and N atoms in dyes andO atoms in surface of adsorbents repulsive force was further thatextensively reduce the amount of removal percentage, hence, pH lowerthan 4.0 was not studied. At pH 4.0, dyes deprotonating lead toenhance in their mass transfer and subsequent adsorption efficiency. At

Fig. 3. XRD pattern (a) TG (b) of MWCNT-Pd-NPs and SEM images of the MWCNT (c) and MWCNT-Pd-NPs (d).

R. Karimi et al. Chemometrics and Intelligent Laboratory Systems 159 (2016) 127–137

132

pH higher than 6.5, due to probable competition of hydroxide withdyes and also formation of similar negative charge on dyes andadsorbents causes significant decrease in dyes removal percentage.Both at high acidic and basic media, expected un-stability of MWCNTand MWCNT-Pd-Plead to reduce in dyes removal percentage andsubsequently lowering life time and reusability of adsorbent.Therefore, all subsequent adsorption experiments were conducted atpH 4.0 for two dyes and adsorbents which process has maximumadsorption efficiency (typically see Fig. 5).

The effect of various initial dyes concentrations on adsorptionprocess at MWCNT and MWCNT-Pd-NPs was shown that increase inthe initial dyes concentration leads to decrease in the adsorptionprocess of the dyes onto two adsorbents. Due to increase in the initial

dyes concentrations gradient between understudy adsorbents and dyes,the adsorption efficiency was high until the system attain its equili-brium and after equilibrium and saturation point of adsorbents, thedyes adsorption percentage due to remains of dyes in to the solutionwas decreased.

The effect of sonication time on the adsorption of understudy dyeswas shown that in the initial stages due to availability of the reactivesite of the MWCNT and MWCNT-Pd-NPs adsorbents removal percen-tage of dyes was increased dramatically, whereas, with increase ofsonication time the removal efficiency gradually gets increased (typi-cally see Fig. 5). Moreover, no significant changes were observed in thepercentage of removal of the dyes after equilibrium. In other word,occur of the maximum adsorption of understudy dyes at short

Table 2The results of ANOVA for the response surface quadratic model.

Source ofvariation

DF MWCNT-Pd-NPs MWCNT

R%AZ R%SY R%AZ R%SY

Sum ofsquares

Meansquare

F value p-valueProb >F

Sum ofsquares

Meansquare

F value p-valueProb >F

Sum ofsquares

Meansquare

F value p-valueProb >F

Sum ofsquares

Meansquare

F value p-valueProb >F

Model 20 1682.93 84.15 6934.14 <0.0001

2073.33 103.67 121.44 <0.0001

0.031 27.97 5571.81 <0.0001

931.46 46.57 119.6 <0.0001

X1 1 238.10 238.10 19620.39 <0.0001

1058.43 1058.43 1239.93 <0.0001

22.65 0.031 6.25 0.0183 88.80 88.80 228.0 <0.0001

X2 1 0.025 0.025 2.02 0.1660 139.35 139.35 163.25 <0.0001

0.049 22.65 4512.08 <0.0001

75.41 75.41 193.6 <0.0001

X3 1 0.71 0.71 58.53 <0.0001

38.93 38.93 45.60 <0.0001

29.10 0.049 9.76 0.0040 29.52 29.52 75.80 <0.0001

X4 1 120.79 120.79 9953.86 <0.0001

100.36 100.36 117.57 <0.0001

5.81 29.10 5797.78 <0.0001

160.32 160.32 411.7 <0.0001

X5 1 29.57 29.57 2436.47 <0.0001

89.28 89.28 104.59 <0.0001

20.64 5.81 1156.68 <0.0001

3.98 3.98 10.22 0.0033

X1×2 1 951.03 951.03 78369.84 <0.0001

14.93 14.93 17.49 0.0002 23.39 20.64 4111.69 <0.0001

1.16 1.16 2.99 0.0946

X1×3 1 0.21 0.21 17.27 0.0003 6.50 6.50 7.61 0.0099 52.94 23.39 4660.00 <0.0001

66.24 66.24 170.1 <0.0001

X1×4 1 0.27 0.27 22.41 <0.0001

8.34 8.34 9.77 0.0040 28.05 52.94 10546.41 <0.0001

16.25 16.25 41.72 <0.0001

X1×5 1 12.29 12.29 1012.63 <0.0001

1.07 1.07 1.26 0.2714 6.88 28.05 5587.76 <0.0001

14.93 14.93 38.35 <0.0001

X2×3 1 5.74 5.74 472.81 <0.0001

0.034 0.034 0.040 0.8437 7.57 6.88 1370.95 <0.0001

90.79 90.79 233.1 <0.0001

X2×4 1 3.91 3.91 322.45 <0.0001

2.02 2.02 2.37 0.1348 24.99 7.57 1507.21 <0.0001

88.78 88.78 227.9 <0.0001

X2×5 1 73.96 73.96 6094.99 <0.0001

5.35 5.35 6.26 0.0182 5.512E−003

24.99 4978.66 <0.0001

74.91 74.91 192.3 <0.0001

X3×4 1 1.13 1.13 93.02 <0.0001

3.59 3.59 4.21 0.0494 104.04 5.512E−003

1.10 0.3033 49.40 49.40 126.8 <0.0001

X3×5 1 52.61 52.61 4335.21 <0.0001

2.29 2.29 2.68 0.1123 2.30 104.04 20725.52 <0.0001

2.45 2.45 6.30 0.0179

X4×5 1 0.43 0.43 35.44 <0.0001

13.36 13.36 15.66 0.0004 72.42 2.30 458.28 <0.0001

20.51 20.51 52.68 <0.0001

X12 1 7.85 7.85 646.62 <

0.000139.59 39.59 46.38 <

0.00014.64 72.42 14426.66 <

0.000119.54 19.54 50.19 <

0.0001X2

2 1 4.72 4.72 388.71 <0.0001

50.38 50.38 59.02 <0.0001

19.50 4.64 923.52 <0.0001

0.29 0.29 0.74 0.3974

X32 1 5.74 5.74 473.09 <

0.000141.57 41.57 48.70 <

0.00016.79 19.50 3884.53 <

0.000157.16 57.16 146.7 <

0.0001X4

2 1 160.59 160.59 13233.53 <0.0001

371.77 371.77 435.52 <0.0001

127.60 6.79 1352.54 <0.0001

22.53 22.53 57.85 <0.0001

X52 1 13.27 13.27 1093.44 <

0.000186.17 86.17 100.95 <

0.00010.15 127.60 25418.83 <

0.000148.49 48.49 124.5 <

0.0001Residual 29 0.35 0.012 24.75 0.85 0.11 5.020E

−00311.29 0.39

Lack ofFit

22 0.31 0.014 2.40 0.1184 21.83 0.99 2.37 0.1218 0.034 5.050E−003

1.02 0.5270 8.52 0.39 0.98 0.5578

PureError

7 0.041 5.884E−003

2.93 0.42 559.55 4.927E−003

2.78 0.40

Cor Total 49 1683.28 2098.08 0.031 27.97 5571.81 <0.0001

942.75 46.57 119.60 <0.0001

R. Karimi et al. Chemometrics and Intelligent Laboratory Systems 159 (2016) 127–137

133

sonication time is due to high contribution of ultrasound power in masstransfer and simultaneously increase available surface area and vacantsites of MWCNT and MWCNT-Pd-NPs. So that, the mass transferincrease is due to raising diffusion coefficient encounter.

The effect of MWCNT and MWCNT-Pd-NPs mass in the adsorptionprocess shows that by varying the adsorbent mass two understudy dyeshave shown similar results that with increase in mass of adsorbentsincreases adsorption efficiency. In other word, enhancement of adsor-bent mass can be increases of number of available adsorption sites aswell as increases of removal efficiency.

3.5. ANN analysis

In this project, to predict the AZ and SY dyes removal behavior ontoMWCNTs and MWCNTs-Pd-NPs, the pH, dyes concentration, sonica-tion time and adsorbents mass used as input variables and the AZ andSY dyes removal are used as target. The ranges of input-outputvariables for each system are given in Table 4. All experimental datapoint that used in this study is 100. The MLP is trained, validated, andtested with random 70%, 15%, and 15%, respectively. The possibility ofover-training is a problem in the ANN and can be overcome by properselection of the number of neurons in hidden layer. The performance ofthe network in training phase should increase with increasing thenumber of neuron, at the same time as the performance of the networkin testing data phase lead to optimum value at an optimal number ofhidden neurons. In this research, the absolute average deviation (AAD%) was chosen as a measure of the performance of the network.Typically, the network with one hidden layer and 17 neurons was usedto predict the AZ and SY dyes removal by MWCNTs and leads to thebest prediction in Fig. 6. Besides, the one hidden layer and 13 neuronscan be predicting the AZ and SY dyes removal by MWCNTs- Pd-NPs.The evolution of training, validation, and test errors as a function of thenumber of training epochs (based on early stopping approach) wasinvestigated and typically results for MWCNT adsorbent was shown inFig. 7. Also, the best epoch for the AZ and SY dyes removal byMWCNTs- Pd-NPs equal to 7.

The optimized neural network model was used to predict the AZand SY dyes removal by MWCNTs. The comparison between thepredictive values and the experimental values is curried out and shownin Fig. 8. The results of Fig. 8a and b show that good agreementbetween the predicted and the experimental values of the AZ and SYdyes removal by MWCNTs and the AADs values are 0.27% and 0.43%,and the R2 values are 0.911 and 0.954, respectively. Fig. 9 shows thatthe error analysis of trained network over number of test for the AZ andSY dyes removal by MWCNTs.

3.6. Adsorption equilibrium

To optimize the design of an adsorption system, it is important toreveals the distribution of the adsorption molecules between the liquidphase and the solid phase when the adsorption process reaches anequilibrium state [33,53]. Therefore, the equilibrium trend related toSY and AZ adsorption onto MWCNT and MWCNT-Pd-NPs was fitted toLangmuir, Freundlich, Temkin and Dubinin–Radushkevich isotherms(see Table 5). It was found that R2 value of Langmuir is about 1.0 thatsuggest the favorable adsorption explain. This observation was shownthat that adsorption of SY and AZ dyes onto MWCNT and MWCNT-Pd-NPs adsorbents can only occur at a finite number of definite localizedsites, and the adsorbed layer is one molecule in thickness or monolayeradsorption. On the other hand, increase in Ka value with rising initialAZ and SY concentration and MWCNT and MWCNT-Pd-NPs massshow high tendency of AZ and SY adsorption.

3.7. Adsorption equilibrium

Various adsorption kinetics models like pseudo-first and second-

Fig. 4. Parity plot showing the correlation between the observed and predicted values (a)and Box Cox plots for R% of AZ and SY by MWCNT and MWCNT-Pd-NPs.

Table 3Quadratic model summary statistics and quality for R% of SY and AZ by MWCNT andMWCNT-Pd-NPs.

Statistics quality of model MWCNT-Pd-NPs MWCNT

R%SY R%AZ R%SY R%AZ

Std. Dev. 0.92 0.11 0.62 0.071Mean 73.97 86.85 65.61 76.86C.V.% 1.25 0.13 0.95 0.092PRESS 83.59 1.18 35.09 0.45R-Squared 0.9882 0.9998 0.9997 0.9997Adj R-Squared 0.9801 0.9996 0.9996 0.9996Predicated R-Squared 0.9602 0.9993 0.9992 0.9992Adeq Precision 38.083 311.389 352.841 352.841

Fig. 5. Response surface plot.

R. Karimi et al. Chemometrics and Intelligent Laboratory Systems 159 (2016) 127–137

134

order model, Elovich equation and intraparticle diffusion model wereapplied to investigate the mechanism of adsorption of understudy dyesonto both adsorbent (see Table 6) [14,51]. Each of the investigatedmodels on the basis of the corresponding determination coefficients(R2) and agreement between experimental and equilibrium adsorptioncapacity was fitted by a linear regression of models to the appropriate-ness of experimental data. R2 values for pseudo-second-order kineticmodel are about 1.0 and the calculated qe values are mainly close to theexperimental data. These two outcome is criterion for judgment aboutapplicability of second-order model for best prediction of adsorptiondata for AZ and SY adsorption onto MWCNT and MWCNT-Pd-NPsadsorbents.

4. Conclusion

In this work, MWCNT and MWCNT-Pd-NPs characterized by XRDand SEM successfully was applied as adsorbents for simultaneousremoval of SY and AZ dyes in aquaes solution. impact operationalparameters on the efficiency of adsorption process successfully wasinvestigated by CCD and ANN and optimal conditions from CCD in

presens of MWCNT was found at 6.01 and 9.98 mg L−1 AZ and SYconcentration, 0.02 of adsorbent mass and 7.0 for pH value and3.8 min sonication time, respectively, while in the presence ofMWCNT-Pd-NPs optimized values were obtained at 10.0 and

Table 4Summary of the input-output dataset characterization.

System Number of test ΔpH AZ concentration SY concentration Sonicated time Adsorbent mass %Removal (AZ) %Removal (SY)

AZ/SY on MWCNT 50 4–8 4–12 1–600 1–5 0.01–0.025 67.1–82.97 53.78–73.69AZ/SY on MWCNT-Pd-NPs 50 4–8 4–12 1–600 1–5 0.01–0.025 74.48–96.69 63.67–87.01

Fig. 6. Effect of the number of hidden layer neurons on AAD% of the AZ and SY dyesremoval by MWCNTs.

Fig. 7. Evolution of training, validation, and test errors as a function of the number oftraining epochs during ANN training.

Fig. 8. Modeling ability of the optimized ANN to predicate a) the AZ dyes removal byMWCNTs: (R2=0.911, AAD=0.27%), b) the SY dyes removal by MWCNTs: (R2=0.954,AAD=0.43%).

Fig. 9. AADs plot of the AZ and SY dyes removal by MWCNTs from experimental data.

R. Karimi et al. Chemometrics and Intelligent Laboratory Systems 159 (2016) 127–137

135

8.0 mg L−1 AZ and SY concentration, 0.018 of adsorbent mass and 5.0for pH value and 3.5 min sonication time, respectively. In theseconditions and at desirability value of 0.99, the maximum adsorptionefficiency for AZ onto MWCNT and MWCNT-Pd-NPs are 82.97 and94.95, respectively. Also, the maximum adsorption efficiency for SYonto MWCNT and MWCNT-Pd-NPs are 73.13 and 85.55, respectively.This values show that MWCNT-Pd-NPs in comparison with MWCNT atthe lower contact time and adsorbent mass was able to remove greateramounts of under study dyes. Based on the ANN model the absoluteaverage deviations (AADs) of AZ and SY dyes adsorption by MWCNTfrom experimental data are 0.27% and 0.43%, and the R2 values are0.911 and 0.954, respectively. Also, the AADs of AZ and SY dyesadsorption by MWCNTs-Pd-NPs are 0.44% and 0.65%, and the R2

values are 0.925% and 0.927%, respectively.

Novelity

In this article, we have focused on the decoration of multiwalledcarbon nanotubes (MWCNTs) with Pd nanoparticles. This material wascharacterized by different techniques such as XRD, SEM and TGA, andsubsequently were used for the adsorption of sunset yellow (SY) andazur (II) (AZ) dyes from aquaes solution. It was clearly demonstratedthat the adsorption dynamics of sunset yellow (SY) and azur (II) (AZ)dyes, the electronic interaction between adsorbents and adsorbates,and the structure and morphology of adsorbent are crucial factors forestablishing efficient adsorption reaction systems. Artificial neuralnetwork and central composite design used for building up to construct

an empirical model to predict the understudy dyes removal behavioronto these adsorbents and the obtained results have good agreementwith experimental data. Finally, it was found that the equilibriumisotherm of adsorption process follows the Langmuir isotherm.

Acknowledgements

The authors thanks of the Research Council of the Yasouj Universityfor financial supporting this work.

References

[1] L. Liu, Z.Y. Gao, X.P. Su, X. Chen, L. Jiang, J.M. Yao, Adsorption removal of dyesfrom single and binary solutions using a cellulose-based bioadsorbent, ACS Sustain.Chem. Eng. 3 (2015) 432–442.

[2] N. Cheng, Q. Hu, Y. Guo, Y. Wang, L. Yu, Efficient and selective removal of dyesusing imidazolium-based supramolecular gels, ACS Appl. Mater. Interfaces 7(2015) 10258–10265.

[3] S. Sansuk, S. Srijaranai, S. Srijaranai, A new approach for removing anionic organicdyes from wastewater based on electrostatically driven assembly, Environ. Sci.Technol. (2016).

[4] X.-S. Wang, J. Liang, L. Li, Z.-J. Lin, P.P. Bag, S.-Y. Gao, Y.-B. Huang, R. Cao, Ananion metal–organic framework with lewis basic sites-rich toward charge-exclusivecationic dyes separation and size-selective catalytic reaction, Inorg. Chem. 55(2016) 2641–2649.

[5] M.N. Amin, R.S. Blackburn, Sustainable chemistry method to improve the wash-offprocess of reactive dyes on cotton, ACS Sustain. Chem. Eng. 3 (2015) 725–732.

[6] A. Asfaram, M. Ghaedi, S. Agarwal, I. Tyagi, V.K. Gupta, Removal of basic dyeAuramine-O by ZnS: cu nanoparticles loaded on activated carbon: optimization ofparameters using response surface methodology with central composite design,RSC Adv. 5 (2015) 18438–18450.

[7] A. Asfaram, M. Ghaedi, A. Goudarzi, M. Rajabi, Response surface methodologyapproach for optimization of simultaneous dye and metal ion ultrasound-assistedadsorption onto Mn doped Fe 3 O 4-NPs loaded on AC: kinetic and isothermalstudies, Dalton Trans. 44 (2015) 14707–14723.

[8] A. Asfaram, M. Ghaedi, S. Hajati, M. Rezaeinejad, A. Goudarzi, M.K. Purkait, Rapidremoval of Auramine-O and Methylene blue by ZnS: cu nanoparticles loaded onactivated carbon: a response surface methodology approach, J. Taiwan Inst. Chem.Eng. 53 (2015) 80–91.

[9] Y. Yang, G. Hu, F. Chen, J. Liu, W. Liu, H. Zhang, B. Wang, An atom-scaleinterfacial coordination strategy to prepare hierarchically porous Fe 3 O 4–graphene frameworks and their application in charge and size selective dyeremoval, Chem. Commun. 51 (2015) 14405–14408.

[10] R. Jin, D. Song, H. Xiong, L. Ai, P. Ma, Y. Sun, Magnetic core/shell Fe3O4/Aunanoparticles for studies of quinolones binding to protein by fluorescencespectroscopy, Luminescence (2015).

[11] D. Robati, B. Mirza, M. Rajabi, O. Moradi, I. Tyagi, S. Agarwal, V.K. Gupta,Removal of hazardous dyes-BR 12 and methyl orange using graphene oxide as anadsorbent from aqueous phase, Chem. Eng. J. 284 (2016) 687–697.

[12] N.N. Nassar, N.N. Marei, G. Vitale, L.A. Arar, Adsorptive removal of dyes fromsynthetic and real textile wastewater using magnetic iron oxide nanoparticles:thermodynamic and mechanistic insights, Can. J. Chem. Eng. 93 (2015)1965–1974.

[13] A. Asfaram, M. Ghaedi, S. Hajati, A. Goudarzi, Ternary dye adsorption onto MnO 2nanoparticle-loaded activated carbon: derivative spectrophotometry and modeling,RSC Adv. 5 (2015) 72300–72320.

Table 5Isotherm constant parameters and correlation coefficients calculated for the dyes adsorption onto MWCNT and MWCNT-Pd-NPs.

Model Equation Parameters MWCNT MWCNT-Pd-NPs

AZ SY AZ SY

0/15 g 0/02 g 0/15 g 0/02 g 0/15 g 0/02 g 0/15 g 0/02 g

Langmuir Ce/qe=1/(KaQm)+Ce/Qm Qm (mg g−1) 46/76 36/10 30/58 36/76 63/97 74/99 36/101 36/63Ka (L mg−1) 1/75 8/66 8/6 1/86 1/50 2/74 6/022 6/66R2 0/996 ./999 0/993 0/990 99/26 99/57 0/9908 0/9819

Frendlich Ln qe=ln KF+(1/n) ln Ce 1/n 0/1184 0/958 0/2173 0/1969 4/06 4/54 0/4615 0/3142KF (L mg−1) 4/58 4/36 20/41 24/68 31/08 40/58 4/877 4/588R2 0/962 0/895 0/7988 0/507 93/00 88/67 0/9151 0/8501

Temkin qe=Bl ln KT+Bl lnCe Bl 4/4395 0/8558 4/7172 5/4213 33/725 32/249 8/437 7/280KT (L mg−1) 1863/1 3856/1 102/51 116/75 103/52 80/62 51/823 87/293R2 0/9450 0/910 0/7758 56/49 93/35 89/55 0/9751 0/9024

Dubinin-Radushkevich Ln qe=ln Qm–Kε2 Qs (mg g−1) 41/96 35/27 15/29 30.65 83/76 81/39 85/21 77/39B×10−7 0/4 −0/2 −0/09 −0/2 −2 −1 −1 −4E 3536 5000 7462/2 4081 1587/30 2272/72 2272/72 1123/59R2 0/783 0/993 0/926 0/991 0/9156 0/9099 0/9437 0/9172

Table 6Kinetic parameters for the adsorption of dyes onto MWCNT and MWCNT-Pd-NPs.

Model Parameters MWCNT MWCNT-Pd-NPs

SY AZ SY AZ

Pseudo-First-order k1(min−1) 0/944 0/887 0/084 0/251kinetic qe (calc) (mg g−1) 33/24 48/68 6/63 14/49

R2 0/661 0/820 0/877 0/9057Pseudo-Second-

order kinetick2 (min−1) 0/0177 0/0374 0/0111 0/005qe(calc) (mg g−1) 40/48 24/93 96/35 59/35R2 0/989 0/982 0/999 0/992

Intraparticle Kdiff (mg g-1 min−1/2)

11/304 6/49 1/34 6/49

diffusion C (mg g−1) 7/2465 7/160 10/77 4/35R2 0/992 0/979 0/898 0/973

Elovich β (g mg−1) 25/99 20/59 18/26 10/59α (mg g−1 min−1) 0/944 0/887 0/084 0/251R2 33/24 48/68 6/63 14/49qe 25/99 20/59 18/26 10/59

R. Karimi et al. Chemometrics and Intelligent Laboratory Systems 159 (2016) 127–137

136

[14] F.N. Azad, M. Ghaedi, K. Dashtian, S. Hajati, A. Goudarzi, M. Jamshidi, Enhancedsimultaneous removal of malachite green and safranin O by ZnO nanorod-loadedactivated carbon: modeling, optimization and adsorption isotherms, N. J. Chem. 39(2015) 7998–8005.

[15] W. Fang, J. Yang, J. Gong, N. Zheng, Photo‐and pH‐triggered release of anticancerdrugs from mesoporous silica‐coated Pd@ Ag nanoparticles, Adv. Funct. Mater. 22(2012) 842–848.

[16] A.A. Rajabi, Y. Yamini, M. Faraji, F. Nourmohammadian, Modified magnetitenanoparticles with cetyltrimethylammonium bromide as superior adsorbent forrapid removal of the disperse dyes from wastewater of textile companies,Nanochem. Res. 1 (2016) 49–56.

[17] M. Ghaedi, R. Hassani, K. Dashtian, G. Shafie, M.K. Purkait, H. Dehghan,Adsorption of methyl red onto palladium nanoparticles loaded on activated carbon:experimental design optimization, Desalination Water Treat. (2016) 1–9.

[18] M. Jamshidi, M. Ghaedi, K. Dashtian, A. Ghaedi, S. Hajati, A. Goudarzi,E. Alipanahpour, Highly efficient simultaneous ultrasonic assisted adsorption ofbrilliant green and eosin B onto ZnS nanoparticles loaded activated carbon:artificial neural network modeling and central composite design optimization,Spectrochim. Acta Part A: Mol. Biomol. Spectrosc. 153 (2016) 257–267.

[19] A. Sarve, S.S. Sonawane, M.N. Varma, Ultrasound assisted biodiesel productionfrom sesame (Sesamum indicum L.) oil using barium hydroxide as a heterogeneouscatalyst: comparative assessment of prediction abilities between response surfacemethodology (RSM) and artificial neural network (ANN), Ultrason. Sonochem. 26(2015) 218–228.

[20] F.N. Azad, M. Ghaedi, A. Asfaram, A. Jamshidi, G. Hassani, A. Goudarzi,M.H.A. Azqhandi, A. Ghaedi, Optimization of the process parameters for theadsorption of ternary dyes by Ni doped FeO (OH)-NWs–AC using response surfacemethodology and an artificial neural network, RSC Adv. 6 (2016) 19768–19779.

[21] M. Ghaedi, N. Zeinali, M. Maghsoudi, M. Purkait, Artificial neural network (ANN)Method for modeling of sunset yellow dye adsorption using nickel sulfidenanoparticle loaded on activated carbon: kinetic and isotherm study, J. Dispers.Sci. Technol. 36 (2015) 1339–1348.

[22] M. Maghsoudi, M. Ghaedi, A. Zinali, A. Ghaedi, M. Habibi, Artificial neural network(ANN) method for modeling of sunset yellow dye adsorption using zinc oxidenanorods loaded on activated carbon: kinetic and isotherm study, Spectrochim.Acta Part A: Mol. Biomol. Spectrosc. 134 (2015) 1–9.

[23] J.P.C. Kleijnen, Response Surface Methodology, Handbook of SimulationOptimization, Springer, New York, 2015, pp. 81–104.

[24] K.A.M. Said, M.A.M. Amin, Overview on the response surface methodology (RSM)in extraction processes, J. Appl. Sci. Process Eng. 2 (2015).

[25] M. Kolaei, K. Dashtian, Z. Rafiee, M. Ghaedi, Ultrasonic-assisted magnetic solidphase extraction of morphine in urine samples by new imprinted polymer-supported on MWCNT-Fe 3 O 4-NPs: central composite design optimization,Ultrason. Sonochem. 33 (2016) 240–248.

[26] A.V. Schenone, L.O. Conte, M.A. Botta, O.M. Alfano, Modeling and optimization ofphoto-Fenton degradation of 2, 4-D using ferrioxalate complex and responsesurface methodology (RSM), J. Environ. Manag. 155 (2015) 177–183.

[27] A. Lee, Z.W. Geem, K.-D. Suh, Determination of optimal initial weights of anartificial neural network by using the harmony search algorithm: application tobreakwater armor stones, Appl. Sci. (6) (2016) 164.

[28] J.J. Cheh, R.S. Weinberg, K.C. Yook, An application of an artificial neural networkinvestment system to predict takeover targets, J. Appl. Bus. Res. (JABR) 15 (2013)33–46.

[29] K. dashtian, F. nasiri azad, M. Ghaedi, A. jamshidi, G. Hassani,M. Montazerozohori, S. Hajati, M. Rajabi, A.A. Bazrafshan, , Preparation andcharacterization of AC-Fe3O4-Au hybrid for simultaneous removal of Cd2+, Pb2+,Cr3+ and Ni2+ ions from aquaes solution via complexation with 2-((2, 4-Dichloro-benzylidene)-amino)-benzenethiol: taguchi optimization, RSC Adv. (2016).

[30] M. Jamshidi, M. Ghaedi, K. Dashtian, S. Hajati, New ion-imprinted polymer-functionalized mesoporous SBA-15 for selective separation and preconcentration ofCr (iii) ions: modeling and optimization, RSC Adv. 5 (2015) 105789–105799.

[31] M. Jamshidi, M. Ghaedi, K. Dashtian, S. Hajati, A. Bazrafshan, Ultrasound-assistedremoval of Al 3+ ions and Alizarin red S by activated carbon engrafted with Agnanoparticles: central composite design and genetic algorithm optimization, RSCAdv. 5 (2015) 59522–59532.

[32] F.N. Azad, M. Ghaedi, K. Dashtian, M. Montazerozohori, S. Hajati,

E. Alipanahpour, Preparation and characterization of MWCNTs functionalized byN-(3-nitrobenzylidene)-N′-trimethoxysilylpropyl-ethane-1, 2-diamine for the re-moval of aluminum (iii) ions via complexation with eriochrome cyanine R:spectrophotometric detection and optimization, RSC Adv. 5 (2015) 61060–61069.

[33] M. Ghaedi, R. Hassani, K. Dashtian, G. Shafie, M. Purkait, H. Dehghan, Adsorptionof methyl red onto palladium nanoparticles loaded on activated carbon: experi-mental design optimization, Desalination Water Treat. (2016) 1–9.

[34] S. Mosleh, M.R. Rahimi, M. Ghaedi, K. Dashtian, Sonophotocatalytic degradationof trypan blue and vesuvine dyes in the presence of blue light photocatalyst of Ag 3PO 4/Bi 2 S 3-HKUST-1-MOF: central composite optimization and synergisticeffect study, Ultrason. Sonochem. (2016).

[35] M. Soleiman, M.R. Rahimi, M. Ghaedi, K. Dashtian, S. Hajati, Photocatalyticdegradation of binary mixture of toxic dyes by HKUST-1 MOF and HKUST-1-SBA-15 in rotating packed bed reactor under blue LED illumination: central compositedesign optimization, RSC Adv. (2016).

[36] S. Haykin, N. Network, A comprehensive foundation, Neural Netw. 2 (2004).[37] P. Xu, S. Xu, H. Yin, Application of self-organizing competitive neural network in

fault diagnosis of suck rod pumping system, J. Pet. Sci. Eng. 58 (2007) 43–48.[38] B. Vaferi, Y. Rahnama, P. Darvishi, A. Toorani, M. Lashkarbolooki, Phase equilibria

modeling of binary systems containing ethanol using optimal feedforward neuralnetwork, J. Supercrit. Fluids 84 (2013) 80–88.

[39] B. Vaferi, M. Karimi, M. Azizi, H. Esmaeili, Comparison between the artificialneural network, SAFT and PRSV approach in obtaining the solubility of solidaromatic compounds in supercritical carbon dioxide, J. Supercrit. Fluids 77 (2013)44–51.

[40] M.T. Vafaei, R. Eslamloueyan, S. Ayatollahi, Simulation of steam distillationprocess using neural networks, Chem. Eng. Res. Des. 87 (2009) 997–1002.

[41] C.M. Bishop, Neural Networks for Pattern Recognition, Oxford university press,1995.

[42] B.D. Ripley, Pattern Recognition and Neural Networks, Cambridge university press,2007.

[43] P.P. Van Der Smagt, Minimisation methods for training feedforward neuralnetworks, Neural Netw. 7 (1994) 1–11.

[44] M. Tuchler, A.C. Singer, R. Koetter, Minimum mean squared error equalizationusing a priori information, IEEE Trans. Signal Process. 50 (2002) 673–683.

[45] H. Konno, H. Yamazaki, Mean-absolute deviation portfolio optimization model andits applications to Tokyo stock market, Manag. Sci. 37 (1991) 519–531.

[46] N.J.D. Nagelkerke, A note on a general definition of the coefficient of determina-tion, Biometrika 78 (1991) 691–692.

[47] G.B. Celli, A. Ghanem, M.S.-L. Brooks, Optimization of ultrasound-assistedextraction of anthocyanins from haskap berries (Lonicera caerulea L.) usingResponse Surface Methodology, Ultrason. Sonochem. 27 (2015) 449–455.

[48] J. Chen, Z. Guo, H.-B. Wang, M. Gong, X.-K. Kong, P. Xia, Q.-W. Chen,Multifunctional Fe 3 O 4@ C@ Ag hybrid nanoparticles as dual modal imagingprobes and near-infrared light-responsive drug delivery platform, Biomaterials 34(2013) 571–581.

[49] B. Bhardwaj, R. Kumar, P.K. Singh, An improved surface roughness predictionmodel using Box-Cox transformation with RSM in end milling of EN 353, J. Mech.Sci. Technol. 28 (2014) 5149–5157.

[50] M.D. Turan, H.S. Altundoğan, Leaching of copper from chalcopyrite concentrate byusing ammonium persulphate in an autoclave: determination of most suitableimpeller type by using response surface methodology, J. Cent. South Univ. 20(2013) 622–628.

[51] F.N. Azad, M. Ghaedi, K. Dashtian, S. Hajati, V. Pezeshkpour, Ultrasonicallyassisted hydrothermal synthesis of activated carbon–HKUST-1-MOF hybrid forefficient simultaneous ultrasound-assisted removal of ternary organic dyes andantibacterial investigation: taguchi optimization, Ultrason. Sonochem. 31 (2016)383–393.

[52] F.N. Azad, M. Ghaedi, K. Dashtian, A. Jamshidi, G. Hassani, M. Montazerozohori,S. Hajati, M. Rajabi, A.A. Bazrafshan, Preparation and characterization of an AC–Fe 3 O 4–Au hybrid for the simultaneous removal of Cd 2+, Pb 2+, Cr 3+ and Ni 2+ions from aqueous solution via complexation with 2-((2, 4-dichloro-benzylidene)-amino)-benzenethiol: taguchi optimization, RSC Adv. 6 (2016) 19780–19791.

[53] S. Hosseini, H. Aghaie, M. Ghaedi, K. Dashtian, M. Purkait, [Lead (II) adsorptionfrom aqueous solutions onto modified ag nanoparticles: modeling and optimiza-tion, Environ. Progress. Sustain. Energy (2016).

R. Karimi et al. Chemometrics and Intelligent Laboratory Systems 159 (2016) 127–137

137