Biosorption of cobalt(II) with sunflower biomass from aqueoussolutions in a fixed bed column and neural networks modelling

Ensar Oguz n, Muhammed ErsoyAtatürk University, Environmental Engineering Department, 25240 Erzurum, Turkey

a r t i c l e i n f o

Article history:Received 18 June 2013Received in revised form25 September 2013Accepted 3 October 2013Available online 5 November 2013

Keywords:BiosorbentCobalt(II)SunflowerNeural networkFixed bed column

a b s t r a c t

The effects of inlet cobalt(II) concentration (20–60 ppm), feed flow rate (8–19 ml/min) and bed height(5–15 cm), initial solution pH (3–5) and particle size (0.25oxo0.5, 0.5oxo1 and 1oxo2 mm) on thebreakthrough curves were investigated. The highest bed capacity of 11.68 mg/g was obtained using 40 ppminlet cobalt(II) concentration, 5 cm bed height and 8 ml/min flow rate, pH 6.5 and 0.25oxo0.5 mm particlesize. According to the BET (N2) measurements, the specific surface area of the shells of sunflower biomasswas found to be 1.82 m2/g. A relationship between the predicted results of the ANN model and experimentaldata was conducted. The ANN model yielded determination coefficient of (R2 0.972), standard deviation ratio(0.166), mean absolute error (0.0158) and root mean square error (0.0141). The results indicated that theshells of the sunflower biomass is a suitable biosorbent for the uptake of cobalt(II) in fixed bed columns.

& 2013 Elsevier Inc. All rights reserved.

1. Introduction

Cobalt compounds are widely used in many industrial applica-tions such as mining, metallurgical, electroplating, paints, pigmentsand electronic (Manohar et al., 2006). Cobalt is also present in thewastewater of nuclear power plants. Toxic heavy metals dischargedfrom the different industrial activities constitute one of the majorcauses of water pollution. Heavy metal ions in contaminated habitatsmay accumulate in microorganisms, aquatic flora and fauna, mayenter into the food chain and lead to health problems. The permis-sible limits of cobalt in the irrigation water and livestock wastewaterare 0.05 and 1.0 ppm, respectively. Higher concentration of cobaltmay cause low blood pressure, paralysis, diarrhoea, lung irritationand bone defects (Gupta et al., 2012).

The uptake of toxic metals ions fromwastewater has attracted agreat deal of attention in last decade for global awareness of thedetriment of toxic metals in the environment. Application oftraditional processes for the uptake of toxic metals carry enormouscosts, which become impracticable and uneconomical because ofcontinuous input of chemicals, causing further environmentaldamage. Therefore, easy, effective, economic and eco-friendlytechniques are required for effluent treatment (Han et al., 2006).

Several methods such as sulphide precipitation, coagulation andion-exchange have been used to remove heavy metals from waste-water, although they have been found to have limited application.

Adsorbents such as fly ash (Nascimento et al., 2009; Papandreouet al., 2011), coal (Xiaobing et al., 2010), alginate beads (Bayramogluet al., 2006), certain clays and clay-minerals (Adebowale et al., 2008;Eloussaief et al., 2009) have been used to uptake heavy metals fromwastewater. The searches for low cost and easily available adsorbentshave led to the investigation of materials of agricultural andbiological origin along with industrial by-products (Baccar et al.,2013; Sihem et al., 2012).

Biosorption, which is an alternative process, is the uptake ofheavy metals from aqueous solutions by biological materials. Thisapproach is competitive, effective and cheap (Volesky, 2000). Bio-sorption of metals by biomass has been greatly explored in recentyears. Different forms of inexpensive, non-living plant material suchas rice husk (Krishnani et al., 2008; Kausara et al.,2013), Coriolusversicolor (Bhatti et al., 2013), waste biomass (Bhatti et al., 2010, 2011,2013), sawdust (Augustine, 2010), and pine bark and canola meal(Gundogdu et al., 2009; Hansen et al., 2010) have been widely usedto remove the heavy metals from wastewater, and Scots pine coneshave been recently investigated as potential biosorbents for heavymetals (Nuhoglu and Oguz, 2003).

A continuous fixed bed adsorber does not run under equilibriumconditions and the effect of flow condition at any cross-section in thecolumn affects the flow behaviour (Singh et al., 2009). In order todesign and operate a fixed-bed biosorption process successfully, thebreakthrough curves under specified operating conditions must bepredictable. The shape of this curve is influenced by the transportprocess in the column (Vazquez et al., 2006). Breakthrough deter-mines bed height and the operating life span of the bed (Walker andWeatherley, 1997).

Contents lists available at ScienceDirect

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

Ecotoxicology and Environmental Safety

0147-6513/$ - see front matter & 2013 Elsevier Inc. All rights reserved.http://dx.doi.org/10.1016/j.ecoenv.2013.10.004

n Corresponding author.E-mail address: eoguz@atauni.edu.tr (E. Oguz).

Ecotoxicology and Environmental Safety 99 (2014) 54–60

The classical optimization method is not only time-consumingand tedious but also does not depict the complete effects of theparameters in the process and ignores the combined interactionsbetween physicochemical parameters. This method can also leadto misinterpretation of results (Montgomery, 2008). To overcomethis difficulty, some statistical methods have been used. In recentyears artificial neural networks (ANNs) have been widely studiedto solve environmental problems because of their reliable andsalient characteristics in capturing the non-linear relationshipsexisting between variables (Turan et al., 2011).

The ANNs approach seems to be completely suitable for theproblems where the relation among variables is not linear andcomplex ANNs are directly inspired by the biology of the humanbrain, where billions of neurons are interconnected to process avariety of complex types of information. Accordingly, a computa-tional neural network consists of simple processing units calledneurons. Each neuron (a processing element) is linked to itsneighbours with varying strengths. The strength of connectionbetween two neurons is called weight, and it is represented bycoefficients of connectivity (Bernard et al., 2004).

The number of input and output neurons is determined by thenature of the problem. The hidden layers act like specific detectors.In theory, there can be more than one hidden layer. Universalapproximation theory suggests that a network with a single hiddenlayer with a sufficiently large number of neurons can interpret anyinput–output structure (Oguz et al., 2008a,2008b).

Input neurons accept the input data characterising a givenobservation (experiment). Output neurons yield the predicted(expected) value. A neuron sums the product of each connectionweight (wjk) from a neuron (j) to the neuron (k) and input (xj) andthe additional weight called the bias to get the value of sum for theneuron. The ith neuron has a gatherer that gathers its weightedinput wij � xj and the bias bi to form its net input ni given by Eq. (1).

ni ¼∑wijxjþbi ð1Þwhere wij denotes the strength of the connection from the jth inputto the ith neuron, xj is the input vector, bi is the ith neuron bias. Thesum of the weighted inputs is further transformed with a transferfunction to get the output value, there are several transfer functions;the most common is the sigmoidal function. To find suitable ws andbiases for each neuron, process training is essential. As the first stepof building an ANN, training means that the weights are corrected toproduce pre-specified target values (Zupan and Gasteiger, 1999).

In the present study, sunflower biomass has been firstly used as anew and cheap biosorbent to explore its potential for the removal ofcobalt(II) from aqueous solution. The breakthrough curves of thesunflower biomass for cobalt(II) have been investigated as a functionof biosortion time, inlet cobalt(II) concentration, feed flow rate, bedheight, initial solution pH and particle size. The model based on theartificial neural networks (ANNs) was conducted to predict the cobalt(II) concentrations removed from aqueous solution. A consistentrelationship between the predicted results of the designed ANNmodel and experimental data was observed.

2. Materials and methods

2.1. Biosorbent preparation

The sunflower biomass used as a biosorbent was gathered in the month ofAugust. The sunflower biomass was dried in the open air for 96 h and cut into smallpieces, ground in a blender and sieved to separate them into different particle sizes(0.25oxo0.5, 0.5oxo1 and 1oxo2 mm).

The surface area of the sunflower biomass was measured by the BET method at77 K using a Quantachrome QS-17 model apparatus (Brunauer et al., 1938). Thesurface area of the sunflower biomass was defined as 1.8 m2/g.

Cobalt(II) solutions were prepared by diluting 400 ppm of CoCl2 �6H2O (Merck)stock solution with deionized water to a desired concentration range between 20

and 60 ppm. The initial concentration of the cobalt(II) in the solution and samplesafter biosorption process were complexometrically determined.

2.2. Comlexometric determination of cobalt(II)

Neutralize a sample of 5 ml and complete to 25 ml volume with distilled water.Add 0.5 ml of 2 M HCl and 100 mg of the indicator (1 wt% of murexid in CaCI, finelymilled). Drop ammonia (2 M) into the red–pink solution until it just becomesyellow. Dilute with water to 50 ml volume (pH¼7.5–8.5) and titrate with 0.01 MEDTA solution from yellow to pink. A volume of 1 ml, 0.01 M EDTA solution isequivalent to 0.5859 mg cobalt(II) (Karge and Weitkamp, 2006).

2.3. Column experiments

Continuous flow biosorption experiments were conducted in Teflon columns of1 cm i.d. and 5, 10 and 15 cm heights. A known quantity of biosorbent was placed inthe column to receive the desired bed height. Cobalt(II) solution with an initialconcentration of 40 ppm was pumped upward through the column at a desiredflow rate by a peristaltic pump. Samples were collected from the exit of the columnat different intervals and analysed for cobalt(II) concentration. Operation of thecolumn was stopped when the effluent cobalt(II) concentration equals influentcobalt(II) concentration. The total quantity of metal biosorbated in the column wascalculated from the area above the breakthrough curve (outlet metal concentrationversus time) multiplied by the flow rate. The total amount of metal ions sent to thecolumn can be calculated from Eq. (2).

mtotal ¼CoFte1000

ð2Þ

where Co is the inlet metal ion concentration (ppm), F the volumetric flow rate(ml/min) and te is the exhaustion time (min). The mass transfer zone can be calculatedfrom the difference between the column exhaustion time (te) and column break-through time (tb). The slope of the breakthrough curve from tb to te was represented bydc/dt. Total metal removal (%) with respect to flow volume can be calculated from theratio of metal mass adsorbed (mad) to the total amount of metal ions sent to thecolumn by Eq. (3):

total metal removal ð%Þ ¼ mad

mtotal� 100 ð3Þ

The amount of metal retained in the column depends on the influent metalconcentration and can be calculated from the area above the breakthrough curve(Eq. (4)):

q¼ CoQm� 1000

Z t

01� Ct

Co

� �dt ð4Þ

where q represents the amount of metal retained (mg cobalt(II)/g adsorbent),Ct and Co are the cobalt concentrations at the column effluent and influent (ppm)respectively, Q is the flow rate (ml/min), m the mass of adsorbent in the column(g) and t is biosorption time (min).

2.4. Data used in the artificial neural networks

The number of experimental data used in the ANN is 294 which divided into thethree sections: the training set (149 data), verification set (74 data) and test set(74 data). Training algorithms do not use the verification or test sets to adjust networkweights. The verification set may optionally be used to track the network's errorperformance to identify the best network and to stop training if over-learning occurs.The test set is not used in training at all, and it is designed to give an independentassessment of the network's performance when an entire network design procedure iscompleted. Fifty percent (50%) of the data set was used to train the network, while theremaining 50% of the data was employed for both verification and testing. Theassignment of cases to the training, verification and test subsets can sometimes affectthe performance of training algorithms. In order to eliminate this situation, the casesshould be shuffled randomly between the subsets. The cases can be left in theiroriginal order, or grouped together in the subsets. In this model, the cases wereshuffled randomly between subsets (training, test and verification).

The training requires sets of pairs (XS, YS) for input: the actual input into thenetwork is a vector (XS), and the corresponding target is labelled (YS) aftersuccessful training. When correct values of YS for each vector of XS from thetraining set are obtained, it is hoped that the network will give correct predictionsof Y for any new object of X, according to the ANN model fundamentals, with use ofmore data for training the network, a better result would be obtained. The mostutilised training method for multi-layered neural network is called back propaga-tion. In this study, one hidden layer was used. Information about errors (differencesbetween target and predicted values) is filtered back through the system and isused to adjust the connections between the layers, thus performance improves.In the early standard algorithm, random initial set of weights were assigned to theneural network. Then, by considering the input data, weights were adjusted, so theoutput error would be its minimum (Pazourek et al., 2005).

E. Oguz, M. Ersoy / Ecotoxicology and Environmental Safety 99 (2014) 54–60 55

3. Results

3.1. Effect of influent cobalt(II) concentration on breakthrough curve

The effect of influent cobalt(II) concentration on the shape ofthe breakthrough curves is shown in Fig. 1. As shown in Fig. 1,in the interval of 30 min, the value of Ct/Co reached 0.7, 0.82 and 0.89when influent concentration was 20, 40 and 60 ppm, respectively.

It is illustrated that the breakthrough time decreased withincreasing influent cobalt(II) concentration. At lower influent cobalt(II) concentration, the breakthrough curve was dispersed and break-through occurred slowly. As the influent concentration increased,sharper breakthrough curves were received. These results demon-strate that the change in the concentration gradient affects thesaturation rate and breakthrough time. The larger the influentconcentration, the steeper the slope of breakthrough curve andsmaller the breakthrough time. This can be explained by the factthat more biosorption sites were being covered with the cobalt(II)ions. These results demonstrate that the change of concentrationgradient affects the saturation rate of biosorbent and breakthroughtime, or in other words, the diffusion process is concentrationdependent. As the influent concentration increases, cobalt(II) loadingrate increases, the driving force also increases for mass transfer,which is a decrease in the biosorption zone length (Goel et al., 2005).

3.2. Effect of flow rate on breakthrough curve

The effect of flow rate on the biosorption of cobalt(II) in thefixed-bed with bed depth of 5 cm was investigated. The flow ratewas changed in the range of 8–19 ml/min while the concentrationof cobalt(II) in influent was kept at 40 ppm. The biosorptionbreakthrough curves obtained at different flow rates are shownin Fig. 2, and for the breakthrough time in the interval of 30 min,the Ct/Co values are given in Fig. 2. The results showed that thebiosorption of cobalt(II) on the sunflower biomass was influencedby the flow rate. All the breakthrough curves had a similar shape.The reduction in the cobalt(II) uptake capacity at higher flow ratesis probably due to the unavailability of sufficient retention time forsolute to interact with the sorbent and the limited diffusivity ofsolute into the sorptive sites or pores. The same findings have been

reported elsewhere (Hasfalina et al., 2010; Futalan et al., 2011;Mondal, 2009; Salamatinia et al.,2008).

The increase of the Ct/Co values from 0.82 to 0.94 for cobalt(II) wasprobably due to insufficient contact time for the metal solution tointeract with the biosorbent. In this case, biosorbate left columnwithout sufficient time to diffuse into pores of the adsorbents. Thus,it results in sharper breakthrough curves as shown in Fig. 2.

Fig. 2 shows that cobalt(II) concentration in the effluentincreased rapidly after the breakthrough time; as the solutioncontinued to flow, the fixed-bed became saturated with cobalt(II),and cobalt(II) concentration in the effluent approached the influ-ent concentration. Equilibrium uptake of cobalt(II) decreased withthe increasing flow rate, and their maximum values were obtainedat the lowest flow rate of 8 ml/min. As shown in Fig. 2, in theinterval of 30 min, the value of Ct/Co reached 0.82, 0.91 and 0.94when flow rate was 8, 14 and 19 ml/min, respectively.

3.3. Effect of bed depth on breakthrough curve

Another important parameter in the fixed bed process is relatedto the bed depth. However, because of the pressure drop and thehandling problems of the smaller particle sizes (o0.5oxo1 mm) inthe column studies, the particle size of 0.5oxo1 mmwas only usedfor the bed depths of 5, 10 and 15 cm. The biosorption performanceof the sunflower biomass was tested at various bed depths such as1.27 g (5 cm), 2.54 g (10 cm) and 3.81 g (15 cm) at a flow rate of 8 ml/min, pH 6.5, particle size of 0.5oxo1 mm and the cobalt(II)concentration of 40 ppm.

Among other factors, the retention of metals in a fixed-bedcolumn depends on the quantity of solid biosorbent used, or onthe bed depth the column works. Fig. 3 shows the breakthroughprofile of cobalt(II) at different bed depths. For the three differentbed depths used, as the bed depth increases, the quantity ofremoved cobalt(II) increases, which is also illustrated by theservice time change. As shown in Fig. 3, in the interval of30 min, the value of Ct/Co reached 0.82, 0.58 and 0.43 when thebed depth was 5, 10 and 15 cm, respectively. In the column studiesat the height of 5 cm, the biosorbent becomes saturated veryquickly because of the fact that the cobalt(II) binding on the sites isfaster.

Fig. 1. Experimental and predicted breakthrough curves of cobalt(II) as a function of inlet cobalt(II) concentration (T 20 1C, pH 5, flow rate 8 ml/min, bed depth 5 cm, particlesize 0.5oxo1 mm).

E. Oguz, M. Ersoy / Ecotoxicology and Environmental Safety 99 (2014) 54–6056

3.4. Effect of pH on breakthrough curve

The pH value of the solution is an important controlling parameterin the fixed bed column process, and the pH value of the aqueoussolution has more influence to uptake cobalt(II) ions in the fixed bed.It influences both the adsorbent surface metal binding sites and themetal chemistry in water. The pH of the feed solution changed from3 to 5. In cobalt(II) biosorption by the sunflower biomass, the highestadsorbent capacity and the longest breakthrough time was receivedat pH 5. As shown in Fig. 4, in the interval of 30 min, the value of Ct/Coreached 0.94, 0.91 and 0.82 when the pH value of the solution was 3,4 and 5, respectively.

At pH (3–5) there are three species present in the solution assuggested by (Krishnan and Anirudhan, 2008). The dominantspecies between pH 4 and 6.5 in the cobalt(II) solution wereCoþ þ and Co(OH)þ , respectively. These species are biosorbateda chemical and electrostatical interaction on the surface of thesunflower biomass. As the pH decreased, the surface of thesunflower biomass exhibits an increasing positive characteristic,therefore the breakthrough time decreased. Obviously, with an

increase of pH in the influent, the breakthrough curves shiftedfrom left to right, which indicates that more metal ions can beremoved.

The biosorption of cobalt(II) on the biosorbent particles wereanalyzed using Langmuir, Freundlich and D–R isotherm models.The D–R isotherm is more general than the Langmuir andFreundlich, because it does not assume a homogeneous surfaceor constant adsorption potential (Ho et al.,2002). The D–R equa-tion is given by Eq. (5):

ln qe ¼ ln qm�Kε2 ð5Þ

where qe is the amount of cobalt(II) biosorbated at the equilibrium,K is a constant related to the biosorption energy, qm is thetheoretical saturation capacity, ε is the Polanyi potential, equalto RT ln (1þ(1/Ce)). The values of qm and K were deduced byplotting ln qe versus ε2. D–R isotherm was found suitable becauseof the high correlation coefficient (R2 0.98). Isotherm studiesindicated that biosorption of cobalt(II) on the sunflower biomasscould not involve different mechanisms. The mean energy of

Fig. 2. Experimental and predicted breakthrough curves of cobalt(II) as a function of flow rate (T 20 1C, pH 5, Co 40 ppm, bed depth 5 cm, particle size 0.5oxo1 mm).

Fig. 3. Experimental and predicted breakthrough curves of cobalt(II) as a function of bed depth (T 20 1C, pH 5, Co 40 ppm, flow rate 8 ml/min, particle size 0.5oxo1 mm).

E. Oguz, M. Ersoy / Ecotoxicology and Environmental Safety 99 (2014) 54–60 57

adsorption (E) can be calculated from Eq. (6):

E¼ 1ffiffiffiffiffiffiffi2K

p ð6Þ

The magnitude of E is useful for estimating the type of biosorptionprocess. It was found to be 20 kJ mol�1, which is bigger then theenergy range of adsorption reactions, 8–16 kJ mol�1. The type ofbiosorption of cobalt(II) on the sunflower biomass was defined aschemical sorption according to (Oguz, 2005, 2007; Bering et al.,1972).

3.5. Effect of particle size on the performance of breakthrough curve

Another important parameter in the fixed bed process isrelated to the particle size of the biosorbent. The particle sizeswere 0.25oxo0.5, 0.5oxo1 and 1oxo2 mm, while the beddepth, influent cobalt(II) concentration and pH were kept at 5 cm,40 ppm and 5 ppm, respectively. The breakthrough curves con-cerning particles are given in Fig. 5. An increase in the particle sizeappeared to increase the sharpness of the breakthrough curve.Furthermore, the breakthrough values (Ct/Co) for the larger particlesize is higher than the smaller particles. A rapid increase in thebreakthrough values was observed with an increase from0.25oxo0.5 to 1oxo2 mm in particle size.

The increase in the breakthrough values (Ct/Co) with theincrease of the particle size was due to the lower surface area ofthe bigger particle sizes. The same behavior was also observed byWalker and Weatherley (1997). As the particle size increase, thethickness of stagnant film around the particles increases, and alsothe total length of the path inside the pore. Under these condi-tions, the overall kinetics of the process is low, because the timefor a molecule of biosorbate to reach the biosorption site is more,as the diffusion path along the pores is large. The strong effect ofparticle size also confirmed that intraparticle diffusion had asignificant impact on column performance. A smaller particle sizewill have a faster pore diffusion rate because the diffusion path isshorter and the resistance to diffusion is lower (Ouvrard et al.,2002; Al-Ghouti et al., 2007). As shown in Fig. 5, in the interval of30 min, the value of Ct/Co reached 0.82, 0.58 and 0.43 when thebed depth was 5, 10 and 15 cm, respectively.

4. Artificial neural network

In this study, a one-layered back propagation neural networkwas used for modelling of the uptake of cobalt(II) ions from aqueoussolutions. In the present work, the input variables to the neuralnetwork are as follows: the treatment time (t), the concentration ofinitial cobalt(II), biosorbent dosage, pH, bed depth and particle size.Cobalt(II) concentration as a function of reaction time was chosen asthe experimental response or output variable. In order to model thecobalt(II) concentrations with ANN, the statistica software pro-gramme was used. The coefficient of determination (R2), the rootmean square error (RMSE), the standard deviation ratio (SDR), andthe mean absolute error (MAE) are the main criterions that are usedto evaluate the performance of ANN, they are defined as follows:

R¼ n ∑obs�pre� �� ∑obs

� �∑pre� �� �

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffin∑obs2� ∑obs

� �2h i� n∑pre2� ∑pre

� �2h ir ð7Þ

RMSE¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiobs�preð Þ2

n

sð8Þ

SDR¼ffiffiffiffiffiffiffiffin∑

pobs2� ∑obs

� �2=n n�1ð Þffiffiffiffiffiffiffiffi

n∑p

error2� ∑error� �2

=nðn�1Þð9Þ

MAE¼ ∑jobs�prejn

ð10Þ

where n is the number of data. Low values of RMSE, SDR and MAEsatisfy the statistical evaluation of prediction for the validation(Celik and Tan, 2005; Tortum, 2003). It was defined that the ANNmodel has a determination coefficient of (0.960), SDR of (0.192),RMSE of (0.41) and MAE of (1.447). The results obtained in thismodel indicate that the ANN model has the ability to predict theremoval of cobalt(II) concentration.

Before the training of the network, both input and outputvariables were normalized within the range 0–1 using theminimax-algorithm. The minimum and maximum of the data setwere found and scaling factors were selected so that these weremapped to desired minimum and maximum values. The minimax

Fig. 4. Experimental and predicted breakthrough curves of cobalt(II) as a function of pH of solution (T 20 1C, bed depth 5 cm, Co 40 ppm, flow rate 8 ml/min, particle size0.5oxo1 mm).

E. Oguz, M. Ersoy / Ecotoxicology and Environmental Safety 99 (2014) 54–6058

algorithm is given by Eq. (11)

H′¼H1

HMAX�HMIN

� HMIN

HMAX�HMIN

ð11Þ

where Hmax and Hmin, respectively indicate the largest and smallestvalues of H, and H' the unified value of the correspondingH. Normalisation of the data greatly improves learning speedand it is beneficial in reducing the error of the trained network.

The weight coefficients and the biases are the values obtainedfor the normalised data, in order to simulate the actual (experi-mental) cobalt(II) concentration. An inverse transformation on thisdata must be performed using shift and scale factors. After longtraining phases, the best result was obtained from the Levenberg–Marquardt algorithm. The hyperbolic tangent function in thehidden layer and the linear activation function in the output layerwere used in the model. It was observed that the optimal networkwas found to be seven inputs, one hidden layer with five neuronsand one output layer.

Sensitivity analysis is a technique used to assess the relativecontribution of the input variables to the performance of a neuralnetwork by testing the neural network when input variables areunavailable. This indicates that the input variables are important byparticular neural network. If the ratio is one or lower, this makes thevariable unavailable either have no effect on the performance of thenetwork, or actually enhance it, because ratios of input parameters ofthe model are more than one, all input variables are meaningful.

The most important parameters that affect the removal ofconcentration of cobalt(II) are; biosorption time, bed depth (Z),biosorbent dosage (M), particle size (P.S), initial cobalt(II) concen-tration, flow rate (Q) and pH, respectively. As the result of thisstudy, the general equation obtained from the optimal networkwas given as the following:

Cþ þt ¼ f 2 w2f 1 w1

t

Cþ þo

M

Vrpm

pH

T 1C

PS

2666666666664

3777777777775þb1

2666666666664

3777777777775þb2

2666666666664

3777777777775

Under different experimental conditions along the reactiontime, the interpretation of the experimental results was basedon the fitting of neural network models for predicting the removalof concentration of cobalt(II). The proposed model based on theANN could predict the cobalt(II) concentration during biosorptiontime. A comparison between the predicted and the observed datawas conducted. According to the ANN model fundamentals, withuse of more data for training the network, a better result would beobtained. In the early standard algorithm, a random initial set ofweights were assigned to the neural network, and then byconsidering the input data, weights were adjusted so the outputerror would be on its minimum.

5. Conclusions

The present investigation shows that the sunflower biomass isan effective and inexpensive biosorbent for the uptake of cobalt(II)from aqueous solutions. The sunflower biomass exhibited highbiosorption capacity. The Ct/Co is a function of biosorption time,the biosorbent dosage, biosorbate concentration, biosorbent par-ticle size, temperature, pH and bed depth. The developed meth-odology is economical, compatible and eco-friendly.

As the influent concentration increased, sharper breakthroughcurves were received. The breakthrough curves shifted to theorigin with the increasing flow rate, and an earlier breakthroughtime and saturation time were observed for a higher flow rate.In the column studies at the height of 5 cm, the biosorbent becomessaturated very quickly because of the fact that the cobalt(II) bindingon the sites is faster. Obviously, with an increase of pH in theinfluent, the breakthrough curves shifted from left to right, whichindicates that more metal ions can be removed. A rapid decrease inthe column biosorption capacity was observed with an increasefrom 0.25oxo0.5 to 1oxo2 mm in particle size.

Artificial neural network modelling has been used to investi-gate the cause-effect relationship in the bed studies of cobalt(II).The ANN model could describe the behaviour of the biosorptionwith the adopted experimental conditions. Simulation based onthe ANN model can then be performed in order to estimate thebehaviour of the system under different conditions.

The model based on ANN could predict the concentrationsof cobalt(II) uptake from aqueous solution during treatment time.

Fig. 5. Experimental and predicted breakthrough curves of cobalt(II) as a function of particle size (T 20 1C, bed depth 5 cm, Co 40 ppm, flow rate 8 ml/min, pH 5).

E. Oguz, M. Ersoy / Ecotoxicology and Environmental Safety 99 (2014) 54–60 59

A relationship between the predicted results of the designed ANNmodel and experimental data was also conducted. At the result ofANN model, the values of determination coefficient (R2), Standarddeviation ratio, mean absolute error and root mean square errorwere obtained as 0.960, 0.192, 1.447and 0.41, respectively. Accord-ing to sensitivity analysis results, the most important parameterseffecting the removal of cobalt(II) were determined as treatmenttime, bed depth, adsorbent dosage, particle size, initial cobalt(II)concentration, flow rate and pH, respectively.

Acknowledgments

The authors are grateful to the research council of AtatürkUniversity for providing financial support under project no: 2008/149.

Appendix A. Supplementary material

Supplementary data associated with this article can be found in theonline version at http://dx.doi.org/10.1016/j.ecoenv.2013.10.004.

References

Adebowale, K.O., Unuabonah, E.I., Olu-Owolabi, B.I., 2008. Kinetic and thermo-dynamic aspects of the adsorption of Pb2þ and Cd2þ ions on tripolyphosphate-modified kaolinite clay. Chem. Eng. J. 136, 99–107.

Al-Ghouti, M.A., Khraisheh, M.A.M., Ahmad, M.N., Allen, S.J., 2007. Microcolumnstudies of dye adsorption onto manganese oxides modified diatomite. J. Hazard.Mater. 146, 316–327.

Augustine, E.O., 2010. Intraparticle diffusion process for lead(II) biosorption ontomansonia wood sawdust. Bioresour. Technol 101, 5868–5876.

Baccar, R., Blánquez, P., Bouzid, J., Feki, M., Attiya, H., Sarrà, M., 2013. Modeling ofadsorption isotherms and kinetics of a tannery dye onto an activated carbonprepared from an agricultural by-product. Fuel Process. Technol. 106, 408–415.

Bayramoglu, G., Tuzun, I., Celik, G., Yilmaz, M., Arica, M.Y., 2006. Biosorption ofmercury (II), cadmium (II) and lead (II) ions from aqueous system by microalgae Chlamydomonas reinhardtii immobilized in alginate beads. Int. J. Miner.Process. 1, 35–43.

Bering, B.P., Dubinin, M.M., Serpinsky, V.V., 1972. On thermodynamics of adsorptionin micropores. J. Colloid Interf. Sci 38, 185–194.

Bernard, L., Bernard, P., Karine, G., Florence, B., Croue, J.P., 2004. Modeling ofbromate formation by ozonation of surface waters in drinking water treatment.Water Res. 38, 2185–2195.

Bhatti, H.N., Bajwa, I.I., Hanif, M.A., Bukhari, I.H., 2010. Removal of lead and cobaltusing lignocellulosic fiber derived from Citrus reticulata waste biomass. KoreanJ. Chem. Eng. 27 (1), 218–227.

Bhatti, H.N., Khadim, R., Hanif, M.A., 2011. Biosorption of Pb(II) and Co(II) on redrose waste biomass. Iran. J. Chem. Chem. Eng. 30 (4), 81–88.

Bhatti, H.N., Saleem, Anila., Hanif, M.A., 2013. Utilization of Mentha arvensis wastebiomass for the removal of Pb(II) and Co(II) from aqueous solutions. Desalin.Water Treat 51 (16–18), 3335–3343.

Brunauer, S., Emmett, P.H., Teller, E., 1938. Biosorption of gases in multimolecularlayers. J. Am. Chem. Soc. 60, 309–319.

Celik, S., Tan, O., 2005. Determination of preconsolidation pressure with artificialneural network. Civ. Eng. Environ. Sys 22 (4), 217–231.

Eloussaief, M., Jarraya, I., Benzina, M., 2009. Adsorption of copper ions on two claysfrom Tunisia: pH and temperature effects. App. Clay Sci 46, 409–413.

Futalan, C.M., Kan, C.C., Dalida, M.L., Pascua, C., Wan, M.W., 2011. Fixed-bed columnstudies on the removal of copper using chitosan immobilized on bentonite.Carbohyd. Polym. 83, 697–704.

Goel, J., Kadirvelu, K., Rajagopal, C.V., Garg, K., 2005. Removal of lead(II) bybiosorption using treated granular activated carbon: batch and column studies.J. Hazard. Mater. B 125, 211–220.

Gundogdu, A., Ozdes, D., Duran, C., Bulut, V.N., Soylak, M., Senturk, H.B., 2009.Biosorption of Pb(II) ions from aqueous solution by pine bark (Pinus brutiaTen.). Chem. Eng. J 153, 62–69.

Gupta, N., Kushwaha, A.K., Chattopadhyaya, M.C., 2012. Adsorptive removal of Pb2þ ,Co2þ and Ni2þ by hydroxyapatite/chitosan composite from aqueous solution.J. Taiwan Inst. Chem. E 43, 125.

Han, R., Zhang, J., Zou, W., Xiao, H., Shi, J., Liu, H., 2006. Biosorption of copper(II) andlead(II) from aqueous solution by chaff in a fixed-bed column. J. Hazard. Mater.B 133, 262–268.

Hansen, H.K., Arancibia, F., Gutiérrez, C., 2010. Adsorption of copper onto agricul-ture waste materials. J. Hazard. Mater. 180, 442–448.

Hasfalina, C.M., Maryam, R.Z., Luqman, C.A., Rashid, M., 2010. The potential use ofkenaf as a bioadsorbent for the removal of copper and nickel from single andbinary aqueous solution. J. Nat. Fibres 7 (4), 267–275.

Ho, Y.S., Porter, J.F., Mckay., G., 2002. Equilibrium isotherm studies for the sorptionof divalent metal ions onto peat: copper, nickel and lead single componentsystems. Water Air Soil Pollut 141, 1–33.

Karge, H.G., Weitkamp, J., 2006. Molecular Sieves Science and Technology, Char-acterization II. 5. Springer-Verlag, Berlin Heidelberg p. 23.

Kausara, A., Bhattia, H.N., MacKinnonb, Gillian, 2013. Equilibrium, kinetic andthermodynamic studies on the removal of U(VI) by low cost agricultural waste.Colloid Surf. B 111, 124–133.

Krishnan, K.A., Anirudhan, T.S., 2008. Kinetic and equilibrium modelling of cobalt(II) biosorption onto bagasse pith based sulphurised activated carbon. Chem.Eng. J. 137, 257–264.

Krishnani, K.K., Meng, X., Christodoulatos, C., Boddu, V.M., 2008. Biosorptionmechanism of nine different heavy metals onto biomatrix from rice husk. J.Hazard. Mater. 153, 1222–1234.

Manohar, D.M., Noeline, B.F., Anirudhan, T.S., 2006. Adsorption performance of Al-pillared bentonite clay for the removal of cobalt(II) from aqueous phase. Appl.Clay Sci. 31, 194–206.

Mondal, M.K., 2009. Removal of Pb(II) from aqueous solution by adsorption usingactivated tea waste. Korean J. Chem. Eng, 1–8.

Montgomery, D.C., 2008. Design and Analysis of Experiments, 7th ed. John Wileyand Sons, New York.

Nascimento, M., Sérgio, P., Soares, M., Paulo de Souza, V., 2009. Adsorption of heavymetal cations using coal fly ash modified by hydrothermal method. Fuel 88,1714–1719.

Nuhoglu, Y., Oguz, E., 2003. Removal of copper(II) from aqueous solutions bybiosorption on the cone biomass of Thuja orientalis. Process Biochem 38,1627–1631.

Oguz, E., Tortum, A., Keskinler, B., 2008a. Determination of the apparent rateconstants of the degradation of humic substances by ozonation and modelingof the removal of humic substances from the aqueous solutions with neuralnetwork. J. Hazard. Mater 157, 455–463.

Oguz, E., Keskinler, B., Tortum, A., 2008b. Determination of the apparent ozonationrate constants of 1:2 metal complex dyestuffs and modeling with a neuralnetwork. Chem. Eng. J. 141, 119–129.

Oguz, Ensar., 2005. Adsorption characteristics and the kinetics of the Cr(VI) on theThuja oriantalis. Colloid Surf. A 252, 121–128.

Oguz, Ensar, 2007. Equilibrium isotherms and kinetics studies for the sorption offluoride on light weight concrete materials. Colloid Surf. A 295, 258–263.

Ouvrard, S., Simonnot, M.O., DeDonato, P., Sardin, M., 2002. Diffusion-controlledadsorption of arsenate on natural manganese oxides. Ind. Eng. Chem. Res. 41,6194–6199.

Papandreou, A.D., Stournaras, C.J., Panias, D., Paspaliaris, I., 2011. Adsorption of Pb(II), Zn(II) and Cr(III) on coal fly ash porous pellets. Miner. Eng. 24, 1495–1501.

Pazourek, J., Gajdosov´a, D., Spanil´a, M., Farkov´a, M., Novotn´a, K., Havel, J., 2005.Analysis of polyphenols in wines: correlation between total polyphenoliccontent and antioxidant potential from photometric measurements predictionof cultivars and vintage from capillary zone electrophoresis fingerprints usingartificial neural network. J. Chromatogr. A 1081, 48–54.

Salamatinia, B., Kamaruddin, A.H., Abdullah, A.Z., 2008. Modeling of the continuouscopper and zinc removal by sorption onto sodiumhydroxide-modified oil palmfrond in a fixed-bed column. Chem. Eng. J 145 (2), 259–266.

Sihem, A., Lehocine, B.M., Miniai, H.A., 2012. Batch adsorptıon of phenol fromindustrıal waste water usıng cereal by-products as a new adsorbent. EnergyProcedia 18, 1135–1144.

Singh, S., Srivastava, V.C., Mall, I.D., 2009. Fixed-bed study for adsorptive removal offurfural by activated carbon. Colloid Surf A 332, 50–56.

Tortum, A., 2003. The Modeling of Mode Choices of Intercity Freight Transportationwith the Artificial Neural Networks and Integrated Neuro-fuzzy System. Ph.D.Thesis. Atatürk University, in Turkish.

Turan, N.G., Mesci, B., Ozgonenel, O., 2011. The use of artificial neural networks(ANN) for modeling of adsorption of Cu(II) from industrial leachate by pumice.Chem. Eng. J 173, 98–105.

Vazquez, G., Alonso, R., Freire, S., Gonzalez-Alvarez, J., Antorrena, G., 2006. Uptakeof phenol from aqueous solutions by biosorption in a Pinus pinaster barkpacked bed. J. Hazard. Mater. B 133, 61–67.

Volesky, B., 2000. Detoxificiation of metal-bearing effluents. Biosorption for thenext century. Hydrometallurgy 59, 203–216.

Walker, G.M., Weatherley, L.R., 1997. Biosorption of acid dyes onto granularactivated carbon in fixed bed. Water Res. 31, 2093–2101.

Xiaobing, L.I., Chunjuan, Z., Jiongtian, L.I.U., 2010. Adsorption of oil from wastewater by coal:characteristics and mechanism. Min. Sci. Technol 20, 778–781.

Zupan, J., Gasteiger, J., 1999. Neural Networks in Chemistry and Drug Design. Wiley-VCH,Weinheim.

E. Oguz, M. Ersoy / Ecotoxicology and Environmental Safety 99 (2014) 54–6060