Accepted Manuscript Title: Removal of Pb(II) ions from aqueous solution using water hyacinth root by fixed-bed column and ANN modeling Author: Tania Mitra Biswajit Singha Nirjhar Bar Sudip Kumar Das PII: DOI: Reference:

S0304-3894(14)00211-8 http://dx.doi.org/doi:10.1016/j.jhazmat.2014.03.025 HAZMAT 15798

To appear in:

Journal of Hazardous Materials

Received date: Revised date: Accepted date:

1-11-2013 4-3-2014 19-3-2014

Please cite this article as: T. Mitra, B. Singha, N. Bar, S.K. Das, Removal of Pb(II) ions from aqueous solution using water hyacinth root by fixedbed column and ANN modeling, Journal of Hazardous Materials (2014), http://dx.doi.org/10.1016/j.jhazmat.2014.03.025 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Removal of Pb(II) ions from aqueous solution using water hyacinth root by fixed-bed column and ANN modeling

ip t

Tania Mitra, Biswajit Singha, Nirjhar Bar and Sudip Kumar Das Chemical Engineering Department University of Calcutta 92, A. P. C. Road, Kolkata – 700 009, India e-mail : [email protected] ABSTRACT

cr

Hyacinth root was used as a biosorbent for generating adsorption data in fixed-bed glass column. The influence of different operating parameters like inlet Pb(II) ion

us

concentration, liquid flow rate and bed height on the breakthrough curves and the

an

performance of the column was studied. The result showed that the adsorption efficiency increased with increase in bed height and decreased with increase in inlet Pb(II) ion

M

concentration and flow rate. Increasing the flow rate resulted in shorter time for bed saturation. The result showed that as the bed height increased the availability of more number

d

of adsorption sites in the bed increased, hence the throughput volume of the aqueous solution

te

also increased. The adsorption kinetics was analyzed using different models. It was observed that maximum adsorption capacity increased with increase in flow rate and initial Pb(II) ion

Ac ce p

concentration but decreased with increase in bed height. Applicability of Artificial Neural Network (ANN) modeling for the prediction of

Pb(II) ion removal was also reported by using multilayer perceptron with Backpropagation, Levenberg-Marquardt and Scaled Conjugate algorithms and four different transfer functions in a hidden layer and a linear output transfer function. Keywords: Adsorption; Hyacinth root; Thomas model; Backpropagation; LevenbergMarquart; Scaled conjugate gradient 1.

Introduction Pollution due to the presence of heavy metals in wastewater is a very serious

environmental problem because of their toxic effects, non-biodegradability, accumulation in

Page 1 of 42

living tissues and effect on the food chain as they can easily enter human and animal bodies. Lead from the atmosphere or soil can contaminate the groundwater and surface water. Lead poisoning in human beings can cause damage to the nervous system, reproductive system,

ip t

liver, kidney and brain. These heavy metals are released from various industries like electroplating, textile, metal finishing, storage batteries, mining, galvanizing, glass and

cr

ceramics, paints, pigments alloy manufacturing and agricultural activities etc. into the water bodies [1-4]. According to the US Environmental Protection Agency the permissible limit of

us

Pb(II) in drinking water is 0.015 mg/L where as its permissible limit according to WHO and

an

IS is 0.01 mg/L and 0.05 mg/L respectively. IS limit for discharge of Pb(II) ion in inland surface water is 0.1 mg/L. Hence, considerable attention has been paid to remove these heavy

M

metals from industrial wastewater. The technique of adsorption process is very promising as it is cost effective, very simple in design and easy to operate. Adsorptive removal of Pb(II)

d

using low cost adsorbents has been recently reviewed by Singha and Das (2012) [4]. The

te

biosorption is a cheap and environmentally safe alternative but its application requires a careful evaluation as they appear to have large variability in the rate and degree of adsorption

Ac ce p

which may limit its usefulness [5].

This work is aimed to investigate the adsorption behavior of hyacinth root for the

treatment of aqueous solution of Pb(II) ion in continuous column operation and the predictability of the removal of Pb(II) ion in the adsorption system using Artificial Neural Network (ANN). 2.

Materials and methods

Hyacinth root is collected from pond alocated near Kolkata, West Bengal, India. It is washed with distilled water and sundried for 6 – 7 days. It is then dried in an oven at 70oC for 5 hours and then crushed and sieved through mesh to get particle sizes of 290-350 µm. 2.1 Preparation of metal solution

Page 2 of 42

The chemicals used for the study are analytical grades Lead Nitrate [Pb(NO3)2], Nitric Acid [HNO3] and Sodium hydroxide [NaOH] (purchased from E. Merck India Limited, Mumbai). Stock Solution of 1000mg/L Pb(II) is prepared by dissolving 1.61g of Pb(NO3)2 in

ip t

distilled water containing a few drops of nitric acid to prevent the precipitation of Pb(NO3)2 by hydrolysis.

cr

2.2 Instruments used

An atomic absorption spectrophotometer (AA 240 VARIAN, Australia) is used for

us

the measurement of Pb(II) ion concentration. Scanning electron microscope (S-3400N;

an

Hitachi, Japan) is used to observe the surface texture of the sorbents. Fourier transform infrared spectroscopy (Alpha FTIR, Bruker, Germany) is used to determine the type

M

offunctional group present in the hyacinth root. The surface area is measured by Micromeritics surface area analyzer (ASAP 2020). The moisture content determinationis

d

carried out with a digital microprocessor-based moisture analyzer (Metteler LP16). The point

te

of zero charge is determined by solid addition method (Srivastava et al. 2006). The pH of the solution is measured usingmulti 340i (WTW) digital microprocessor-based pH meter

Ac ce p

previously calibratedwith standard buffer solutions. A peristaltic pump (Cole-Parmer, model7535-04, USA) is used for providing constant flow of aqueous solution and desorbing solution in the fixed-bed. Different physical characteristics of hyacinth root are shown in Table 1.

3. Experimental

The experiment was carried out in glass columns with internal diameter of 1.8cm and

length of 50cm. At the bottom of the column, 0.05 cm thick glass wool was placed to prevent any loss of the biosorbent and was finally guarded by a stainless steel sieve plate to give mechanical support. The whole experiment was carried out at room temperature, i.e, 28±2o C. The effect of process parameters like bed height, initial Pb(II) ion concentration and flow rate

Page 3 of 42

through the bed, on the adsorption process was investigated. Process flow diagram is shown in Fig. 1. A known quantity of hyacinth root was placed and then distilled water was added to

ip t

settle the bed at constant height. After the desired height was obtained, the water was drained out. Metal solution having known initial concentration was pumped through the column at

cr

desired flow rate by peristaltic pump. Samples were collected at specific time intervals from the bottom of the column to analyze the Pb(II) ion concentration using AAS. The

us

experimentwas completed when the Pb(II) ion concentration in the bed became nearly equal

an

to the initial Pb(II) ion concentration. 3.1 Characteristics of the adsorbent

M

The detailed characteristics of the hyacinth root have been reported in earlier studies [4]. Table 1 shows the characteristics of the hyacinth root.

d

4. Results and Discussion

te

The column performance was evaluated by plotting relative concentration of Pb(II) ion, which is defined as the ratio of the concentration of Pb(II) ion in effluent to the

Ac ce p

concentration in Pb(II) ion in influent (Ct/C0) with respect to flow time, t. The total Pb(II) ion adsorbed, qt, in the fixed-bed column for a given solute concentration and flow rate is calculated from the following equation, qt 

v t t s Cad dt 1000 t  0

(1)

The equilibrium uptake of Pb(II) ions, qeq, based on the weight of the biomass can be calculated from the following equation, qeq 

qt  qto x

(2)

The percentage of removal (R) can be calculated from the following equation,

Page 4 of 42

R

C 0  Ct  100 C0

(3)

4.1 Effect of initial pH [4] Bio-sorption of Pb(II) on hyacinth root was affected by the solution pH. From our

ip t

previous batch study it was observed that the removal capacity of the adsorbent increases

cr

significantly by increasing the solution pH at 2.0–5.0. At low pH, the removal capacity of the adsorbent is low because of the competition of the H+ and Pb(II) ion for the occupancy of the

us

binding sites of the adsorbent and surface functional group, which are responsible for adsorption and which may be activated as the pH increases. As the pH starts to increase from

an

pH 5.0, the Pb(II) ion starts to precipitate as lead hydroxide and the removal is based on the combination of the adsorption and precipitation. Hence the present study was carried out at

M

pH 5.0. 4.2 Effect of Flow Rate

d

The flow rate plays an important role on the performance of adsorption process.

te

Experiments were carried out by varying the flow rate between 10 to 40 ml/min. The effect of

Ac ce p

flow rate on breakthrough curve is shown in Fig. 2. It is clear from the figure that the breakthrough time decreases as the flow rate increases. At the lowest flow rate, the metal ions have more time to be in contact with adsorbent than at higher flow rate, which results in the higher removal of metal ions from solution. Steep breakthrough curve is observed at higher flow rates, which causes the breakthrough time to be fast. This behavior probability is due to insufficient residence time of the solute in the column and the diffusion limitations of the solute into the pores of the sorbent at higher flow rate [6]. However, the effect of flow rate on the metal uptake capacity has been observed to be the best at the lowest flow rate. 4.3 Effect of Influent Concentration Fig. 3 shows the effect of influent Pb(II) ion concentration for the removal on the hyacinth root by plotting the outlet Pb(II) ion concentration against the contact time. It is

Page 5 of 42

clear from the figure that the adsorption capacity of the adsorbent increases with the decrease in influent metal ion concentration and it has taken more time to reach the saturation time. So, as the influent concentration increases, the saturation time decreases becausethe adsorption

ip t

sites are covered rapidly by the metal ion. The sharp breakthrough curve and smaller breakthrough time is the result of larger influent concentration.

cr

4.4 Effect of bed height

The bed height (Z) is varied from 4–6 cm for the experiment. Fig. 4 shows the effect

us

of bed height for Pb(II) ion adsorption on hyacinth root at constant flow rate and influent

an

concentration. As the bed height increases, the curve tends to shift to the right side of the time axis, i.e., the breakthrough time increases due to the fact that more adsorption sites are

M

available. The equilibrium adsorption capacity decreases with the increase in the bed height. This indicates that at the smaller bed height the adsorbate concentration ratio (Ct/C0)

d

increases more rapidly than compared to the higher bed height due to the availability of less

te

adsorbent. These trends are very common and reported by many researchers [7-9]. For smaller bed height the amount of adsorbent requirement is less and the bed is also saturated

Ac ce p

quickly.

4.5 Kinetics and Isotherm

The interaction between adsorbate and adsorbent are normally interpreted using the

standard isotherm model, i.e., Langmuir and Freundlich model. The Langmuir model assumes all sites of the adsorbent be homogeneous, identical, energetically equivalent and monolayer adsorption, whereas, the Freundlich isotherm assumes the multilayer adsorption on the adsorbent surface. The kinetics and isotherm for batch adsorption process using the hyacinth root are reported in our earlier studied [4] and the best kinetic and isotherm models are the pseudo-2nd order model, and Frendlich isotherm model (Table 2). 4.6 Modeling of the column study

Page 6 of 42

The fixed-bed column design is needed for the prediction of breakthrough curves and the adsorption capacities under different operating conditions. The well-known simple empirical and semi-empirical models have been used to analyze the experimental data.

ip t

4.6.1 Bohart-Admas model [10] The Bohart-Admas model is as follows,

cr

C  Z ln  t   K ABC0t  K AB N 0 v  C0 

(4)

us

Fig. 5 represents the Bohart-Admas model curve for different flow rate and influent concentration of Pb(II) ion. Table 3 shows the parameters, KAB and N0 along with the

an

correlation coefficients. It is clear from the table that the kinetic constant, KAB, and maximum adsorption capacity, N0, are dependent on the flow rate and influent concentration. This

M

model assumes the adsorption rate to be proportional to both the residual capacity of adsorbent and the concentration of the metal ions. It is often used to describe the initial part

te

d

of the breakthrough curve [11-12]. The value of KAB increases by increasing the value of both influent concentration and flow rate, but it decreases with the increase in the bed height.

Ac ce p

Hence, it indicates that the overall system kinetics is controlled by the external mass transfer in the initial part of the adsorption process within the column [11]. 4.6.2 Wolborska model [13-14]

This model is based on the concept of mass transfer in diffusion mechanism, which is

applicable only for low metal ion concentration, negligible axial diffusion and only the external diffusion with a constant kinetic coefficient. The expression of the model is, ln

Ct  a .C0  H  t a C0 q v

(5)

Fig. 5 is used to calculate the Bohart-Admas parameters and also the Wolborska parameters. Table 3 shows the model parameters along with the correlation coefficients. 4.6.3 Thomas model [15]

Page 7 of 42

The expression of the Thomas model for adsorption is given as,

Ct  C0

 kth  q0 x  C0Veff 1  exp  v 

(6)

  

C  k q x k CV ln  0  1  th 0  th 0 eff v v  Ct 

ip t

or,

1

(7)

cr

The kinetic coefficient, kth and the adsorption capacity of the column, q0 are determined from

us

the plot of ln [(C0/Ct) - 1] against Veff at a given flow rate. Fig. 6 represents the Thomas model and Table 4 shows the model parameters along with the correlation coefficients. This model

an

agrees well with the experimental data in the range of (Ct/C0) < 0.3. This model assumes that the external and internal diffusion is not the limiting step and the Langmuir isotherm is valid.

M

But the adsorption is generally controlled by the interphase mass transfer and also the existence of an axial dispersion. From Table 4 it is clear that as the influent concentration

d

increases, the value of q0 also increases, but the value of kinetic coefficient, kth, decreases.

te

This is due to the high driving force for adsorption between the Pb(II) ions on the hyacinth

Ac ce p

root and the Pb(II) ion concentration in the solution, and the result shows better column performance. The maximum adsorption capacity, q0 increases with increase in flow rate and initial Pb(II) ion concentration but decreases with bed height. The rate constant, kth increases with increase in flow rate but decreases with increase in initial Pb(II) ion concentration and bed height. The correlation coefficient values range from 0.9475 to 0.9964.These resultsindicate a good agreement between the experimental value and the data generated by using Thomas model. 4.6.4 Yoon-Nelson model [16] This model assumedthat the rate of decrease in the probability of adsorption for each adsorbate molecule is proportional to the probability of adsorbate and also the probability of adsorbate breakthrough on the adsorbent [9]. The expression of the model is,

Page 8 of 42

Ct  exp  kYN t  kYN  (C0  Ct )

(8)

From the linearized plot of ln[Ct/(C0-Ct)] vs t at different flow rates, initial Pb(II) ion concentrations and bed heights kYN and τ were calculated. Fig. 7 shows the model plot and

ip t

Table 5 shows the model parameters. The rate constant kYN increases and the 50%

cr

breakthrough time, i.e., τ, decreases with both increasing the flow rate and Pb(II) ion influent concentration. With the increase of bed height the value of τ increases while the value ofkYN

us

decreases. The increase in kYN with flow rate indicates a decrease in the mass transfer resistance and the decrease in τ with flow rate indicates the faster saturation of the adsorption

an

bed. The mass transfer resistance is proportional to the axial dispersion and thickness of the liquid film on the particle surface. In this case one can assume that the axial dispersion is

M

negligible as the flow rates are small enough; hence, with increasing the flow rate the driving

4.7 FTIR studies [4, 19-20]

te

researchers [17-18].

d

force of the mass transfer in the liquid film increases. Similar dependency is also observed by

Ac ce p

FTIR studies are used to investigate the changes in the vibration frequency in the functional groups of the hyacinth roots due to Pb(II) ion adsorption. It has been scanned in the spectral range of 4000 – 400 cm-1. The functional groups are one of the key factor to understand the mechanism of the Pb(II) adsorption process on the hyacinth roots. Fig. 8 shows the FTIR plot for fresh and Pb(II) ions loaded hyacinth roots respectively. Table 6 represents the shift from the wave number for the dominant peak associated with the fresh and Pb(II) ions loaded hyacinth roots in the FTIR plot. Surface O-H stretching, Aliphatic C-H stretching, Aliphatic acid C=O Stretching, Unsaturated group like alkene, Aromatic C-NO2 stretching, Sulphonic acid S=O stretching are responsible for the Pb(II) ion adsorption on the hyacinth roots. 4.8Artificial Neural Network (ANN) modeling

Page 9 of 42

Fig. 9 shows the schematic diagram of multilayer perceptron (MLP). Bar and Das (2012) [21] had pointed out that out of all types of ANN the MLP the most popular and widely used network. It has three layers: an input layer, hidden layer(s) and an output layer.

ip t

Literature survey suggested that a network with single hidden layer using different popular transfer functions like sigmoid, hyperbolic tangent etc. are extensively used for prediction

cr

and they can besuccessfully used to perform various problems related to prediction. Hence this study is based on MLP using a single hidden layer. Presently there are many popular

us

training algorithms in use like Backpropagation (BP), Delta-Bar-Delta (DBD), Quickprop

an

(QP), Conjugate gradient (CG), scaled conjugate gradient (SCG), Levenberg-Marquart (LM) etc. Out of the above mentioned algorithms Backpropagation (BP), Quickprop (QP) and

M

Delta-Bar-Delta (DBD) are first-order algorithms. Scaled conjugate gradient (SCG) and Levenberg-Marquart (LM) algorithms are second-order algorithms. The advantage of Scaled

d

conjugate gradient (SCG) and Levenberg-Marquart (LM) is that they are user independent

te

whereas Backpropagation (BP), Quickprop (QP) and Delta-Bar-Delta (DBD) are user dependent. Considering the popularity and usefulness of the algorithms the present study is

Ac ce p

done using Backpropagation (BP), Scaled conjugate gradient (SCG) and Levenberg-Marquart (LM) algorithm.

For the networks using Backpropagation algorithm, each run was set for a maximum

of 32000 epochs. If no improvement was observed in the value of cross-validation MSE for 20000 epochs then the training was set to stop. For the hidden layer of BP network the value of learning rate was 0.7 and that of momentum coefficient was 1. For network using Levenberg-Marquardt algorithm each run was set for 1000 epochs. If no improvement was observed in the value of cross-validation MSE for 300 epochs then the training was set to stop.

Page 10 of 42

For network using Scaled Conjugate Gradient algorithm each run was set for 32000 epochs. If no improvement was observed in the value of cross-validation MSE for 10000 epochs then the training was set to stop.

ip t

Also the minimum value of MSE of cross-validation (the threshold value) was set at 0.001for all the training of all the 3 networks.

cr

4.8.1 ANN Performance

Sola and Sevilla [22] reported the effects of data normalization on the ANN process

us

and concluded normalization of data yields better prediction. Hence, to avoid numerical

an

overflow, i.e. the effect due to very large and very small values all raw data used as input variables or output are normalized. The normalization (limiting the values in between 10 and

M

20) was done using the following formula,

  input value - minimum input value Normalized value   10   10  maximum input value - minimum input value 

(9)

te

d

Table 7 represents the range of variables investigated. The input variables are initial concentration, bed height, flow rate, and time and the output variable is the outlet

Ac ce p

concentration of Pb(II) ions in the exit of the column. The total number of data points is 534. The performance of the network is checked using the following parameters, Mean square error (MSE),

Average Absolute Relative Error (AARE),

AARE 

1 N ( yi  xi )  x N i 1 i

(10)

Standard Deviation (σ),  1  ( yi  xi )    AARE   xi i 1 N  1   N

2

(11)

Cross-correlation coefficient (R),

Page 11 of 42

N

R

 ( x  x)( y i

i 1

N

i

 y)

(12)

N

 ( x  x)  ( y  y ) 2

i 1

i

i 1

2

i

N

2  

( xi  yi )2 yi

(13)

cr

i 1

ip t

Chi-square test (χ²),

To explain the good performance of the network the MSE, AARE and Standard Deviation

us

should be as small as possible. It has also been verified that the Cross-correlation coefficient between input and output should be close to unity for better predictability. When more than

an

one model is acceptable statistically then the Chi-square (χ²) test is required to be performed to find the best-fit model. The lowest value indicates the best model.

M

Fig. 10 shows the variation of minimum value of cross-validation MSE with the number of nodes for different transfer functions in the hidden layer for Backpropagation

d

algorithm. Similar procedure has been followed for the other 2 networks also. The optimum

te

number of processing element is estimated on the basis of the lowest value of theminimum

Ac ce p

value MSE for cross-validation. Table 8 presents the optimum number of processing elements for all 3 networks. Fig. 11 represents the cross-validation curve of the neural networks where training was done using Levenberg-Marquardt algorithm. Table 9 represents the minimum value of Cross–validation MSE reached during

training with four different transfer functions. Similar procedure has also been reported in earlier studies [21, 23].

Testing was done just before the final prediction to check the effectiveness of training using the above four mentioned networks. Table 10 represents the performance of the testing. The low values of AARE and cross correlation co-efficient (R) which is nearly 0.99 in each case indicates that the training was good and the network can be used for final prediction.

Page 12 of 42

Fig. 12 shows the comparison between the experimental and the predicted outlet concentration with optimum number of processing elements in the hidden layer for testing for training with Scaled Conjugate Gradient algorithm.

ip t

Table 11 represents the performance of the neural networks for final prediction. It is clear from the table that the cross correlation co-efficient (R) value is more than 0.998 for

cr

each of the four cases. The low value of the average absolute relative error (AARE), Standard Deviation (σ) and Mean square error (MSE) also shows the accuracy of the result in the

us

different systems.

an

The high value of cross correlation co-efficient (R) and the low value of chi square (χ²) for the transfer function 2 with 20 optimum number of processing elements gives the

M

most accurate prediction of the percentage removal using Backpropagation algorithm. Fig. 12 shows the comparison between the experimental and the predicted outlet concentration.

te

5. Conclusions

d

This result indicates that the performance of the network output is excellent.

In this study, the sorption of Pb(II) ion onto hyacinth root in fixed-bed column

Ac ce p

operated in continuous mode are reported. Applicability of ANN modeling was also investigated.

Effect of different adsorption parameters, flow rate, bed height and initial Pb(II) ion

concentration were studied in fixed-bed column using hyacinth root. The Pb(II) ion adsorption increased with increase in bio-sorbent dose and decreased with the increase of initial Pb(II) ion concentration or bed height. Thomas model, Bohart-Admas model, Yoonnelson model and Wolborska model are applied to the experimental data obtained from this study to predict the breakthrough curve and also to determine the characteristic parameters of the column. From the statistical analysis, the Thomas model gives the better prediction.

Page 13 of 42

A neural network based model has been developed for the prediction of percentage removal. A multilayer perceptron with Backpropagation, Levenberg-Marquardt and Scaled Conjugate algorithms and four different standard transfer functions in a hidden layer and a

ip t

linear output function have been used for the analysis. Optimization for each transfer function has been carried out in all cases. The ANN model with Backpropagation algorithm, using a

cr

hidden layer with transfer function 2 and 20 processing elements gives better predictability of the outlet concentration.

us

Acknowledgment The authors acknowledge DST, Govt. of India for the financial support for

an

the project work (file no. DST/WTI/2K9/141, dated 19.05.2010). Nomenclature Adsorbed concentration (mg L-1)

C0

Influent metal ion concentration at t= 0 (mg L-1)

Ct

Effluent metal ion concentration at time t (mg L-1)

te

Ct/C0 Relative concentration

d

M

Cad

Kinetic constant for Bohart-Admas model (L mg-1)

kth

Thomas rate constant (L mg-1 min-1)

Ac ce p

KAB

KYN

Yoon-Nelson model constant (min-1)

No

Saturation concentration (mg L-1)

qo qeq qt

Maximum metal uptake per gram of the adsorbent (mg g-1) Equilibrium metal uptake by adsorbent (mg g-1) Total adsorbed Pb(II) at time t (mg)

qto

Total adsorbed Pb(II) at time t=0 (mg)

R

Percentage removal of metal ion (%)

R2

Correlation factor

t

Flow time (min)

Page 14 of 42

Breakthrough time (minute)

v

Flow rate (ml min-1)

Veff

Effluent volumn (ml)

x

Amount of adsorbent present in the column (g)

Z

Bed depth (cm)

a

Wolborska rate constant (min-1)

τ

Time for Yoon-Nelson model (min)

us

cr

ip t

tb

References

an

[1] T. K. Naiya, A. K. Bhattacharya, S. K. Das,. Adsorption of Cd(II) and Pb(II) from aqueous solutions on activated alumina, J. Colloid Interface Sci. 333 (2009a)14–26.

M

[2] T. K. Naiya, A. K. Bhattacharya, S. N. Mandal, S. K. Das, The sorption of lead(II) ions on rice husk, J. Hazard. Mater. 163 (2009b) 1254–1264.

d

[3] X. S. Wang, Z. P. Lu., H. H. Miao, W. He, H. L. Shen, Kinetics of Pb(II) adsorption on

te

black carbon derived from wheat residue, Chem. Eng. J. 166 (2011) 986–993. [4] B. Singha, S. K. Das, Removal of Pb(II) ions from aqueous solution and industrial

Ac ce p

effluent using natural biosorbents, Environ. Sci. Pollut. Res., 19 (2012) 2212–2226. [5] A. Demirbas, Heavy metal adsorption onto agro-based waste materials : a review, J. Hazard. Mater. 157 (2008) 220–229. [6] D. C.K. Ko, J. F. Porter, G. McKay, Optimised correlations for the fixed-bed adsorption of metal ions on bone char, Chem. Eng. Sci., 55 (2000) 5819-5829. [7] K. S. Rao, S. Anand, P. Venkateswarlu, Modeling the kinetics of Cd(II) adsorption on Syzygiumcumini L leaf powder in a fixed bed mini column, J. Ind. Engg. Chem. 17 (2011) 174-181. [8] J. T. Nwabanne, P. K. Igbokwe, Adsorption performance of packedbedcolumn for the removal of Lead(II) usingoil palm fibre, Int. J. App. Sci. Technol. 2(5) (2012) 106-115.

Page 15 of 42

[9] R. Han, J. Zhang, W. Z. H. Xiao, J. Shi, H. Liu, Biosorption of Copper(II) and Lead(II) from aqueous solution by chaff in a fixed-bed column, J. Hazard. Mat. 133 (2006) 262–268. [10] G. Bohart, E. N. Admas, Some aspects of the behavior of charcoal with respect to

ip t

chlorine, J. Am. Chem. Soc. 42 (1920) 523–544. [11] A. C. Taxier, Y. Andrès, C. F. Brasquet, P. I. Cloirec, Fixed-bed study for anthanide (La

cr

Eu, Yb) in removal from aqueous solution by immobilized Pseudomonas aeruginosa: experiental data and modelization, Chemosphere 47 (2002) 333–342.

us

[12] Z. Aksu, F. Gönen, Biosorption of phenol by immobilized activated sludge in a

an

continuous packed bed: prediction of breakthrough curves, Process Biochem. 39 (2004) 599– 613.

M

[13] A. Wolborska, Adsorption on activated carbon of p-nitrophenol from aqueous solution, Water Res. 23(1989) 85–91.

d

[14] A. Wolborska, P. Pustelnik, A simplified method for determination of breakthrough time

te

of an adsorbent layer, Water Res. 30 (1996) 2643–2650. [15] H. G. Thomas, Chromatography: a problem in kinetics, Ann. N. Y. Acad. Sci. 49 (1948)

Ac ce p

161–182.

[16] Y. H. Yoon, J. H. Nelson, Application of gas adsorption kinetics II. A theoretical model for respirator cartridge service life, Am. Ind. Hyg. Assoc. J. 45 (1984) 509–516. [17]

E.

Malkoc,

Y.

Nuhoglu,

Y.

Abali,

Cr(VI)

adsorption

by waste

acorn

of Quercusithaburensis in fixed beds: Prediction of breakthrough curves, Chem. Engg. J., 119 (2006) 61-68

[18] O. Hamdaoui, Removal of Copper(II) from aqueous phase by purolite C100-MB cation exchange resin in fixed bed columns : Modeling, J. Hazard. Mat. 161 (2009) 737-746. [19] M. Ibrahim, Z. Al-Fifi, Mechanism of Pollution Control for Aquatic Plant Water Hyacinth", Open Spectroscopy J. 4 (2010) 10-15.

Page 16 of 42

[20] H. S. Ibrahim, N. S. Ammar, M. Soylak, M. Ibrahim, Removal of Cd(II) and Pb(II) from aqueous solution using dried water hyacinth as a biosorbent, Spectrochimica Acta Part A. 96 (2012) 413-420

ip t

[21] N. Bar, S. K. Das, Gas-non-Newtonian liquid flow through horizontal pipe – gas holdup and pressure drop prediction using multilayer perceptron, Am. J. Fluid Dynamics 2(3) (2012)

cr

7–16.

[22] J. Sola, J. Sevilla, Importance of input data normalization for the application of neural

us

networks to complex industrial problems, IEEE Trans. Nuclear Sci. 44(3) (1997) 1464–1468.

an

[23] N. Bar, S. K. Das, Comparative study of friction factor by prediction of frictional pressure drop per unit length using empirical correlation and ANN for gas-non-Newtonian

Ac ce p

te

d

M

liquid flow through 180° circular bend, Int. Rev. Chem. Engg. 3(6) (2011) 628–643.

Page 17 of 42

List of Figures Schematic diagram of Experimental setup 1. Feed storage, 2. Pump, 3.Glass column, 4. Fixed-bed

Figure 2

Variation of the effect of bed height for Pb(II) adsorption on hyacinth root at constant flow rate and influent concentration

Figure 3

Variation of the effect of liquid flow rate for Pb(II) removal on hyacinth root at constant bed height and influent concentration

Figure 4

Variation of the outlet Pb(II) ion concentration with the contact time

Figure 5

Bohart-Admas model plot

Figure 6

Thomas model plot

Figure 7

Yoon-Nelson model plot

Figure 8

FTIR plot for raw and Pb(II) loaded hyacinth root (a) Fresh hyacinth root, (b) Pb(II) loaded hyacinth root

Figure 9

Schematic diagram of MLP ANN

Figure 10

Variation of cross-validation MSE with the number of processing elements in the hidden layer

Figure 11

Cross-validation curve when the network is trained using LevenbergMarquardt algorithm

Ac ce p

te

d

M

an

us

cr

ip t

Figure 1

Figure 12

Comparison between the experimental to the predicted Outlet concentration with optimum numberof processing elements in the hidden layer for testing

Figure 13

Comparison between the experimental to the predicted Outlet concentration with optimum number elements in the hidden layer

Page 18 of 42

ip t

te

d

M

an

us

cr

Table 1- Characteristics of the adsorbent Name of adsorbent Water hyacinth roots Particle size (µm) 250 - 350 Surface area (m2/g) 5.78 Bulk density (g/cm3) 0.48 Dry matter (%) 88.75 Moisture content (%) 11.25 Ash content (%) 10.74 Point of zero charge 6.59

Ac ce p

Table 2 Kinetic and Isotherm model constants (Singha and Das, 2012) dq 2 Pseudo-Second Order model : t  k2  qe  qt  dt

k2 (mg/g/min) Correlation coefficient, r2 Langmuir model:

0.1658 0.9999 Ce C 1   e qe qmax .b qmax

qmax (mg g-1) b (l mg-1) Correlation coefficient, r2

24.9376 0.0701 0.9259 1

Freundlich model : qe  K f .  ce  n  Kf (mg g-1) (mg l-1)(1/n) n Correlation coefficient, r2

2.0363 1.7185 0.9977

Page 19 of 42

cr

ip t 10 20 30 Bed depth

Z cm

4 5 6

us

an

KAB × 10 a L mg-1 min-1 Bed depth: 5 cm; Influent concentration: 10 mg L-1 0.90 34.17 30.82 34.17 2.37 14.56 34.52 14.58 4.79 5.20 24.93 5.20 Bohart-Admas model Wolborska model Rate Saturation Rate Metal uptake constant concentration constant capacity 3

N0 g L-1

q g L min-1 -1

KAB × 10 a L mg-1 min-1 Bed depth: 6 cm; Flow rate: 30 ml min-1 2.10 14.56 30.57 14.56 1.41 21.97 30.98 24.97 0.57 34.90 19.89 35.10 Bohart-Admas model Wolborska model Saturation Rate Metal uptake Rate concentration constant capacity constant

Ac ce p

C0 mg l-1

q g L min-1 -1

M

10 30 40 Influent concentration

N0 g L-1

3

d

v ml min-1

Table 3 Bohart–Admas and Wolborska model parameters Bohart-Admas model Wolborska model Rate Saturation Rate Metal uptake constant concentration constant capacity

te

Flow rate

N0 g L-1

q g L-1 min-1

KAB × 103 a L mg-1 min-1 Influent concentration: 30 mg L-1; Flow rate: 40 ml min-1 2.94 17.86 52.52 17.86 2.60 14.25 37.04 14.25 0.53 22.44 12.63 27.67

Correlation coefficient R2

0.9063 0.8966 0.6714 Correlation coefficient R2

0.9600 0.7555 0.7946 Correlation coefficient R2

0.7484 0.7521 0.7530

Page 20 of 42

10 20 30 Bed depth

4 5 6

ip t R2

cr

us

kth x 103 q0 qeq -1 -1 -1 L mg min mg g mg g-1 Bed depth: 5 cm; Flow rate: 10 ml min-1 1.70 3.49 4.94 1.44 6.41 10.61 1.41 8.80 14.45 Maximum Equilibrium metal Rate constant metal uptake uptake kth x 103 q0 qeq L mg-1 min-1 mg g-1 mg g-1 Influent concentration: 30 mg L-1, Flow rate: 10 ml min-1 1.94 9.71 15.23 1.41 8.80 14.45 1.19 8.65 12.77

Ac ce p

Z cm

an

C0 mg L-1

kth x 103 q0 qeq -1 -1 -1 L mg min mg g mg g-1 Bed depth: 6 cm; Influent concentration: 20 mgL-1 1.39 6.20 8.26 3.61 5.93 13.18 6.17 8.63 9.57 Rate constant Maximum Equilibrium metal metal uptake uptake

M

10 30 40 Influent concentration

Correlation Coefficient

0.9830 0.8945 0.9022

R2 0.9950 0.9685 0.9700

d

v ml min-1

te

Flow rate

Table 4 Thomas model parameters Rate constant Maximum Equilibrium metal metal uptake uptake

R2 0.9863 0.9700 0.9798

Table 5 Yoon-Nelson model parameters Flow rate

Model constant

Time

Metal uptake capacity

Correlation Coefficient

v

KYN x 102

τ

q

R2

Page 21 of 42

10 20 30 Bed depth

R2

an

0.9614 0.9506 0.9747

R2 0.9062 0.9422 0.9404

Ac ce p

te

d

4 5 6

KYN x 102 τ q -1 min min mg g-1 Influent concentration: 20 mg L-1; Flow rate: 30 ml min-1 3.96 88.96 10.94 6.85 69.75 6.86 7.86 86.39 5.60

M

Z cm

KYN x 102 τ q min-1 min mg g-1 Bed depth: 4 cm; Flow rate: 10 ml min-1 1.11 222.77 4.56 8.83 162.61 6.66 5.10 153.62 9.44 Model constant Metal uptake Time capacity

ip t

C0 mg L-1

0.9922 0.9736 0.7748

cr

10 20 30 Influent concentration

min-1 min mg g-1 Bed depth: 6 cm; Influent concentration: 10 mg L-1 1.63 216.84 3.55 4.41 93.51 4.60 11.75 29.05 1.90 Model constant Time Metal uptake capacity

us

ml min-1

Table 6 Wave number (cm-1) for the dominant peak from FTIR for Pb(II) ions sorption

Functional groups

Raw hyacinth roots

Pb(II) ions loaded hyacinth roots

Surface O-H stretching Aliphatic C-H stretching Aliphatic acid C=O Stretching Unsaturated group like alkene Aromatic C-NO2 stretching Sulphonic acid S=O stretching

3328.53 2924.52 1713.44 1644.02 1514.81 1055.84

3273.90 2922.26 1713.44 1645.64 1541.10 997.12

Page 22 of 42

ip t cr us

Table 7 Range of data

Range 10–30 4–6 10–40 10–480 0.01–29.91 534

Ac ce p

te

d

M

an

Measurement Type Initial Pb(II) ion Concentration (mg/L) Bed height (cm) Flow rate (ml/min) Time (min) Outlet concentration (mg/L) Total number of data points

Page 23 of 42

ip t

f2 h ( x )   x Where

Transfer Function 3

f3 h ( x )   x Where

x

x

e e

 x  1 for  x  1

 x  1 for  x  1  x  0 for  x  0  x  1 for  x  1

1

1 e

te

f4 h ( x) 

 x

BP LM SCG BP LM SCG BP LM SCG BP LM SCG

Ac ce p

Transfer Function 4

 x

d

Transfer Function 2

x

e e

an

f1h ( x )  tanh  x 

Algorithm

M

Transfer Function 1

Equation

Optimum Number of Processing Elements 21 17 23 20 11 23 23 25 22 11 11 20

us

Transfer Function in Hidden Layer

cr

Table 8 Optimum numbers of processing elements in the hidden layer for four different transfer functions

Page 24 of 42

Table 9 Performance of the best Neural Network on the basis of minimum value of MSE reached during Cross-validation

ip t

Algorithm LevenbergMarquardt

Scaled Conjugate Gradient

0.001401

0.001115

0.001320

0.001547

0.001995

0.001529

0.001861

0.006114

0.003472

0.001689

us

cr

Backpropagation

0.003207

an

0.001611

Ac ce p

te

d

M

Transfer Function in Hidden Layer Transfer Function 1 Transfer Function 2 Transfer Function 3 Transfer Function 4

Page 25 of 42

Table 10 Performance of the different algorithms for prediction of Concentration by the optimized neural network for different transfer functions in case of Testing

Transfer Function 3

ip t

cr

0.018425 0.017619 0.098629 0.994407 0.024703 0.023266 0.204663 0.987331 0.039176 0.036262 0.420359 0.975309 0.018636 0.022347 0.141734 0.991617

Ac ce p

te

d

Transfer Function 4

0.015295 0.013362 0.078650 0.995576 0.018053 0.015764 0.109637 0.993498 0.024790 0.019160 0.171588 0.989066 0.022818 0.020319 0.163826 0.989860

Scaled Conjugate Gradient 0.017674 0.019061 0.121066 0.993902 0.018454 0.017237 0.121333 0.993335 0.030700 0.025191 0.274959 0.983333 0.021645 0.017121 0.139167 0.991390

us

Transfer Function 2

AARE SD (σ) MSE CCC (R) AARE SD (σ) MSE CCC (R) AARE SD (σ) MSE CCC (R) AARE SD (σ) MSE CCC (R)

BackLevenbergpropagation Marquardt

an

Transfer Function 1

Algorithm Measurement Type

M

Transfer Function in Hidden Layer

Table 11 Performance of the different algorithms for prediction of Concentration by the optimized neural network for different transfer functions in case of final prediction

Page 26 of 42

Transfer Function 3

ip t

cr

Ac ce p

te

d

Transfer Function 4

us

Transfer Function 2

AARE SD (σ) MSE CCC (R) χ² AARE SD (σ) MSE CCC (R) χ² AARE SD (σ) MSE CCC (R) χ² AARE SD (σ) MSE CCC (R) χ²

Scaled BackLevenbergConjugate propagation Marquardt Gradient 0.018567 0.019336 0.017812 0.019172 0.018563 0.016772 0.172677 0.178382 0.132727 0.999632 0.999602 0.999707 0.605563 0.606916 0.475305 0.024784 0.020458 0.017961 0.021914 0.021144 0.015227 0.249541 0.180766 0.125521 0.999451 0.999598 0.999734 0.883741 0.670449 0.445399 0.023086 0.039027 0.031414 0.020929 0.037442 0.025984 0.217732 0.544096 0.322914 0.999588 0.998777 0.999297 0.792973 2.096347 1.224743 0.025149 0.019663 0.019860 0.026463 0.021084 0.019619 0.292204 0.187322 0.182402 0.999373 0.999580 0.999621 1.086233 0.668037 0.647444

an

Transfer Function 1

Algorithm Measurement Type

M

Transfer Function in Hidden Layer

Page 27 of 42

Highlights Hyacinth root used for removal of Pb(II) from aqueous solution. The sorption mechanism was illustrated using FTIR studies. Thomas model predicted the column sorption curves successfully.

ip t

The ANN model with Backpropagation algorithm, using single hidden layer with

Ac ce p

te

d

M

an

us

cr

transfer function 2 and 20 processing elements predicted the outlet concentration successfully.

Page 28 of 42

Ac

ce pt

ed

M

an

us

cr

ip t

Graphical Abstract (for review)

Page 29 of 42

Figure 1

Ac

ce pt

ed

M

an

us

cr

ip t

Figure

Page 30 of 42

ip t

1.0

us

cr

0.8

an

C t / Co

0.6

M

0.4

ed

0.2

ce pt

0.0

100

200

300

400

500

Time, min

Figure 2

Ac

0

-1

Influent Pb(II) Conc.: 20 mg L Bed height: 6 cm Symbol Flow rate -1 (ml min ) 10 30 40

Page 31 of 42

ip t

30

cr

25

us

20

10

M

Co, mg L

-1

an

15

ed

5

ce pt

0

Flow rate: Bed height: Symbol

-5

30 ml min-1 5 cm Influent Pb(II) conc. (mg L-1) 10 20 30

-10

Ac

0

100

200

300

400

500

Time, min

Figure 3

Page 32 of 42

ip t

1.0

us

cr

0.8

an

0.4

M

C t / C0

0.6

ed

0.2

ce pt

0.0

Ac

0

100

200

-1 Influent Pb(II) Conc.: 30 mg L -1 Flow rate: 30 ml min System Bed height (cm) 4 5 6

300

400

500

Time, min

Figure 4

Page 33 of 42

ip t

0.0

cr

-3.0

us

Symbol Influent FLow Bed conc. rate depth -1 -1 (mg L ) (ml min )(cm) 10 10 4 10 10 5 10 10 6 20 10 6 20 30 4 30 40 5

-4.5

-6.0 0

100

200

an

ln(Ct/C0)

-1.5

300

400

M

Time, min

Ac

ce pt

ed

Figure 5 Bohart-Admas and Wolborska model plot

Page 34 of 42

8

System

7

Influent

Flow

conc.

rate

-1

Bed depth

-1

4

10

4

10

6

20

10

5

20

10

6

20

40

5

20

30

6

30

10

4

30

30

6

3

us

0

ln[(Ct/C )-1]

5

10 10

cr

6

ip t

(mg L ) (ml min ) (cm)

1

0 0

20

40

60

80

100

an

2

120

140

160

180

200

M

Time, min

Ac

ce pt

ed

Figure 6 Thomas model plot

Page 35 of 42

0.0

ip t Symbol

Influent

Flow

conc.

rate

-1 (mg L )

-4.5

10 10 10 -6.0

20

30

0

100

Bed

depth -1 (ml min ) (cm) 10

5

10

6

30

4

10

6

40

6

10

4

an

20

-7.5

cr

-3.0

us

ln[Ct/(C0-Ct)]

-1.5

200

300

M

Time, min

Ac

ce pt

ed

Figure 7 Yoon-Nelson model plot

Page 36 of 42

ip t cr us an M

ed

a. Fresh hyacinth root b. Pb(II) loaded hyacinth root

Ac

ce pt

Figure 8

Page 37 of 42

ip t cr us

Ac

ce pt

ed

M

an

Figure 9

Page 38 of 42

0.030

Algorithm : Backpropagation Transfer Function 1 Transfer Function 2 Transfer Function 3 Transfer Function 4

ip t cr

0.020

0.015

us

Average MSE for Cross-validation

0.025

0.010

10

an

0.005

20

M

Number of Processing Elements in hidden layer

Ac

ce pt

ed

Figure 10

Page 39 of 42

2

0

10

-1

10

-2

10

-3

10

-4

ip t

10

cr

1

us

10

Algorithm : Levenberg-Marquardt Transfer Function 1 Transfer Function 2 Transfer Function 3 Transfer Function 4

1

an

Average MSE for Cross-validation

10

10

100

1000

M

Number of epochs

Ac

ce pt

ed

Figure 11

Page 40 of 42

ip t cr us

10

10

an

Outlet Concentration, Predicted mg/l

20

Algorithm : Scaled Conjugate Gradient Transfer Function 1 Transfer Function 2 Transfer Function 3 Transfer Function 4

20

M

Outlet Concentration, Experimental mg/l

Ac

ce pt

ed

Figure 12

Page 41 of 42

an

us

cr

Outlet Concentration, Predicted mg/l

ip t

Algorithm : Backpropagation Number of hidden layer: 1 Transfer Function: 2 Number of Processing Elements: 20

0.01 0.01

M

Outlet Concentration, Experimental mg/l

Ac

ce pt

ed

Figure 13

Page 42 of 42