Chemical Engineering Journal 228 (2013) 10981109

Contents lists available at SciVerse ScienceDirect

Chemical Engineering Journal

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

Modeling of xed-bed adsorption of uoride on bone char using a hybrid

neural network approach
R. Tovar-Gmez a, M.R. Moreno-Virgen a, J.A. Dena-Aguilar a, V. Hernndez-Montoya a,
A. Bonilla-Petriciolet a,, M.A. Montes-Morn b
a
Instituto Tecnolgico de Aguascalientes, Aguascalientes, Mexico
b
Instituto Nacional del Carbn, Oviedo, Spain

h i g h l i g h t s g r a p h i c a l a b s t r a c t

 A model based on Thomas equation

and ANNs has been proposed to
predict breakthrough curves.
 This model has been used to predict
uoride adsorption breakthrough
curves using bone char.
 This modeling approach is useful for
process systems engineering of xed-
bed columns.

a r t i c l e i n f o a b s t r a c t

Article history: This study introduces a hybrid model based on the Thomas equation and articial neural networks
Received 26 January 2013 (ANNs) for the modeling of unsymmetrical breakthrough curves obtained from the uoride adsorption
Received in revised form 17 May 2013 on bone char. Experimental results of kinetics, isotherms and breakthrough curves of uoride adsorption
Accepted 18 May 2013
on two commercial bone chars have been used for analyzing the capabilities and limitations of this hybrid
Available online 30 May 2013
ANN model. Performance of this hybrid model has been studied and compared with respect to the results
of traditional linear regression of the Thomas breakthrough equation at different operating conditions of
Keywords:
packed-bed adsorption columns. Results showed that an improvement in the modeling capabilities of
Bone char
Fluoride
Thomas model can be obtained using ANNs. Specically, the hybrid ANNsThomas model showed deter-
Adsorption mination coefcients higher than 0.9 and its average mean square errors are lower, up to 86%, than those
Neural network model obtained with the linear modeling approach. In fact, the present study illustrates that the improper
Water treatment handling of the Thomas model using traditional regression approach may lead to imprecise values of
design parameters and erroneous conclusions of adsorption performance. On the other hand, the hybrid
ANNsThomas model is useful for determining reasonable and accurate design parameters of packed-bed
adsorption columns. This modeling approach can be useful for the process system engineering of
dynamic adsorption systems involved in the eld of water treatment and purication.
2013 Elsevier B.V. All rights reserved.

1. Introduction composition of water wells) and anthropogenic sources (e.g., the

Fluoride pollution in water resources for human consumption is
considered a signicant environmental problem in several devel-
oping countries of America, Africa and Asia [1,2]. Water pollution
by uoride is caused by both natural factors (e.g., geochemical
Corresponding author.
E-mail address: petriciolet@hotmail.com (A. Bonilla-Petriciolet).
use of fertilizers and semiconductors). In particular, water con-
sumption with uoride concentrations higher than 1.5 mg/L is
toxic for human beings and, therefore, this uoride concentration
limit has been established as the guideline value by the WHO for
drinking water [3]. Traditional symptoms of a chronic exposure
to uoride are the dental and skeletal uorosis. However, studies
reported in literature have documented different lesions in other
organs of human being such as the endocrine glands, liver and

1385-8947/$ - see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.cej.2013.05.080
 R. Tovar-Gmez et al. / Chemical Engineering Journal 228 (2013) 10981109
thyroid [4]. Therefore, water deuoridation is fundamental to meet
environmental regulations for water quality, to prevent public
health risks and to reduce the prevalence of dental and skeletal
uorosis and other uoride-related diseases in endemic areas.
Adsorption process is considered the most effective and rent-
able method for uoride removal from water and it offers advanta-
ges with respect to results obtained with other water treatment
and purication technologies such as the electrodialysis, mem-
brane processes and chemical precipitation [4,5]. Literature on
adsorption process indicates that a wide variety of natural and syn-
thetic adsorbents has been tested and applied for the removal of
uoride ions from aqueous solution [521]. These adsorbents in-
clude several types of activated carbons [79], polymers [10], clays
[11], industrial wastes [12], biomasses [1315] and others materi-
als [1621]. In particular, the modication of surface chemistry of
adsorbents using multivalent metallic species (e.g., lanthanum,
magnesium, zirconium) appears to be a promising approach for
developing novel materials for water deuoridation [14,22,23].
However, these promising adsorbents are not still commercially
available. In the case of commercial options available in the world-
wide market, the activated alumina [4,24] and the bone char [25
27] are the most used and effective adsorbents for uoride removal
from water.
Water deuoridation systems based on the adsorption process
generally use the conguration of packed-bed columns
[10,24,2832]. This separation system is very exible and easy in
terms of design and operation; and it offers several advantages
such as there is an effective contact between the adsorbent and
the uid to be treated, the adsorption rates are favored due to
the adsorbent is continuously in contact with a fresh solution, a
large-scale water treatment is feasible in a reduced time, the
adsorbent regeneration can be performed in the same column,
and the equipment and supplies required for the column operation
are not expensive. Overall, this treatment system shows a high de-
gree of stability and reliability at a wide range of operating condi-
tions and it can practically meet strict regulations of water quality
[33].
Performance of packed-bed adsorption columns are character-
ized via the breakthrough curve, i.e., the efuent concentration
prole versus time (for a constant ow rate) or volume of treated
efuent. The quantitative characterization of the performance of
the packed-bed adsorption columns involves the modeling and
prediction of this breakthrough curve. Note that this modeling
stage is fundamental for the design and optimization of this sepa-
ration system. In fact, extensive studies in pilot plant scale can be
avoided if the breakthrough curves for adsorption columns can be
reliably predicted using laboratory measurements [33].
The reliable modeling of breakthrough curves requires the
application of proper numerical procedures. The modeling ap-
proaches used for data tting of breakthrough curves include the-
oretical (e.g., mass transfer and kinetic models), empirical and
semi-empirical equations. Note that the modeling of packed-bed
adsorption processes using a theoretical and rigorous model usu-
ally requires the application of numerical methods, which may
show convergence problems due to poor initialization and non-lin-
earity of the problem to be solved. Alternatively, several simple
breakthrough equations have proposed for data correlation of dy-
namic adsorption experiments. For example, these equations in-
clude BohartAdams model [34], Thomas model [35], Yoon
Nelson model [36], Clark model [37] and the error-function model
[38]. However, these models are suitable for the data correlation of
symmetrical breakthrough curves but they may be inadequate and
fail to describe the performance of packed-bed columns with a
complex adsorption process (i.e., unsymmetrical breakthrough
curves). It is convenient to remark that the modeling of break-
through curves is challenging due to the nonlinearity of equations
1099
used for describing the equilibrium, kinetics and mass transport
phenomena involved in dynamic adsorption processes [39]. In fact,
some authors have recognized that the development of break-
through models for describing accurately the performance of
packed-bed columns is a difcult task [40].
Based on the fact that accurate and simple breakthrough mod-
els are required for the proper design of adsorption process, in
this study we propose a hybrid neural network approach for
the modeling of unsymmetrical adsorption breakthrough curves.
Specically, articial neural networks (ANNs) are used to improve
the modeling capabilities of traditional breakthrough equations
for adsorption systems involving priority water pollutants. Litera-
ture indicates that ANNs have been widely used to model com-
plex relationships between inputs and outputs or to nd
patterns in selected data [41]. For the case of dynamic adsorption
systems, some studies have reported the application of ANNs as a
black-box model (i.e., an empirical model that relies on measure-
ments only) for data tting of breakthrough curves [33,42,43]. In
this study, a hybrid approach based on the combination of Tho-
mas breakthrough equation [35] and ANNs (i.e., a gray-box mod-
el) has been used to model the performance of packed-bed
adsorption columns for the uoride removal from water using
bone char. This modeling approach has been used for data analy-
sis in other science elds [44,45] and, to the best of our knowl-
edge; this is the rst application of ANNs to improve the
performance of traditional breakthrough equations in adsorption
research. Capabilities and limitations of this approach for model-
ing dynamic adsorption processes are discussed and analyzed.
Our results show that an improvement in the modeling capabili-
ties of Thomas model can be obtained using ANNs. Finally, the
proposed approach can be useful for the process system engineer-
ing of dynamic adsorption systems involved in the eld of water
treatment and purication.
2. Methodology
2.1. Bone chars used in adsorption experiments
Fluoride adsorption experiments were performed using two
commercial bone chars: BCM from Carbones Mexicanos (Mexico)
and BCB (Brimac 216) from Brimac Carbon Services (United King-
dom). Table 1 provides the general specications of these adsor-
bents provided by the supplier. Physicochemical properties and
textural parameters of these adsorbents were determined in this
study. Specically, the elemental composition of bone chars was
determined with a LECO CHNS-932 elemental analyzer and the
oxygen content with a LECO VTF-900. Textural parameters were
calculated from the nitrogen adsorption isotherms at 77 K, which
were obtained with a Micromeritics TriStar II 3020. Organic
functional groups were identied by FT-IR spectroscopy using a
Thermo Nicolet 6700 spectrophotometer and the diffraction pat-
terns of these adsorbents were recorded in a Bruker D8 Advance
diffractometer equipped with a Cu Ka X-ray source operated at
40 kV and 40 mA. A single Gbel mirror conguration was used
Table 1
Supplier specications of commercial bone chars used in packed-bed adsorption
columns for uoride removal from water.
Propertya Bone char
Carbones Mexicanos (BCM) Brimac 216 (BCB)
Carbon content 1015% 911%
Tricalcium phosphate 8085% 7076%
Specic surface area 104 m2/g 129 m2/g
Humidity <4% <5%
a
Adsorbent specications provided by the supplier.
1100 R. Tovar-Gmez et al. / Chemical Engineering Journal 228 (2013) 10981109
to monochromatise and focus the X-rays on the samples attaining
highly efcient parallel beam geometry. Step scanning, with a step
size of 0.02 2h and a scan step time of 5 s, was used for collecting
the diffraction data. Additionally, the physical morphology of bone
chars was observed using a FE-SEM system (Quanta FEG 650, FEI)
equipped with an EDX analyzer (Ametek-EDAX). Adsorbent parti-
cles were dispersed on a graphite adhesive tab placed on an alumi-
num stub and no further coating was required. A semi-quantitative
analysis of the inorganic elements was obtained by EDX and an
average of three punctual analyses on the bone char surface at
20 lm was reported.
2.2. Fluoride adsorption experiments in packed-bed columns using
bone chars
Dynamic adsorption experiments were performed in a Pyrex
glass column with 2.5 cm of internal diameter and 13 cm of
length. This column was packed with 27.69 g (i.e., 7.5 cm of pack-
ing height) of each commercial boner char (particle size: 2035
mesh fraction) and the bed porosity ranged from 25% to 30% for
both adsorbents in all adsorption experiments. Fluoride break-
through curves were obtained using each bone char and different
feed uoride concentrations (940 mg/L) and feed ow rates (3.3
and 6.6 mL/min). Note that these pollutant concentrations have
been selected by considering that the groundwater extracted in
several regions of Mexico and other countries may show uoride
concentrations up to more than 30 mg/L [2,5,9]. On the other
hand, the range of feed ow rates corresponds to 1.53.0 min
of residence time, which are common conditions for the operation
of packed-bed adsorption columns [46,47]. Our experience in
adsorption process used dynamic conditions indicated that resi-
dence times higher than 3 min do not improve signicantly the
performance of removal process due to limitations caused by
mass transfer phenomena. Fluoride solutions were prepared using
analytical grade sodium uoride (NaF) and deionized water. All
adsorption experiments were performed at pH 7 and 30 C using
a peristaltic pump and an upow operation mode, which was
used to avoid an excessive head-loss and channeling. Several
efuent samples were collected from the packed-bed column at
regular time intervals and the uoride concentration in all sam-
ples was determined to obtain the breakthrough curve. Fluoride
concentrations were quantied using a uoride ion selective elec-
trode (Denver Instruments, UP-25) and TISAB buffer (Hach) fol-
lowing the methodology described in the Standard Methods of
Examination of Water and Wastewater [48]. A linear calibration
curve was used for determining the uoride concentration in all
samples. All experiments were performed twice and the average
results were used for data analysis. Finally, uoride adsorption
kinetics and isotherms were obtained using batch reactors for
both bone chars at 30 C, pH 7 and a mass-volume ratio of
10 g/L. These kinetics and isotherms were used in adsorption data
analysis, for validating the application of Thomas model and for
comparing the adsorbent performance in the packed-bed columns
with respect to the results obtained using batch adsorption
systems.
2.3. Analysis and modeling of the breakthrough curves of uoride
adsorption on bone char
2.3.1. Calculation of design parameters of breakthrough curves
Design parameters of the packed-bed columns were calculated
from the experimental data of the uoride breakthrough curves
using both bone chars. Specically, the mass transfer zone (MTZ)
was calculated for different operating conditions of packed-bed
columns using the next expression
 
te  tb
MTZ L 1
te
where tb is the breakthrough time, te is the exhaustion time and L is
the bed height in cm. For this analysis, tb was dened as the time
when [F]/[F]0 = 0.05, while te corresponds to [F]/[F]0 = 0.95
where [F] is the efuent uoride concentration and [F]0 is the
feed uoride concentration both in mg/L, respectively. The overall
adsorption zone (Dt) was also determined using
Dt t e  t b 2
On the other hand, the retardation factor (rf) was used to deter-
mine the rate at which the pollutant moves within the packed-bed
column [46]. This parameter has been obtained from the treated
volume per void volume that gives [F]/[F]0 = 0.5 or, equivalently,
the operating time (t50%) of the packed-bed column when the efu-
ent concentration has reached 50% of the inuent concentration,
then
V 50%
rf 3
ALe
where e is the void volume of the packed-bed column, A is the cross
sectional area of the column and L is the bed height of the column,
respectively. Retardation factors were estimated for all operating
conditions used in uoride adsorption experiments.
Finally, the bed adsorption capacity (qbed) and the maximum
adsorption capacity obtained from uoride adsorption isotherms
(qe,max) were used for the estimation of the lter material usage
rate (Fq)
qbed
Fq 4
qe;max
where qbed and qe,max are given in mg/g and are obtained at the same
conditions of temperature and pH. Note that qbed can be calculated

via the integration of the experimental data FF t versus t, or using a
0
specic breakthrough equation (e.g., Thomas model). All these de-
sign parameters were used for characterizing the performance of
both bone chars in uoride adsorption columns at different operat-
ing conditions.
2.3.2. Data tting of uoride breakthrough curves using Thomas model
Thomas model [35] was used for modeling the uoride adsorp-
tion breakthrough curves using commercial bone chars. This
breakthrough model is based on Langmuir-type adsorption
desorption and assumes isothermal and isobaric operating
conditions, a constant column void fraction, and the radial and
axial dispersion are negligible in the packed bed column. For the
case of uoride adsorption, this model is given by
F  t 1
  5
F  0 1 exp kTh q 
Q bed;T W bed  F 0 V eff
where Veff is the treated volume of uoride solution in L, Q is the
feed ow rate in L/h, Wbed is the amount of bone char packed in
the columns and is given in g, qbed,T is the adsorption capacity of
the packed-bed column estimated by the model given in mg/g
and kTh is the Thomas rate constant reported in L/h mg, respectively.
Parameters kTh and qbed,T can be obtained from a linear regression
approach using the following linearized form of Thomas model
  
F 0 kTh qbed;T W bed kTh F  0 V eff
ln  1  6
F t Q Q
 

A plot of ln FF 0  1 versus t at the given operating conditions
t
(i.e., feed ow rate, temperature and pH) is used for this linear
 R. Tovar-Gmez et al. / Chemical Engineering Journal 228 (2013) 10981109
regression. Excel Solver has been used for performing the linear
regression of Thomas model using the uoride breakthrough
curves of tested bone chars.
2.3.3. A gray-box model using ANNs and Thomas equation for data
tting of breakthrough curves
As stated, the linear regression approach is extensively used for
determining the parameters of Thomas model and other traditional
breakthrough equations. However, this approach may fail to t the
experimental adsorption data of unsymmetrical breakthrough
curves. In fact, our experience and results of other studies on
adsorption processes for water treatment indicate that the experi-
mental breakthrough curves obtained for the removal of uorides
and other priority pollutants (e.g., heavy metals, dyes) are com-
monly unsymmetrical, e.g., [42,43,46,47]. Based on this fact, we
have used a hybrid approach for modeling the breakthrough curves
using ANNs and Thomas equation. Herein, it is important to remark
that some authors have used ANNs for modeling adsorption break-
through curves [33,39,43]. For example, Cavas et al. [43] have re-
ported the comparison of the results obtained with the Thomas
model and those results from the application of ANNs for modeling
breakthrough curves. However, no attempts have been performed
to combine both ANNs and breakthrough equations for modeling
the experimental data of dynamic adsorption systems.
ANNs are black-box (i.e., empirical) computational models in-
spired by the biological neural networks. ANNs have the capabili-
ties for predicting the performance of complex and non-linear
problems and for nding patterns in data analysis relying only
on measurements (inputs and outputs) [41]. A neural network con-
sists of input, hidden and output layers, which are constituted by
Fig. 1. Illustration of a neuron in an ANN model.
Fig. 2. Articial neural network architecture used for modeling the uoride adsorption breakthrough curves using bone char.
1101
an interconnected group of articial neurons. These articial neu-
rons are interconnected to each other via connection weights,
which represent the relative strength of an input neuron in con-
tributing to the output neuron, see Fig. 1. Formally, the net input
Yij of the neuron j in the layer i is given by
ni1
X
Y ij wijk V i1;k hij 7
k1
V ij gY ij 8
where wijk is the connection weight, Vik is the neuron input and hij is
the neuron bias. An activation function g(Yij) is used to calculate the
neuron output Vij given the set of neuron inputs. The most common
activation function is the sigmoid function, which has been used in
this study for the neurons of hidden layers; while an identity func-
tion (i.e., Vij = Yij) was used in the output layer. A training process is
used to determine suitable values of w and h for each neuron of the
ANNs model. Note that this training is performed using input data
and target output values obtained from the system to be studied.
So, we have used the classical backpropagation algorithm for ANNs
training, which is a rst order gradient descent method.
In this study, a hybrid model using a feed-forward ANNs archi-
tecture and Thomas breakthrough equation has been proposed for
data tting of uoride breakthrough curves, see Fig. 2. In particular,
this hybrid model imposes non-negativity constraints for the val-
ues of parameters of Thomas model: kTh and qbed,T. The handling
of these constraints has been performed using the procedure pro-
posed by Dena-Aguilar [45]. Before the ANNs training process, the
values of input neurons were normalized. Parameters of Thomas

model, which were used to calculate the values of FF t ; were con-
0
sidered as the targets of the ANNs model subject to the non-nega-
tivity restrictions: kTh > 0 and qbed,T > 0. These restrictions have
been considered for the hybrid model because both parameters
of Thomas equation have a physical and theoretical meaning.
Without these restrictions, the hybrid model may nd negative
values for these model parameters thus losing its physical
signicance.
The training of the hybrid model was performed via the minimi-
zation of the following objective function
0  12
 exp   calc 
1X nn X
ndat
@ F  t  F  
F obj  t  A 9
2 m1 j1 F  0 j F 0 
m;j
1102 R. Tovar-Gmez et al. / Chemical Engineering Journal 228 (2013) 10981109

where ndat is the number of experimental data used for the training
of hybrid model, nn is the number of output neurons and super-
scripts calc and exp indicate the calculated and experimental values,
respectively. Fortran programs were used for data modeling of
uoride breakthrough curves using this hybrid model based on
ANNs and Thomas equation.
For ANNs training, the input data include the feed uoride con-
centration ([F]0), the operation time of packed bed column (t) and
the feed ow (Q) obtained from uoride breakthrough curves. Note
that previous studies have used these input variables for the mod-
eling of breakthrough curves using black-box ANNs models
[33,39,43] because they have a statistical inuence on the perfor-
mance of dynamic adsorption process. Additionally, this set of in-
put variables is uncorrelated, which is in agreement to the basis
of the principal component analysis approach. The architecture of
hybrid ANN model was determined based on preliminary calcula-
tions using a trial and error procedure where the combination of
input variables, the number of hidden layers and neurons were
changed until nding a proper ANN architecture. Several ANN
models were analyzed and they included different congurations
of the input variables, i.e.: tQ, t[F]0, [F]0Q and tQ[F]0,
where the hybrid ANN model with three input variables offered
the best performance. On the other hand, the results of sensitivity
analysis indicated the following trend for the most inuential
variables: t > [F]0 > Q. Therefore, the ANNs architecture used for
modeling the uoride adsorption breakthrough curves consisted
of an input layer with three neurons tQ[F]0, two hidden layers
with 18 neurons for each layer, and one output layer, see Fig. 2. It is
convenient to remark that our calculations indicated that high
number of hidden neurons (i.e., >18) caused overtting of the mod-
el. Finally, the results obtained from the linear regression approach
and those obtained from the hybrid model have been compared
and analyzed using statistical criterions and design parameter cal-
culations of packed-bed columns.
3. Results and discussion
3.1. Breakthrough curves of uoride adsorption using commercial bone
chars
Experimental data of uoride breakthrough curves for both
bone chars at different operating conditions of packed-bed col-
umns are reported in Fig. 3. As expected, the characteristics and
shape of uoride breakthrough curves depend on the operating
conditions of adsorption columns (i.e., feed ow rate, uoride ini-
tial concentration and adsorbent type). However, it is clear that
all the uoride breakthrough curves of both adsorbents show a
common S-shape but they are unsymmetrical. Note that all break-
through curves obtained for the commercial bone chars showed a

1 1
(a) (b)

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

ANNs-Thomas model
0 0
[F -]/[F -]0 0 20 40 60 80 0 10 20 30 40

1 1

(c) (d)
0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2
9 mg/L
20 mg/L
40 mg/L
0 0
0 50 100 150 200 0 20 40 60 80

Time, h

Fig. 3. Fluoride breakthrough curves using packed-bed adsorption columns and commercial bone chars. BCB: (a) 3.3 and (b) 6.6 mL/min; BCM: (c) 3.3 and (d) 6.6 mL/min.
 R. Tovar-Gmez et al. / Chemical Engineering Journal 228 (2013) 10981109 1103

slow approach of [F]t/[F]0 ? 1.0, which is a common characteris-

tic of adsorption processes in a liquid phase where the diffusion
phenomena are the rate-limiting mass transport process [40].
In particular, the feed ow rate Q is a crucial parameter for the
operation of packed-bed adsorption columns. This operating
parameter has a direct impact on the mass transfer phenomena in-
volved in the adsorption process under dynamic conditions. Fig. 3
also shows the effect of feed ow rate on the performance of both
commercial bone chars for the uoride removal from aqueous
solution in the xed-bed columns. It is clear that there is a reduc-
tion in the values of both tb and te with the increment of the feed
ow rate because the contact time between the adsorbent and
the uid to be treated is not enough for reaching the adsorption
equilibrium thus limiting the diffusion and mass transfer of uo-
ride ions into the adsorbent. On the other hand, the feed uoride
concentration also inuences the performance of adsorption col-
umns using both bone chars. The adsorption columns operated at
the lowest feed uoride concentration (i.e., 9 mg/L) exhibited high-
er operating times than those obtained for the columns receiving
feed uoride concentrations of 20 or 40 mg/L. Note that the change
of concentration gradient affects the adsorbent performance in
packed bed columns [33] and, consequently, the saturation rate
and the breakthrough time depends on the pollutant feed concen-
tration. As expected, the breakthrough time of adsorption columns
increased with a decrement in the feed uoride concentration. In
summary, the exhaustion of uoride adsorption columns using
both bone chars is faster if a high feed uoride concentration
and/or high feed ow rate are used, see Fig. 3. The service time
of bone char packed columns used for uoride removal decreased
with both feed ow rate and inlet uoride concentration. In partic-
ular, the uoride breakthrough curves of BCM required the longest
operating time to reach the breakthrough and exhaustion points Fig. 4. Fluoride adsorption kinetics using commercial bone chars: (a) BCB and b)
than those breakthrough curves of BCB at the same operating con- BCM at 30 C and pH 7. Initial uoride concentration: (s) 9 and (e) 40 mg/L. ()
Pseudo-second order kinetic model.
ditions. Experimental operating times of adsorption columns for
the exhaustion of BCB ranged from 12.7 to 58.3 h, and from 12.4
to 178.0 h for the BCM at tested experimental conditions. There-
3
fore, characteristics, shape and position of the breakthrough curves Brimac
indicate that the capacity of adsorption columns is different for Carbones Mexicanos
Adsorption capacity, mg/g

each adsorbent and more uoride ions can be removed using Langmuir model
BCM in comparison to the packed-bed columns using BCB. 2
With illustrative purposes, Figs. 4 and 5 show the uoride
adsorption kinetics and isotherms obtained at batch conditions
for both bone chars at 30 C and pH 7. Results of uoride adsorp-
tion kinetics have been modeled using both pseudo rst- and sec- 1
ond-order kinetic models, see Table 2. It is clear that the pseudo-
second order model offers the best correlation for the uoride
adsorption rates on the commercial bone chars BCB and BCM. On
the other hand, uoride adsorption isotherms of Fig. 5 were tted 0
0 5 10 15 20
to Langmuir, Freundlich and Sips models and results of data corre-
lation are reported in Table 3. Both Langmuir and Sips are suitable Equilibrium fluoride concentration, mg/L
for the correlation of uoride adsorption isotherms on commercial Fig. 5. Fluoride adsorption isotherms using commercial bone chars at 30 C and
bone chars. Based on this fact, we have used the Langmuir pH 7.

Table 2
Results of kinetic data modeling of uoride adsorption on commercial bone chars.

Bone char [F]0 (mg/L) Results of kinetic data of uoride adsorption on bone char
Pseudo rst-order modela Pseudo second-order modelb
2
1
k1 (h ) qte (mg/g) R k2 (g/mg h) qte (mg/g) R2
BCB 9 0.218 0.936 0.994 0.141 1.248 0.996
40 1.514 1.872 0.966 0.869 2.082 0.978
BCM 9 0.218 0.939 0.990 0.191 1.152 0.996
40 0.243 2.763 0.986 0.052 3.692 0.991
a
qt qte 1  ek1 t .
b q2 k t
qt 1qte 2k2 t.
te
1104 R. Tovar-Gmez et al. / Chemical Engineering Journal 228 (2013) 10981109

Table 3 suggest that although the linear regression method is the tradi-
Results of isotherm data modeling of uoride adsorption on commercial bone chars.
tional and widely used approach for modeling breakthrough data
Model Bone char with Thomas equation, it may not be adequate in describing and
BCB BCM predicting the experimental data of unsymmetrical breakthrough
Langmuir
curves. Note that the calculated values of kTh and qbed,T may vary
K L qm C e
qe 1K KL, L/mg 0.027 0.129 signicantly for each operating condition used for adsorption col-
L Ce
qm, mg/g 3.614 4.510 umns, see Table 4. Specically, the values of kTh ranged from
R2 0.935 0.943 0.0088 to 0.0198 L/h mg for BCB and from 0.0024 to 0.0132 L/
Fobj 0.090 0.166
h mg for BCM, while the calculated values of qbed,T ranged from
Freundlich 0.47 to 2.47 mg/g and from 2.26 to 5.00 mg/g for these bone chars,
qe K f C 1=n1 Kf 0.114 0.519
e respectively.
n1 1.267 1.461
R2 0.856 0.847 On the other hand, results of data tting of uoride dynamic
Fobj 0.175 0.268 adsorption experiments using the hybrid model (i.e., ANNsTho-
Sips mas model) are also reported in Table 4. Specically, the 75% of
q a Cn
m s e
qe 1a qm, mg/g 1.815 2.525 the experimental data of two bone chars were used for the training
Cns e
as 0.026 0.232 stage of this hybrid model: 186 and 198 experimental data for BCB
n 1.579 1.941
and BCM bone chars, respectively; while the 25% remaining data
R2 0.995 0.970
Fobj 0.008 0.076
(i.e., 62 and 66 experimental points) were used as the verication
and test sets. We have used two approaches for determining the
parameters of the hybrid model: (a) all uoride breakthrough
curves of each bone char has been simultaneously used for the
isotherm model for determining the monolayer uoride adsorption training of the ANNsThomas model, and (b) each individual
capacities (qmax): BCM: qmax = 4.51 mg/g and BCB: qmax = 3.62 mg/g, breakthrough curve has been used alone for the training of the
respectively. Results of uoride breakthrough curves are consistent ANNsThomas model. The same ANNs architecture given in Fig. 2
with the batch adsorption results and the monolayer adsorption has been used for both variations of the ANNsThomas model. As
capacities obtained from Langmuir model, which indicate that expected, both modeling approaches provide different results for
BCM has the greatest uoride adsorption capacity in both static the modeling of packed bed adsorption columns, see Table 4. In
and dynamic conditions. Note that these maximum adsorption particular, the hybrid model based on all breakthrough curves for
capacities from Langmuir model has been used for the calculation ANNs training offers the exibility of reliably predicting the perfor-
of the lter material usage rate (Fq) during the analysis of design mance of packed-bed adsorption columns at different operating
parameters of bone char packed-bed columns (see Table 4). conditions using constant and unique values of both kTh and qbed,T
for Thomas equation. In contrast, the modeling of each individual
breakthrough curve using the hybrid model provides specic val-
3.2. Results of data modeling of breakthrough curves for uoride
ues for the parameters of Thomas model, see results reported in
adsorption using bone chars
Table 4. However, the calculated values of Thomas model by this
hybrid approach showed a less variability than those obtained
Results of the linear regression of Thomas model for both bone
from the application of the linear regression approach specially
chars are given in Fig. 6, while Table 4 provides the model param-
for the parameter qbed,T. In general, the kinetic coefcients kTh cal-
eters. Herein, it is convenient to note that the uoride adsorption
culated by all models increased with the inlet uoride concentra-
kinetics and isotherms conrmed that the assumptions of Thomas
tion and the feed ow rate suggesting that the driving force of
model are satised and, consequently, this model is valid for pre-
uoride mass transfer in the liquid lm is also increased. Similar
senting and analyzing the uoride breakthrough results. In general,
trends have been observed for parameters of Thomas model in
the linear approach of Thomas model gave a suitable t for data
other studies on the removal of water pollutants using packed-
modeling of the uoride adsorption experiments using BCB and
bed adsorption columns [43].
the best correlation results are obtained for adsorption experi-
For comparing tested models, we have calculated the average
ments performed at the lowest feed ow rate. However, the Tho-
Mean Square Error (MSE) between experimental and predicted
mas model gave a poor t of some breakthrough curves obtained
breakthrough curves of uoride adsorption using each bone char
for BCM especially at higher feed ow rate, see Fig. 6. These results

Table 4
Results of data modeling of uoride breakthrough curves using linear regression of Thomas model and the gray-box model based on ANNs and Thomas equation.

Bone char Flow rate (L/h) [F]0 (mg/L) Results of uoride breakthrough data modeling for
Thomas model ANNThomas model using all ANNThomas model using individual
breakthrough curves breakthrough curves
kTh (L/h mg) qbed,T (mg/g) MSE kTh (L/h mg) qbed,T (mg/g) MSE kTh (L/h mg) qbed,T (mg/g) MSE
BCB 0.198 9 0.0147 1.33 0.0168 0.0191 1.28 0.0138 0.0175 1.11 0.0057
0.396 0.0198 1.84 0.0268 0.0114 0.0237 1.24 0.0104
0.198 20 0.0088 2.47 0.0219 0.0147 0.0105 1.65 0.0059
0.396 0.0191 1.59 0.0153 0.0084 0.0332 1.13 0.0060
0.198 40 0.0095 2.00 0.0140 0.0113 0.0093 1.70 0.0055
0.396 0.0271 0.47 0.0621 0.0068 0.0247 1.38 0.0036
BCM 0.198 9 0.0036 3.59 0.0140 0.0088 2.14 0.0105 0.0050 2.60 0.0063
0.396 0.0053 3.09 0.0142 0.0173 0.0052 2.49 0.0120
0.198 20 0.0024 5.00 0.0307 0.0095 0.0102 2.00 0.0090
0.396 0.0091 3.27 0.0263 0.0137 0.0102 2.45 0.0135
0.198 40 0.0029 4.89 0.0408 0.0097 0.0098 2.01 0.0097
0.396 0.0132 2.26 0.0166 0.0205 0.0209 1.66 0.0036
 R. Tovar-Gmez et al. / Chemical Engineering Journal 228 (2013) 10981109 1105

6 8
(a) 9 mg/L (b)
20 mg/L
6
4 40 mg/L

4
2

0
0 20 40 60 80
0
0 10 20 30 40
-2
-2

-4 2
-4 2 R = 0.6384
2
R = 0.8014 R = 0.7225
2
2
2
R = 0.8481 R = 0.8818
R = 0.8081
-6 -6

6 8
(c) (d)
6
4

4
2

0
0 50 100 150 200
0
0 20 40 60 80
-2
-2 2
R = 0.5856
2
-4
2
R = 0.6581 R = 0.6313
R2 = 0.774 -4
2
R = 0.6071 2
R = 0.6923

-6 -6
Time, h

Fig. 6. Linear regression of Thomas model for uoride breakthrough curves using packed-bed adsorption columns and commercial bone chars. BCB: (a) 3.3 and (b) 6.6 mL/
min; BCM: (c) 3.3 and (d) 6.6 mL/min.

!2
F  exp  calc
1 X ndat
t;i  F t;i
MSE  10
ndat i1 F 0
MSE values of the different modeling approaches are reported
in Table 4. In general, the ANNsThomas models provided a better
t (i.e., lowest values of MSE) of the experimental uoride break-
through curves than that obtained with the linear regression of
Thomas model. However, it is convenient to remark that the major
deviations between predicted and experimental breakthrough
curves were observed at early operating times of adsorption col-
umns (i.e., the beginning of column operation) for all tested mod-
els. For illustration, Fig. 3 shows the performance and comparison
of the results obtained from the hybrid model, using all uoride
breakthrough curves for ANNs training, in the data modeling of
adsorption columns for uoride removal using both bone chars.
Overall, the performance of the hybrid ANNs model is worse in
the zone [F]t/[F]0 < 0.1 of the uoride breakthrough curve while
this model provides satisfactory predictions for uoride adsorption
in the region 0.1 < [F]t/[F]0 < 1.0 (i.e., R2 > 0.9). It is convenient to
remark that the failures of Thomas model for representing ade-
quately the initial part of the breakthrough curve (i.e., t ? 0) has
been documented in other studies [49]. However, there is a signif-
icant improvement on the modeling capabilities of the ANNsTho-
mas model in comparison to results obtained with the traditional
Thomas model for this zone of the breakthrough curve. For further
studies, it is convenient to study the application of this hybrid
ANNs approach using other more exible breakthrough equations
such as the Yan model [49].
3.3. Analysis of design parameters of breakthrough curves for uoride
adsorption using bone chars
Design parameters of packed-bed columns were calculated
using the model parameters obtained from the Thomas equation
and the hybrid models and these results are reported in Table 5.
In particular, we have analyzed the effect of the modeling approach
on the calculation of design parameters of packed-bed columns.
Based on the fact that tested models showed the major uncertainty
for predicting the performance of uoride breakthrough curves at
low values of column operating time t, the breakthrough time tb
of packed columns has been determined from the experimental
data of uoride breakthrough curves and the exhaustion time te
was calculated from Thomas equation applying the condition
[F]/[F]0 = 0.95 in Eq. (5), where
qbed;T W bed ln19
te 11
Q F  0 kTh F  0
1106 R. Tovar-Gmez et al. / Chemical Engineering Journal 228 (2013) 10981109

Table 5
Calculated design parameters of packed-bed adsorption columns for the uoride removal from water using bone chars.

Bone char Flow rate (L/h) [F]0 (mg/L) Results of uoride breakthrough data modeling for
Thomas model ANNThomas model using all breakthrough ANNThomas model using individual
curves breakthrough curves
Dt (h) MTZ (cm) rf Fq Dt (h) MTZ (cm) rf Fq Dt (h) MTZ (cm) rf Fq
BCB 0.198 9 37.00 6.47 404.1 0.37 31.09 6.30 388.9 0.35 30.02 6.26 337.3 0.31
0.396 28.70 6.98 559.1 0.51 24.95 6.91 388.9 0.35 21.32 6.82 376.8 0.34
0.198 20 30.00 6.62 337.7 0.68 12.66 5.7 175.0 0.35 21.56 6.33 225.6 0.46
0.396 12.17 6.88 217.4 0.44 11.08 6.82 175.0 0.35 7.28 6.52 154.5 0.31
0.198 40 13.29 6.76 136.7 0.55 6.88 6.19 87.5 0.35 12.41 6.72 116.2 0.47
0.396 2.96 6.27 32.1 0.13 5.51 6.79 87.5 0.35 4.81 6.69 94.3 0.38
BCM 0.198 9 136.15 6.96 1090.8 0.80 59.92 6.38 650.2 0.47 95.33 6.76 790.0 0.58
0.396 83.51 7.31 938.9 0.69 51.58 7.19 650.2 0.47 80.04 7.30 756.6 0.55
0.198 20 92.19 7.18 683.6 1.11 27.58 6.53 292.6 0.47 24.31 6.42 273.5 0.44
0.396 26.11 7.09 447.1 0.73 22.71 7.04 292.6 0.47 21.50 7.01 335.0 0.54
0.198 40 39.63 7.00 334.3 1.08 13.00 6.15 146.3 0.47 11.69 6.03 137.4 0.45
0.396 8.83 6.95 154.5 0.50 11.41 7.07 146.3 0.47 5.73 6.69 113.5 0.37

Calculated values of te were used for the determination of MTZ,
Dt and rf for the different operating conditions of uoride adsorp-
tion columns using both bone chars (see Table 5). Note that Tho-
mas equation has been also used for the calculation of
retardation factors replacing [F]/[F]0 = 0.5 in Eq. (5) and rear-
ranging, then rf is given by
qbed;T qads
rf
F  0 e
The time (t50%) used for the calculation of retardation factors has
been also obtained from Thomas equation using
qbed W bed
t 50%
QF  0
For both adsorbents, the calculated values of MTZ increased
with the ow rate of uoride solution at given feed uoride con-
centration, while Dt decreased with an increment of both feed ow
rate and inlet uoride concentration. These trends have been ob-
served for all modeling approaches used in adsorption data analy-
sis. As stated in literature [40], the ideal breakthrough curve is a
step function where there should be an instantaneous increment
of the efuent pollutant concentration from zero to the feed pollu-
tant concentration when the maximum adsorption capacity of the
column is reached. Therefore, a short MTZ and a low value of Dt
(i.e., tbte ? 0) are desirable for optimizing the performance of dy-
namic adsorption process. On the other hand, the calculated values
of rf are higher for BCM than those obtained for BCB indicating that
uoride ions are adsorbed more effectively in the rst adsorbent,
see results given in Table 5. In fact, uoride ions will move faster
inside the adsorption column with lower rf than those packed-
bed columns with higher rf. As expected, the retardation factors
should decrease with the inlet uoride concentration. However,
it is convenient to note that calculated values of rf are constant,
for a given feed ow rate, using the hybrid model with the ANNs
training based on all breakthrough curves (see Table 5). This result
is because this ANNsThomas model provides unique and con-
stants values for parameters of Thomas equation and rf is only
dependent on the inlet pollutant concentration, see Eq. (12).
Finally, the estimated lter material usage ratios are also re-
ported in Table 5. In particular, this parameter is used to highlight
the capabilities and limitations of the models used for predicting
the performance of breakthrough curves. Fp has been calculated
using Eq. (4) while qmax and qbed,T obtained from Langmuir model
and Thomas equation, respectively. Calculated values of Fp using
the results of linear regression showed a high variability and incon-
sistent results. Based on the fact that diffusion phenomena are
affected by hydraulic conditions of adsorption columns and the
adsorption equilibrium is usually not reached in these separation
systems, it is expected that the adsorbent performance should be
reduced in xed-bed columns in comparison to the batch adsorp-
tion systems (i.e., low values of Fp are expected). However, if the re-
sults from linear regression of Thomas model are used, we
overpredicted the lter material usage rates and incorrect estima-
tions are obtained (i.e., Fp > 1 or qbed,T > qmax); see Table 5 especially
12
for BCM adsorbent. Therefore, the estimated values of the design
parameters may be not representative of the true adsorption pro-
cess and any decision based on these values would be in error, thus
potentially leading to ineffective process design decisions. In fact,
these results highlight some of the limitations of the use of linear-
13 ized Thomas equation for breakthrough data modeling and, conse-
quently, for adsorption process design. As stated in other studies,
e.g. [47], Thomas model is reliable to predict the performance of
breakthrough curves and to estimate the maximum pollutant up-
take by the adsorbent at dynamic operation conditions. However,
this study shows that the improper handling of this model may
lead to imprecise values of design parameters and erroneous
conclusions of adsorption performance. On the other hand,
ANNsThomas models provide reasonable values of Fp for the dif-
ferent operating conditions of uoride breakthrough curves and
these calculated ratios of lter material usage are consistent with
calculations performed in other studies, e.g. [47]. As stated, the
hybrid model trained using all breakthrough curves predicted con-
stant values of Fp for both bone chars, i.e., 0.35 for BCB and 0.47 for
BCM. In contrast, the values of Fp calculated from the other hybrid
model varied from 0.31 to 0.47 and from 0.37 to 0.58 for BCB and
BCM, respectively. In summary, the design parameters calculated
by linear regression and ANNsThomas models showed signicant
deviations. These deviations are caused by the characteristics of
the modeling approach used for adsorption data analysis. However,
ANNsThomas hybrid models provide reasonable and accurate
values of column design parameters.
3.4. Comparison of performance of commercial bone chars and the
uoride removal mechanism
Finally, the differences in the adsorption performance of tested
bone chars are mainly related to its chemical composition. Results
of elemental composition and the inorganic constituents identied
by EDX in both bone chars are reported in Table 6. In this case, it is
important to highlight the difference that exists in the percentage
of carbon, hydrogen and oxygen between these two adsorbents.
Specically, BCM showed the highest content of oxygen and hydro-
gen and, probably, this adsorbent may have a higher quantity of
 R. Tovar-Gmez et al. / Chemical Engineering Journal 228 (2013) 10981109 1107

Table 6 identied in BCM and BCB according to EDX analysis, see Table 6
Results of ultimate analysis and principals elements identied by EDX in the and Fig. 7. Also, the diffraction patterns of the crystalline form of
commercial bone chars used for the uoride removal from water. hydroxyapatite were observed in the X-ray results of both BCM
Ultimate analysis EDX resultsa and BCB (Fig. 8a). In general, the diffraction patterns of these
Element (w%) BCM BCB Element (w%) BCM BCB adsorbents are very similar to those reported for other bone chars
obtained from swine [53] and bovine bones [51]. On the other
C 6.1 7.3 Mg 1.26 1.155
H 0.5 0.1 P 16.64 15.48
hand, some characteristics peaks of hydroxyapatite were identi-
N 0.8 0.7 Ca 30.72 37.78 ed in the FT-IR spectrum of BCM and BCB at 1030 and
S 0.0 0.0 600 cm1, which can be assigned to vibrations of stretching and
O 10.3 6.9 bending of phosphate [54], see Fig. 8b. Additionally, other peaks
a
Average of three determinations. were identied and the signal around 3400 cm1 can be attribut-
able to the existence of hydroxyl groups and the peaks at
1460 cm1 can be assigned to CO stretching vibrations of car-
hydroxyl groups than that of BCB. These hydroxyl groups are in- boxyl groups [25]. Finally, the particles of BCM and BCB have
volved in the uoride removal according to the reaction [50] an irregular form and few pores are observed on the char surface
(see Fig. 7). This morphology is associated with the low specic
Ca10 PO4 6 OH2 2F  ! Ca10 PO4 6 F2 2OH 14 surface of these adsorbents, which was 104 and 129 m2/g for
where Ca10(PO4)6(OH)2 is the chemical structure of hydroxyapa- BCM and BCB, respectively. These characterization results are
tite, which is the principal component of bone chars according similar to those obtained for other commercial bone chars re-
to results reported in the literature [5052]. Our results ported in the literature such as the bone chars from Carvo Ati-
corroborated that Ca and P are the principal inorganic elements vado do Brasil and Barnebey Sutclife Corporation [25].

Fig. 7. SEM images and EDX results of the adsorbents: BCM and BCB.
1108 R. Tovar-Gmez et al. / Chemical Engineering Journal 228 (2013) 10981109
(a) % (b)
%: Hydroxylapatite
Lin (Counts) %%
% %
10 20 30 40 50 60
2-Theta - Scale
Fig. 8. (a) XRD patterns and (b) FT-IR spectra of BCM (black line) and BCB (gray line).
4. Conclusions
A hybrid neural network approach has been proposed for data
modeling of asymmetric experimental breakthrough curves of
the uoride removal from water using packed-bed adsorption col-
umns and bone chars. Specically, it is a gray-box model obtained
from the combination of articial neural networks and Thomas
breakthrough equation. Capabilities and limitations of this hybrid
model has been compared and analyzed with respect to the perfor-
mance of traditional linear regression of Thomas model. Results
showed that the breakthrough data modeling using the linear
regression of Thomas model may overestimate design parameters
of adsorption columns causing erroneous conclusions for design
and operation policy of adsorption processes. On the other, the hy-
brid model is useful for tting and predicting satisfactorily the
uoride breakthrough curves. This hybrid model can predict prop-
erly the uoride breakthrough curves at different operating condi-
tions using a unique set of values of parameters kTh and qbed,T of
Thomas equation. In summary, the proposed modeling approach
can be useful for process engineering of dynamic adsorption sys-
tems involved in the water and wastewater treatment.
Acknowledgements
CONACYT, DGEST, Gobierno del Estado de Aguascalientes and
Municipio de Aguascalientes, Instituto Tecnolgico de Aguascalien-
tes and Instituto Nacional del Carbn have provided the nancial
support for this study.
References
[1] Meenakshi, R.C. Maheshwari, Fluoride in drinking water and its removal, J.
Hazard. Mater. B137 (2006) pp. 456463.
[2] M.T. Alarcn-Herrera, J. Bundschuh, B. Nath, H.B. Nicolli, M. Gutierrez, V.M.
Reyes-Gomez, D. Nuez, I.R. Martn-Dominguez, O. Sracek, Co-ocurrence of
arsenic and uoride in groundwater of semi-arid regions in Latin America:
Genesis, mobility and remediation, J. Hazard. Mater, 2012 (in press).
[3] Guidelines for drinking-water quality, World Health Organization, 4th ed.,
2011.
[4] S.S. Tripathy, J.L. Bersillon, K. Gopal, Removal of uoride from drinking water
by adsorption onto alum-impregnated activated alumina, Sep. Purif. Technol.
50 (2006) 310317.
[5] A. Bhatnagar, E. Kumar, M. Sillanpaa, Fluoride removal form water by
adsorption a review, Chem. Eng. J. 171 (2011) 811840.
BCM
BCB 3420
1480
Transmittance (%)
604
1040
3390
1460
604
1040
70 80 4000 3000 2000 1000
Wavenumber (cm-1)
[6] M. Mohapatra, S. Anand, B.K. Mishra, D.E. Giles, P. Singh, Review of uoride
removal from drinking water, J. Environ. Manage. 91 (2009) 6777.
[7] I. Abe, S. Iwasaki, T. Tokimoto, N. Kawasaki, T. Nakamura, S. Tanada, Adsorption
of uoride ions onto carbonaceous materials, J. Colloid Interface Sci. 275 (2004)
3539.
[8] A.A.M. Daifullah, S.M. Yakout, S.A. Elreefy, Adsorption of uoride in aqueous
solutions using KMnO4-modied activated carbon derived from steam
pyrolysis of rice straw, J. Hazard. Mater. 147 (2007) 633643.
[9] V. Hernndez-Montoya, L.A. Ramrez-Montoya, A. Bonilla-Petriciolet, M.A.
Montes-Moran, Optimizing the removal of uoride from water using new
carbons obtained by modication of nut shell with a calcium solution from egg
shell, Biochem. Eng. J. 62 (2012) 17.
[10] J.L. Dvila-Rodrguez, V.A. Escobar-Barrios, J.R. Rangel-Mndez, Removal of
uoride from drinking water by a chitin-based biocomposite in xed-bed
columns, J. Fluorine Chem. 140 (2012) 99103.
[11] M. Srimurali, A. Pragathi, J. Karthikeyan, A study on removal of uorides from
drinking water by adsorption onto low-cost materials, Environ. Pollut. 99
(1998) 285289.
[12] W. Nigussie, F. Zewge, B.S. Chandravanshi, Removal of excess uoride water
using waste residue from alum manufacturing process, J. Hazard. Mater. 147
(2007) 954963.
[13] V. Sivasankar, T. Ramachandramoorthy, A. Chandramohan, Fluoride removal
from water using activated and MnO2-coated Tamarind fruit (Tamarindus
indica) shell: batch and column studies, J. Hazard. Mater. 177 (2010) 719
729.
[14] H. Paudyal, B. Pangeni, K. Inoue, H. Kawakita, K. Ohto, H. Harada, S. Alam,
Adsorptive removal of uoride from aqueous solution using orange waste
loaded with multi-valent metal ions, J. Hazard. Mater. 192 (2011) 676
682.
[15] A.B. Nasr, K. Walha, C. Charcosset, R.B. Amar, Removal of uoride ions using
cuttlesh bones, J. Fluorine Chem. 132 (2011) 5762.
[16] A.M. Raichur, M.J. Basu, Adsorption of uoride onto mixed rare earth oxides,
Sep. Purif. Technol. 24 (2001) 121127.
[17] X. Fan, D.J. Parker, M.D. Smith, Adsorption kinetics of uoride on low cost
materials, Water Res. 37 (2003) 49294937.
[18] A. Tor, N. Danaoglu, G. Arslan, Y. Cengeloglu, Removal of uoride from water
by using granular red mud: batch and column studies, J. Hazard. Mater. 164
(2009) 271278.
[19] N. Chen, Z. Zhang, C. Feng, N. Sugiura, M. Li, R. Chen, Fluoride removal from
water by granular ceramic adsorption, J. Colloid Interface Sci. 348 (2010) 579
584.
[20] R. Liu, W. Gong, H. Lan, T. Yang, H. Liu, J. Qu, Simultaneous removal of arsenate
and uoride by iron and aluminum binary oxide: competitive adsorption
effects, Sep. Purif. Technol. 92 (2012) 100105.
[21] M. Mourabet, A. El Rhilassi, H. El Boujaady, M. Bennani-Ziatni, R. El Hamri, A.
Taitai, Removal of uoride from aqueous solution by adsorption on apatitic
tricalcium phosphate using Box-Behnken design and desirability function,
Appl. Surf. Sci. 258 (2012) 44024410.
[22] G. Alagumuthu, M. Rajan, Equilibrium and kinetics of adsorption of uoride
onto zirconium impregnated cashew nut shell carbon, Chem. Eng. J. 158 (2010)
451457.
[23] E. Tchomgui-Kamga, V. Alonzo, C.P. Nanseu-Njiki, N. Audebrand, E. Ngameni,
A. Darchen, Preparation and characterization of charcoals that contain
dispersed aluminum oxide as adsorbents for removal of uoride from
drinking water, Carbon 48 (2010) 333343.
 R. Tovar-Gmez et al. / Chemical Engineering Journal 228 (2013) 10981109
[24] S. Ghorai, K.K. Pant, Equilibrium, kinetics and breakthrough studies for
adsorption of uoride on activated alumina, Sep. Purif. Technol. 42 (2005)
265271.
[25] V. Hernndez-Montoya, M.P. Elizalde-Gonzlez, R. Trejo-Vzquez, Screening of
commercial sorbents for removal of uoride in synthetic and groundwater,
Environ. Technol. 28 (2007) 595607.
[26] N.A. Medellin-Castillo, R. Leyva-Ramos, R. Ocampo-Perez, R.F. Garcia de la
Cruz, A. Aragon-Pia, J.M. Martnez-Rosales, R.M. Guerrero-Coronado, L.
Fuentes-Rubio, Adsorption of uoride from water solution on bone char, Ind.
Eng. Chem. Res. 46 (2007) 92059212.
[27] R. Leyva-Ramos, J. Rivera-Utrilla, N.A. Medellin-Castillo, M. Sanchez-Polo,
Kinetic modeling of uoride adsorption from aqueous solution onto bone char,
Chem. Eng. J. 158 (2010) 458467.
[28] D.S. Bhargava, D.J. Killedar, Fluoride adsorption on shbone charcoal through a
moving media adsorber, Water Res. 26 (1992) 781788.
[29] S. Ayoob, A.K. Gupta, Sorptive response prole of an adsorbent in the
deuoridation of drinking water, Chem. Eng. J. 133 (2007) 273281.
[30] M. Sarkar, A. Banerjee, P.P. Pramanick, A.R. Sarkar, Design and operation of
xed bed laterite column for the removal of uoride from water, Chem. Eng. J.
131 (2007) 329335.
[31] N. Chen, Z. Zhang, C. Feng, M. Li, R. Chen, N. Sugiura, Investigations on the
batch and xed-bed column performance of uoride adsorption by Kanuma
mud, Desalination 268 (2011) 7682.
[32] S. Ghorai, K.K. Pant, Investigations on the column performance of uoride
adsorption by activated alumina in a xed-bed, Chem. Eng. J. 98 (2004) 165
173.
[33] E. Oguz, M. Ersoy, Removal of Cu2+ from aqueous solution by adsorption in a
xed bed column and neural network modelling, Chem. Eng. J. 164 (2010) 56
62.
[34] G. Bohart, E.Q. Adams, Some aspects of the behavior of charcoal with respect to
chlorine, J. Am. Chem. Soc. 42 (1920) 523544.
[35] H.C. Thomas, Heterogeneous ion exchange in a owing system, J. Am. Chem.
Soc. 66 (1944) 16641666.
[36] Y.H. Yoon, J.H. Nelson, Application of gas adsorption kinetics, Am. Ind. Hyg.
Assoc. J. 45 (1984) 509516.
[37] R.M. Clark, Evaluating the cost and performance of eld-scale granular
activated carbon systems, Environ. Sci. Technol. 21 (1987) 573580.
[38] K.H. Chu, Improved xed bed models for metal biosorption, Chem. Eng. J. 97
(2004) 233239.
[39] C. Faur, A. Cogunaud, G. Dreyfus, P.L. Cloirec, Modelling the breakthrough of
activated carbon lters by pesticides in surface waters with static and
recurrent neural networks, Chem. Eng. J. 145 (2008) 715.
The author has requested enhancement of the downloaded file. All in-text references underlined in blue are linked to publications on ResearchGate.
1109
[40] Z. Aksu, F. Gnen, Binary biosorption of phenol and chromium (VI) onto
immobilized activated sludge in a packed bed: prediction of kinetic
parameters and breakthrough curves, Sep. Purif. Technol. 29 (2006) 205216.
[41] I.A. Basheer, M. Hajmeer, Articial neural networks: fundamentals, computing,
design and application, J. Microbiol. Meth. 43 (2000) 331.
[42] A.C. Texier, Y. Andrs, C. Faur-Brasquet, P.L. Cloirec, Fixed-bed study for
lanthanide (La, Eu, Yb) ions removal from aqueous solutions by immobilized
Pseudomonas aeruginosa: experimental data and modelization, Chemosphere
47 (2002) 333342.
[43] L. Cavas, Z. Karabay, H. Alyuruk, H. Dogan, G.K. Demir, Thomas and articial
neural network models for the xed-bed adsorption of methylene blue by a
beach waste Posidonia oceanica (L.) dead leaves, Chem. Eng. J. 171 (2011) 557
562.
[44] U.I. Bravo-Snchez, R. Rico-Martnez, G.A. Iglesias-Silva, Improvement of the
empiricism in the Back equation of state via hybrid neural networks, Ind. Eng.
Chem. Res. 41 (2002) 37053713.
[45] J.A. Dena-Aguilar, Master Thesis, Instituto Tecnolgico de Aguascalientes,
Mxico, 2007.
[46] R. Apiratikul, P. Pavasant, Batch and column studies of biosorption of heavy
metals by Caulerpa lentillifera, Bioresour. Technol. 99 (2008) 27662777.
[47] I.A. Aguayo-Villarreal, A. Bonilla-Petriciolet, V. Hernndez-Montoya, M.A.
Montes-Morn, H.E. Reynel-Avila, Batch and column studies of Zn2+ removal
from aqueous solution using chicken feathers as sorbents, Chem. Eng. J. 167
(2011) 6776.
[48] Standard Methods for Examination of Water and Wastewater, American Public
Health Association, 20th ed., 1998.
[49] D. Pokhrel, T. Viraraghavan, Arsenic removal in an iron oxide-coated fungal
biomass column: analysis of breakthrough studies, Bioresour. Technol. 99
(2008) 20672071.
[50] S.M. Rao, B.V.V. Reddy, S. Lakshmikanth, N.S. Ambika, Re-use of uoride
contaminated bone char sludge in concrete, J. Hazard. Mater. 166 (2009) 751
756.
[51] K.K.H. Choy, G. McKay, Sorption of cadmium, copper, and zinc ions onto bone
char using Crank diffusion model, Chemosphere 60 (2005) 11411150.
[52] M.C. Bennett, J.C. Abram, Adsorption from Solution on the carbon and
hydroxyapatite components of bone char, J. Colloid Interface Sci. 23 (1957)
513521.
[53] X. Pan, J. Wang, D. Zhang, Sorption of cobalt to bone char: kinetics, competitive
sorption and mechanism, Desalination 249 (2009) 609614.
[54] S. Dimovic, I. Smiciklas, I. Pleacas, D. Antonovic, M. Mitric, Comparative study
of differently treated animal bones for Co2+ removal, J. Hazard. Mater. 164
(2009) 279287.