Journal of Molecular Liquids 208 (2015) 114121

Contents lists available at ScienceDirect

Journal of Molecular Liquids

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

Breakthrough curve modeling of liquid-phase adsorption of uoride ions

on aluminum-doped bone char using micro-columns: Effectiveness of
data tting approaches
C.K. Rojas-Mayorga a, A. Bonilla-Petriciolet a,, F.J. Snchez-Ruiz a, J. Moreno-Prez a, H.E. Reynel-vila a,
I.A. Aguayo-Villarreal a, D.I. Mendoza-Castillo a,b
a
b

Instituto Tecnolgico de Aguascalientes, Aguascalientes 20256, Mexico

Consejo Nacional de Ciencia y Tecnologa, Ctedras CONACYT, Mxico D.F. 03940, Mexico

a r t i c l e

i n f o

Available online 28 April 2015

Keywords:
Packed-bed columns
Fluoride
Aluminum-modied bone char
Breakthrough curve modeling
Liquid-phase adsorption

a b s t r a c t
A comparative numerical analysis of three approaches for the breakthrough curve modeling has been performed
using the experimental data of packed-bed uoride adsorption on a novel aluminum-doped bone char using
micro-columns. The performance of traditional Thomas and Yan breakthrough equations, a mass transfer
model for a mobile uid owing through a porous media, and an articial neural network with the optimal
brain surgeon approach have been studied and discussed in the data tting of asymmetric uoride breakthrough
curves. Results of this study highlighted the relative merits of tested breakthrough curve models for the nonlinear adsorption data analysis involved in water deuoridation using a new adsorbent. In particular, the application of articial neural networks is reliable for tting highly non-linear adsorption patterns of priority water
pollutants.
2015 Elsevier B.V. All rights reserved.

1. Introduction
Adsorption processes can be performed using batch and continuous
type congurations [1,2]. Both operation modes offer different advantages but the dynamic adsorption systems are preferred for large-scale
applications in water treatment due to their exibility and easy operation [1,3]. The liquid-phase adsorption at continuous ow conditions
is recognized as a non-linear and multivariable process that can be analyzed by the breakthrough curve. Specically, the breakthrough curve
describes the adsorption column dynamics and provides relevant information and key parameters for the design, operation and optimization
of the separation system [46]. For example, the sharpness of the breakthrough curves is an indicator of the effectiveness of the adsorption process [4]. The characteristics of breakthrough curves depend on the bed
conguration, feed ow, adsorbate concentration, pH, temperature, adsorbent mass, and other relevant variables [1,58]. Therefore, it is expected that the adsorption pattern may vary signicantly for each
adsorbentadsorbate system in packed-bed columns.
Results reported in the literature indicate that the breakthrough
curves obtained in the adsorption of water pollutants usually show an
asymmetric sigmoidal prole even for mono-component solutions
(i.e., solutions with only one adsorbate/pollutant). The modeling of
these breakthrough curves for liquid-phase adsorption is fundamental
Corresponding author.
E-mail address: petriciolet@hotmail.com (A. Bonilla-Petriciolet).

http://dx.doi.org/10.1016/j.molliq.2015.04.045
0167-7322/ 2015 Elsevier B.V. All rights reserved.

for water treatment design in real-life applications [8]. The principal objective of the breakthrough data modeling stage is the correlation and
prediction of adsorption behavior based on the parameters obtained
from experimental studies. However, the tting of adsorption patterns
in packed-bed columns is more difcult than the modeling in batch processes because the concentration proles in the adsorbent and liquid
phases for dynamic systems are determined by the mass transfer phenomena, adsorption kinetics and thermodynamics [4]. The breakthrough curve modeling is considered as a challenging mathematical
problem, which implies non-linear data regressions for the determination of model parameters using an error function. Depending on the
model complexity, the application of other computational strategies is
often necessary in breakthrough curve tting, e.g., the use of numerical
methods for solving partial differential equations.
To date, several models have been proposed for the correlation of
breakthrough curve data obtained in dynamic adsorption processes.
These models include the Thomas [9], Bohart-Adams [10], Yon-Nelson
[11], Clark [12], Yan [13] and other empirical or semi-empirical breakthrough equations [2,4]. These models are usually analytical equations
that require a low computational time for data processing and are popular for the correlation of adsorption data in water treatment due to
their simple form and acceptable-tting performance. Alternatively, a
variety of mass transfer models can be applied for data tting of adsorption results at dynamic conditions [2,14]. This type of models shows different degrees of mathematical complexity depending on the transport
phenomena considered for the adsorption analysis [2]. Besides the

C.K. Rojas-Mayorga et al. / Journal of Molecular Liquids 208 (2015) 114121

analytical and mass transfer models, the breakthrough curves can be

also modeled using computationally intelligent data processing algorithms such as articial neural networks [68,15,16]. These strategies
are considered as effective empirical approaches for non-linear data
analysis [17].
Herein, it is convenient to highlight that the desirable characteristics
for the breakthrough curve models include a suitable mathematical complexity, a capability for providing an acceptable accuracy in the correlation
and estimation of adsorption patterns at different operating conditions,
and the model should be useful for assessing the effect of operating variables on the adsorption column dynamics [2]. Depending on the type of
breakthrough model, some of these desirable characteristics can be fullled. However, the numerical performance of available models cannot
be determined a priori based on the fact that they may show limitations
for data analysis depending on the adsorption system under study and
its experimental conditions. In fact, some studies have reported the performance of different approaches for breakthrough curve modeling but
using concentration patterns with symmetric behavior [18]. In order to
improve the performance of breakthrough models and to develop reliable
approaches for adsorption data processing, it is important to identify and
to compare the relative advantages of available models in the analysis of
dynamic processes for the removal of priority water pollutants involving
highly asymmetric adsorption proles.
In this study, a comparative numerical analysis of three approaches
for breakthrough curve modeling has been performed using the experimental data of packed-bed uoride adsorption on a novel aluminumdoped bone char. Specically, the performance of traditional Thomas
and Yan breakthrough equations, the mass transfer model for a mobile
uid owing through a porous media, and an articial neural network
approach have been studied and discussed in the data tting of asymmetric uoride breakthrough curves obtained with micro-columns. To
the best of the authors' knowledge, a systematic study of the capabilities
and limitations of breakthrough models for the correlation and prediction of adsorption patterns in packed beds for water deuoridation
has not been reported. In summary, the results of this study highlight
the relative merits of tested breakthrough curve modeling strategies
for the adsorption data analysis involved in water deuoridation using
a new bone char.
2. Methodology

115

second stage, the surface modication of bone char was performed

with an aluminum sulfate solution 0.1 M. The adsorbent was mixed
with this metallic solution at 50 C during 6 h. The samples of
aluminum-doped bone char were washed using deionized water,
dried and stored for dynamic adsorption experiments. The physicochemical properties of the modied adsorbent were determined using
the results from FTIR, XPS and XRD analyses. Diffraction pattern of
bone char was obtained using an X-ray diffractometer Bruker D8Advance with a mirror Gebel that has a tube with cooper anode RX
and radiation Cu K ( = 1.5406 ). Functional groups of the adsorbent
were identied via FTIR spectroscopy using a Bruker IFS 66/S spectrophotometer and all samples were analyzed using spectroscopic grade
KBr. XPS studies were performed with a Prevac photoelectron spectrometer equipped with a hemispherical analyzer (VG SCIENTA
R3000). Finally, the textural properties of metal-modied bone char
were determined via N2 adsorptiondesorption isotherm at 77 K using
a N2 Gsorb-6 equipment.
2.2. Determination of breakthrough curves for uoride adsorption on
aluminum-doped bone char using packed-bed micro-columns
Fluoride adsorption experiments were performed in packed-bed
micro-columns with 5.4 cm of length and 1.8 cm of internal diameter.
These columns were packed with 10.35 g of metal-doped bone
char with a particle size of 2035 mesh fractions. Dynamic adsorption experiments were performed at pH 7 and 30 C using different uoride feed
concentrations from 10 to 100 mg/L and feed ow rates of 0.18 and 0.36 L/
h. All adsorption micro-columns were operated at up ow operating
mode using a peristaltic pump. The mean bed porosity () in all adsorption columns was 30%. For the determination of experimental uoride
breakthrough curves, several samples were collected from the adsorption
column outlet at regular time intervals (i.e., 10 min) and the uoride concentrations were quantied. Adsorbate concentrations were determined
using a uoride ion selective electrode and TISAB chemical reagent [21]
where linear calibration curves were used. All experiments were performed twice, including the uoride measurements, and the mean results
were used for adsorption calculations.
Analysis of uoride breakthrough curves implied the determination
of the breakthrough time (tb in h), the bed adsorption capacity (qbed
in mg/g), the retardation factor (rf) and the degree of bed utilization
(Rq in %). These parameters were calculated using

2.1. Synthesis of aluminum-doped bone char used in uoride adsorption

with packed-bed columns

t b t oper where 0:2F #feed

Packed-bed adsorption experiments for uoride removal from water

were performed using an aluminum-doped bone char. This novel adsorbent was obtained via metal surface modication of bone char using an
aqueous solution of aluminum sulfate. First, the bone char was obtained
from the pyrolysis of bovine bones using the optimum synthesis conditions reported by Rojas-Mayorga et al. [19]. These conditions involved a
pyrolysis process with N2 at 700 C during 2 h of residence time. In a

qbed t0oper

tt

!
"
F #feed F #outlet
Q dt
mbed

1
2

rf

V 50%
AL

Rq

qbed
qbatch

1. Train a reasonably large network to minimum error

2. Compute H -1
3. Find the q that gives the smallest saliency, Eq. (16). If this candidate error increase is lower than
the error function used in training process, then the qth weight should be deleted, and continue
with step 4 of this algorithm; otherwise step 5 applies.
4. Use the q from step 3 to update all weights using Eq. (15). Go to step 2.
5. No more weights can be deleted without large increase in the error function of training process.
At this point it may be desirable to retrain the network.
Fig. 1. Algorithm reported by Hassibi et al. [22 for the optimal brain surgeon procedure in articial neural network models.

116

C.K. Rojas-Mayorga et al. / Journal of Molecular Liquids 208 (2015) 114121

a)

where [F]feed and [F]outlet are the uoride concentrations of the feed
and outlet of adsorption columns given in mg/L, Q is the feed ow in
L/h, mbed is the mass of bone char packed in the adsorption columns
given in g, t is the operating time of adsorption columns in h, is the column void fraction of the packed-bed column (~0.3), L is the bed height
of the adsorption column given in cm, A is the cross sectional area of the
adsorption column reported in cm2, V50% is the treated volume when
the efuent concentration has reached 50% of the inuent concentration
and it is given in cm3, and qbatch is the uoride adsorption capacity
of bone char obtained at batch mode using the same operating conditions (i.e., pH and temperature) of dynamic experiments and it is
given in mg/g, respectively. All parameters were determined for tested
conditions of xed-bed adsorption studies. Note that the integration of
Eq. (2) for the calculation of the bed adsorption capacity was performed
using the trapezoidal rule.

* Hydroxyapatite

Intensity (a.u.)

*
*
*

*
*

* *
*

**

BC-AlS
10

20

30

40

50

60

70

80

b)

In this study, three approaches have been used for data tting of
uoride breakthrough curves on the metal-modied bone char. These
approaches include the application of: a) analytical breakthrough equations, b) a mass transfer model and c) an articial neural network
model. In the following subsections, a brief description of each modeling
approach is provided.

Transmittance (%)

2 ()

2.3. Modeling of the uoride xed-bed adsorption on aluminum-doped

bone char

++

3309

873

1619 1450

+
+

599

559

2.3.1. Analytical breakthrough equations: Thomas and Yan models

Two analytical breakthrough curve equations were employed for the
data modeling of dynamic uoride adsorption on the modied bone
char. These models include the Thomas [9] and Yan [13] equations,
which are dened as

1018

BC-AlS
4000 3500 3000 2500 2000 1500 1000 500
-1

F #outlet

F #feed

!
$"
k #

1 exp Th qbed;T mbed F #feed V eff

Q

F #outlet
1
F #feed

Q2t
ky qbed;Y mbed

$k F #
Y
feed=Q

Wavenumber (cm )
5

where Veff is the treated volume of the uoride solution in L, qbed,T is the
adsorption capacity estimated by the Thomas model given in mg/g, kTh
is the Thomas rate constant reported in L/h mg, kY is the Yan rate constant given in L/h mg, and qbed,Y is the adsorption capacity estimated
by the Yan model given in mg/g, respectively. Parameters of these analytical models (kTh, qbed,T) and (kY, qbed,Y) were obtained from a nonlinear regression of the uoride breakthrough curves using a global optimization method.
2.3.2. Mass transfer model for a porous medium
A general liquid phase mass transfer model for a porous medium
was used in data tting of the uoride breakthrough curves. This
model is based on a macroscopic mass conservation equation applied

Fig. 3. Results of a) XRD and b) FTIR analyses of aluminum-modied bone char used for
uoride adsorption in packed-bed micro-columns.

for a control volume and it is given by [2]

2
#
$ c
c
c
b K p
u DDap 2
t
x
x

where Kp is an adsorption constant given in m3/Kg, DDap is an

apparent dispersion coefcient in m2/h, x is the distance along the bed
length in m, b is the bulk density in Kg/m3, t is the operation time in
h, c = [F] is the adsorbate concentration in the uid phase given in
mol/m3, and u is the uid velocity in m/h, respectively. The model parameters Kp and DDap were also determined via the non-linear data
tting of experimental uoride breakthrough curves. The initial and

Table 1
Functional groups and XPS surface composition of aluminum-doped bone char used for
uoride adsorption in packed-bed micro-columns.
FTIR interpretation

Fig. 2. Architecture of articial neural network (ANN) model for the correlation of uoride
breakthrough curves on aluminum-doped bone char.

XPS surface chemical

composition

Wavenumber, cm1

Functional groups

Element

Wt, %

3309
1619
1450
1018
873
599
559

\
\OH
C_O
C_C
PO3
4
CO2
3
PO3
4
Ca

C
O
P
Ca
Al

14.91
38.12
19.68
19.49
7.80

117

C.K. Rojas-Mayorga et al. / Journal of Molecular Liquids 208 (2015) 114121

Table 2
Parameters of breakthrough curves for the uoride adsorption on aluminum-doped bone
char in packed-bed micro-columns.
Feed conditions

Design parameters of breakthrough curves

[F]feed, mg/L

Q, L/h

tb at 0.2[F]feed, h

rf

qbed, mg/g

Rq, %

10

0.18
0.36
0.18
0.36
0.18
0.36
0.18
0.36

9.50
4.83
8.67
4.33
8.00
4.17
5.67
3.33

702.3
704.1
647.3
593.6
464.2
493.1
319.4
420.4

3.3
4.2
7.7
8.4
13.5
11.7
14.2
18.5

10.7
13.5
24.9
27.2
43.6
37.6
45.9
59.7

25
50
100

boundary conditions used for solving this mass transfer model are
t 0; 0 b x b L; c 0
t N0; x 0; c cf
t N0;

x L;

c0

where cf = [F]outlet is the nal uoride concentration in the column

outlet obtained from the experimental breakthrough curves. Comsol
Multiphysics was used for solving this model and the best values of
Kp and DDap were determined one-at-a-time via a parametric study
using this software.
2.3.3. Articial neural network model for uoride breakthrough curves
An articial neural network (ANN) model with the optimal
brain surgeon procedure [22] was employed for the correlation of
uoride breakthrough curves. In general, a basic neural network
model has nI input units, one hidden layer with nH units, and a single linear output unit. The response yp of the network to an input vector s can
be written as

yp

nH
X

W jh

j1

nI
X
i1

1.0

wji si bj

9
1.0

0.18 L/h
100 mg/L
50 mg/L
25 mg/L
10 mg/L

0.8
0.6
0.4

[F -]outlet/[ F - ]feed

0.2
0.0
1.0

12

16

20

N #
$2
1 X
yyp
i
2N i1

10

where N is the number of data used in the ANN training. To avoid the
over-tting and to guarantee the generalization ability of ANN model,
a validation set is used to compute the objective function. This set consists of data that are not used in the training set.
On the other hand, the optimal brain surgeon procedure is used
to estimate the increment in the training error when some weights
are deleted using the information in the second-order derivatives of
the error surface. This approach can be employed to minimize the
ANN complexity via the reduction of the number of connection weights.
It is also useful to prevent the overtting problems without a signicant
impact on the ANN model generalization. In this approach, a local model
of the error surface is constructed to evaluate the effect of weight
deletion [22]
F

"
#
$
F T
1 T
3
w w Hw O kwk
2
w

eq w wq 0

12
1.0

0.18 L/h

0.8

0.8

0.6

0.6

0.4

0.4

24

0.2

0.0

1.0

0.36 L/h

12

16

20

24

0.0

0.36 L/h

0.8

0.6

0.6

0.6

0.4

0.4

0.4

0.2

0.2

0.2

12

16

20

Artificial neural
networks (ANNs)

24

0.0

1.0

0.8

0.18 L/h

Mass transfer model

0.8

0.0

11

where H2F/w2 is the Hessian matrix and the superscript T denotes

the vector transpose. For a network trained to a local minimum in the
error function, the rst (linear) term vanishes and the third and all
higher order terms are also discarded. Then, a set of the weights (wp)
can be discarded (i.e., their values can be set to zero) for minimizing
the increment in the error given by Eq. (11). Note that

0.2

Thomas model
Yan model

where wji is the weight for the link from input i to the hidden cell j, Wj is
the weight for the link from hidden cell j to the output, h is a tangent hyperbolic or linear transfer function, bj and B are the biases of the model,
respectively. The weights and biases are the ANN adjustable parameters
and they are determined from the data under analysis via the training
process. The objective of this training procedure is to nd a set of possible weights and biases so that the ANN model will provide predictions
(yp) as similar as possible to the known outputs (y). The objective function used for this training process is given by

12

16

20

24

0.0

12

16

20

24

12

16

20

24

0.36 L/h

Time, h
Fig. 4. Breakthrough curves of uoride adsorption on aluminum-doped bone char in packed-bed micro-columns and results of data modeling approaches.

118

C.K. Rojas-Mayorga et al. / Journal of Molecular Liquids 208 (2015) 114121

where eq is the unit vector in the weight space corresponding to (scalar)

weight wp. Therefore, the problem to be solved is given by
minq

&

'
!
"%
1 T
% T
w Hw %eq w wq 0 :
minw
2

13

A Lagrangian formulation is considered for solving this optimization

problem
L

#
$
1 T
T
w Hw eq w wq
2

Table 3
Results of the Thomas and Yan models for the data tting of uoride breakthrough curves.

Model
Thomas

Feed conditions

Model parameters

Results of
data tting

[F]feed,
mg/L

Q, L/h

kTh or kY,
L/mg h

qbed,T or
qbed,Y, mg/g

R2

E, %

0.18
0.36
0.18
0.36
0.18
0.36
0.18
0.36
0.18
0.36
0.18
0.36
0.18
0.36
0.18
0.36

3.13E02
4.19E02
7.63E03
1.73E02
8.39E03
1.29E02
7.75E03
8.42E03
4.53E02
5.37E02
1.18E02
2.69E02
8.39E03
1.75E02
7.40E03
9.51E03

3.09
3.21
8.02
6.94
4.74
11.71
13.89
19.02
1.18
2.14
4.56
3.50
4.74
4.58
3.31
6.92

0.54
0.72
0.85
0.86
0.89
0.96
0.96
0.99
0.95
0.97
0.97
0.96
0.97
0.92
0.99
0.90

34.94
21.94
15.56
13.70
18.70
10.50
19.34
7.94
23.30
11.28
17.87
12.91
21.33
23.28
22.30
26.75

10
25
50

14

where is an undetermined Lagrange multiplier. According to Hassibi

et al. [22], the optimal weight change and the change in error are
given by

100
Yan

10
25
50

wq
1
w ( 1 ) H eq
H
qq
Lq

2
1 wq
( 1 )
2 H
qq

100

15

3.2. Analysis of uoride breakthrough curves

16

where neither H nor H1 need to be diagonal. In this method, the magnitude of all the weights in the ANN model is recalculated using Eq. (15).
Note that Lq is dened as the saliency of weight q, which represents
the increment in the error that results when the weight has been
discarded in the ANN model. For illustration, the general procedure of
the optimal brain surgeon approach is reported in Fig. 1 and interested
readers can found a detailed description of this strategy in the paper of
Hassibi et al. [22].
Finally, Fig. 2 shows the architecture of ANN model used for uoride
breakthrough curve tting. The input variables of this model were the
operating time, the feed uoride concentration and the feed ow,
while the output variable was the ratio of uoride concentrations in
the outlet and inlet of packed-bed micro-column. Two neurons in the
hidden layer were used for data modeling. It is important to remark
that different numbers of hidden neurons were analyzed (i.e., from 2
to 5 neurons) and the best conguration without overtting was selected for the ANN model. All uoride breakthrough curves were used for
the ANN modeling and they included 948 experimental data (50% training, 25% validation and 25% testing).
3. Results and discussion
3.1. Physicochemical characteristics of the adsorbent used in uoride
removal
Fig. 3 and Table 1 show the results of the analyses performed
for the adsorbent characterization. XRD analysis of metal-doped
bone char conrmed the crystalline structure of hydroxyapatite
[Ca10(PO4)6(OH)2] and the FTIR spectrum showed the absorption
bands that correspond to the structural composition of bone char [19,
20,23,24], see Fig. 3. In particular, the vibration bands correspond to
the inorganic and organic phases of bone char structure and they are described in Table 1. The bone char is a mixed adsorbent constituted by
carbon (~ 10%) and calcium phosphate (~ 70%) in the hydroxyapatite
form [19]. According to XPS results, the surface chemical composition
shows the typical composition of bone chars besides the presence of
aluminum, which is caused by the impregnation process of the adsorbent using aluminum sulfate. Note that this impregnation process is
useful to improve the uoride adsorption properties of bone char. This
bone char is a mesoporous adsorbent with a mean pore size of ~8 nm
and a BET surface area of ~85 m2/g.

Fig. 3 shows the experimental uoride breakthrough curves using

the metal-doped bone char in packed-bed micro-columns at tested operating conditions. All curves showed the typical S-concentration prole
but they are clearly asymmetric sigmoidal functions. The operating conditions have a signicant impact on the asymmetry degree of uoride
adsorption patterns, which is more evident at low feed ow and uoride
concentration; see results reported in Fig. 3. Overall, the breakthrough
curves shifted towards the origin with increasing Q or [F]feed. Note
that these parameters are related to the driving forces for adsorption
process and, consequently, they determine the adsorbent performance
in continuous mode operation. In particular, the gradient concentration
affects the adsorption rates and the mass transfer phenomena, while the
feed ow is related to column hydrodynamics and determines the contact time between the adsorbent and uoride solution (i.e., the residence time of uid inside the column).
Table 2 contains the calculated values of breakthrough times, retardation factors, bed adsorption capacities and bed utilization percentages
for all dynamic adsorption experiments in the micro-columns. It is clear
that these characteristics of uoride breakthrough curves are determined by both Q and [F]feed. In particular, tb values ranged from 5.67
to 9.50 h and from 3.33 to 4.83 h for Q = 0.18 and 0.36 L/h, respectively.
These breakthrough times decreased with both feed ow and column
inlet concentration. The location of breakthrough point depends on
the adsorbatesolute system and its values are relevant for the application of packed-bed columns in water treatment where high tb values are
desirable. On the other hand, the retardation factors (rf) are determined
by the characteristics of the column feed and they ranged from 319.4 to
704.1, see Table 2. As expected, rf decreased with feed uoride

Table 4
Results of mass transfer model for data tting of uoride breakthrough curves.
Operating conditions

Mass transfer parameters

Results of data
tting

Q, L/h

[F]feed, mg/L

DDap, cm2/s

Kp, L/g

R2

E, %

0.18

10
25
50
100
10
25
50
100

0.031
0.056
0.044
0.012
0.115
0.090
0.046
0.035

0.78
0.77
0.76
0.43
1.1
0.90
0.70
0.59

0.99
0.98
0.99
0.99
0.99
0.99
0.99
0.99

22.35
26.23
23.96
30.34
15.19
14.69
25.54
31.91

0.36

119

C.K. Rojas-Mayorga et al. / Journal of Molecular Liquids 208 (2015) 114121

the curve-tting capability of tested models may vary signicantly depending on the operating condition of micro-columns. Determination
coefcients (R2) and average errors (E, %) of all models are reported in
Tables 35 and these results have been used for a comparative analysis.
Specically, the errors in uoride adsorption tting were calculated
using

Table 5
Results of ANN model for data tting of uoride breakthrough curves.
Operating conditions

Results of ANN model

Feed ow, L/h

[F]feed, mg/L

R2

E, %

0.18

10
25
50
100
10
25
50
100

0.99
0.99
0.99
0.99
0.99
0.99
0.99
0.99

12.4
3.3
4.8
9.4
5.3
3.2
4.0
4.5

0.36

%
% %
%
%
% %

%
ndat %F #outlet =F #feed %i; exp F #outlet =F #feed %
%
X
100
%
i;cal %
%
E
%
%
%
%
%

ndat i1
F #outlet =F #feed %i; exp
%
%

where the subscripts exp and cal correspond to experimental and calculated values of breakthrough curve data and ndat is the number of experimental points, respectively.
In general, the analytical breakthrough curve models showed systematic discrepancies for data correlation of uoride adsorption patterns, see Fig. 4. In particular, Thomas and Yan models showed the
worst performance for data tting of uoride breakthrough curves
where their determination coefcients ranged from 0.54 to 0.99 with
mean errors from 7.9 to 34.9%, see Table 3. Although the Yan model is
better than the Thomas model, the Yan-based adsorption bed capacities
(qbed,Y) differ signicantly from the bed capacities calculated via
the integration of experimental data. In contrast, Thomas model provides reasonable estimations of the bed adsorption capacities for several
breakthrough curves. It is convenient to remark that the concentrationtime proles obtained in the xed-bed experiments suggest that the
axial dispersion may play an important role on uoride adsorption at
tested conditions. Therefore, it could be expected that the Thomas
model may fail for breakthrough curve data correlation because this
model assumes an adsorptiondesorption process with negligible
axial dispersion [2,9]. On the other hand, the Yan model is considered
as an empirical equation that could overcome some drawbacks of
Thomas model [25]. Unfortunately, both models may fail for the accurate modeling of asymmetric breakthrough curves as those reported
in this study.

concentration independently of the feed ow used. Note that high

values of retardation factor indicated a better removal performance of
the packed-bed column. The estimated values of bed capacities for uoride adsorption ranged from 3.3 to 18.5 mg/g at tested experimental
conditions. These bed capacities correspond to a degree of bed utilization of 10.759.7%. Results of these calculations are consistent with
the dynamic performance reported for water deuoridation using commercial bone chars in micro-columns where the bed utilization may
range from 35 to 50% [6]. Bed capacities for uoride adsorption and,
consequently the degree of bed utilization, increased with both feed
uoride concentration and ow. It is convenient to remark that the uoride uptakes of the novel aluminum-doped bone char are higher (up to
500%) than those values reported for packed-bed micro-columns using
commercial bone chars, i.e., 1.03.0 mg/g [6].
3.3. Numerical performance of tested breakthrough curve models
Proposed models were tted to the experimental uoride adsorption data and the model-calculated breakthrough curves are reported
in Fig. 4. As expected, each model has different capabilities for
representing the transient behavior of uoride adsorption pattern in
the packed-bed columns of aluminum-modied bone char. However,

Thomas model

0.9
0.8

0.8

0.7

0.7

0.6

0.6

0.5

0.5

0.4

0.4

0.3

5%
10%

0.2
0.1

a)

0.2
0.1

0
0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

b)

1
cal

0
0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

0.5

0.6

0.7

0.8

0.9

1.0

Mass transfer model

0.9

c)

Yan model

0.9

0.3

[ F ]oulet
[ F ] feed

17

0.8

0.8

0.7

0.7

0.6

0.6

0.5

0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0
0.0

0.1

0.2

0.3

0.4

0.5

ANNs model

0.9

0.6

0.7

0.8

0.9

1.0

d)

[ F ]oulet
[ F ] feed

0
0.0

0.1

0.2

0.3

0.4

exp

Fig. 5. Experimental (exp) and predicted (cal) breakthrough curves for uoride adsorption on aluminum-doped bone char in packed-bed micro-columns. Model: a) Thomas equation,
b) Yan equation, c) mass transfer model and c) articial neural networks.

120

C.K. Rojas-Mayorga et al. / Journal of Molecular Liquids 208 (2015) 114121

Table 6
Error analysis of tested tting approaches for different regions of the uoride breakthrough curve.
Error (%) for
F #outlet
F #feed

Model

0:0 b

Thomas equation
Yan equation
Mass transfer equation
ANN

21.4 17.4
40.3 31.1
66.3 36.7
14.5 18.8

b 0:2

0:2 b

F #outlet
F #feed

b 0:8

19.9 13.6
10.9 6.7
4.18 7.2
2.2 1.8

0:8 b

F #outlet
F #feed

b 1:0

14.2 6.2
6.3 5.6
1.82 1.8
2.2 1.3

The uoride breakthrough curves are better predicted by ANN and

mass transfer models, see Fig. 5 and Table 6. Eq. (7) is a simple mass
transfer equation that may offer a good performance for asymmetric
breakthrough curve modeling. In fact, this model can outperform the
traditional Thomas and Yan models especially for experimental data obtained at low feed ow and uoride concentration. In general, the determination coefcients of mass transfer model are higher than 0.98 with a
mean error from 15.2 to 31.9%, see Table 4. The calculated values of the
apparent dispersion coefcient DDap ranged from 0.012 to 0.115 cm2/s,
while the adsorption constant Kp ranged from 0.43 to 1.1 L/g. On the
other hand, the articial neural network model provides outstanding
correlations for the uoride breakthrough curves and its performance
is better than those obtained for Thomas and Yang equations and the
mass transfer model, see Fig. 4. For all curves, the determination coefcients of ANN model are 0.99 and the tting errors ranged from 3.2 to
12.4%, see Table 5.
A detailed error analysis has been performed for different regions of
the breakthrough curves where [F]outlet / [F]feed has been divided in
three zones: (00.2), (0.20.8) and (0.81.0). Results of this analysis
are reported in Table 6 and Fig. 5. Herein, it is convenient to note that
Fig. 5 contains baselines for data analysis considering relative errors in
the uoride concentration patterns of 5 and 10%, respectively. All
models showed the worst tting performance and signicant discrepancies between experimental and predicted adsorption patterns for
the relative concentration region b 0.2 (i.e., the breakthrough zone). In
fact, all models tended to underestimate the dynamic adsorption
0.4

pattern in this zone. Thomas and Yan models can show mean percentage errors up to 40%. It is interesting to remark that the mass transfer
model showed the worst performance for breakthrough modeling in
this zone with a mean error of 66%. Mass transfer-type models usually
offer an acceptable accuracy for the breakthrough curve correlation
and prediction. However, this result indicated that this mass transfer
model is not valid at [F]outlet / [F]feed ratios lower than the breakthrough point in highly asymmetric curves. The ANN outperformed
other breakthrough curve models in this region but its errors were
still high, see Table 6 and Fig. 5.
The capabilities of tested models improved for the region
0.2 b [F ] outlet / [F ] feed b 0.8 where both the mass transfer model
and ANN showed better data ts (i.e., errors b 5%). Thomas and Yan
equations tended to underestimate the adsorption prole at this zone
of breakthrough curve and their modeling errors were higher than
10%. At the saturation zone, [F]outlet / [F]feed N 0.8, the calculated adsorption proles of the mass transfer model and ANN were in good
agreement with the experimental data. In fact, these models showed
error values lower than 2% inside this region and they outperformed
both analytical Thomas and Yan equations. Note that mass transfer
model may outperform the ANN in this region of the breakthrough
curve, while Thomas equation showed the worst predictions for uoride adsorption pattern.
Finally, this error analysis has been complemented using the plots of
residuals (ei) of the breakthrough models, which are reported in Fig. 6.
In particular, the residual ei is given by
%
%
%
%

ei F #outlet =F #feed %i; exp F #outlet =F #feed %

i;cal

These residual plots have been used to assess the quality of adsorption data tting. In a proper data tting, the residuals would have to be
randomly distributed around zero and they should not show any trend.
It is clear that the residual plot of ANN showed a randomly distribution
in comparison to the other breakthrough models. Note that the residual
plots of Thomas and Yan models showed non-random patterns and they
have an increasing trend that suggests that the error variance increases
0.4

Thomas model

0.3

0.3

0.2

0.2

0.1

0.1

0.0

Yan model

0
0.0

Residual ei

18

0.2

0.4

0.6

0.8

1.0

0.0

-0.1

-0.1

-0.2

-0.2

-0.3

-0.3

a) -0.4

b) -0.4

0.4

0.4

Mass transfer model

0.3

0.3

0.2

0.2

0.1

0.1

0.2

0.4

0.6

0.8

1.0

0.4

0.6

0.8

ANNs model

0
0

0.2

0.4

0.6

0.8

-0.1

-0.1

-0.2

-0.2

-0.3

-0.3

c) -0.4

0.2

d) -0.4
-

[ F ]oulet
[ F - ] feed
Fig. 6. Residual plots of tested modeling approaches used in data tting of the uoride breakthrough curves.

C.K. Rojas-Mayorga et al. / Journal of Molecular Liquids 208 (2015) 114121

with respect to [F]outlet / [F]feed, see Fig. 6. These results conrmed that
the best adsorption data tting has been obtained using the ANN model.
In summary, ANN model offered a good agreement between predicted and experimental uoride adsorption proles. It is convenient to remark that Du et al. [18], reported the comparison of a general rate model
and a traditional ANN model for data tting of adsorption of solanesol
using resins in packed-bed chromatographic columns. These authors
concluded that the ANN model may offer a better performance than
the tested general rate model. However, the breakthrough curves used
in data modeling showed a low degree of asymmetry. Results reported
in this study conrmed that an ANN model, with the optimal brain surgeon procedure, may offer a better data tting performance than those
obtained for a mass transfer model but using highly asymmetric breakthrough curves.
Finally, this study highlighted the challenging task of nding a
breakthrough model of general applicability. It is clear that breakthrough curve experiments are time consuming and, for some adsorbentpollutant systems, they could be expensive. Therefore, the
modeling and prediction of column performance at micro-scale level
is an interesting alternative if reliable models are applied. Based on
these facts, the articial neural networks offer additional advantages
for the packed-bed column design because these black-box models
may predict the adsorption performance at other operating conditions
used in its training process. In the absence of a universal breakthrough
curve model, ANN-based models emerge as promising approach for
the data tting of dynamic adsorption processes involved in water
treatment.

4. Conclusions
This study has analyzed and compared the capabilities and limitations of three approaches for the modeling of asymmetric breakthrough
curves of the uoride adsorption on a metal-modied bone char using
micro-columns. Overall, results indicated that each model has different
capabilities for representing the transient behavior of uoride adsorption pattern in the packed-bed columns of aluminum-modied bone
char and its performance depends on the operating conditions. Thomas
and Yan equations showed systematic discrepancies for data correlation
of uoride adsorption patterns and they offered the worst modeling
performance. On the other hand, the ANN and the mass transfer
model are better options for the data tting of uoride breakthrough
curves where ANN provides the best correlation properties. A detailed
error analysis indicated that all models may fail to predict the uoride
adsorption patterns at the breakthrough zone but ANN still offers the
best performance for the data tting. This study showed that ANN can
be effectively used for modeling and predicting the asymmetric breakthrough curves in the adsorption of priority water pollutants. This intelligent data processing algorithm offers additional advantages for the
packed-bed column design using experiments at micro-scale and ANN
model emerges as promising approach for non-linear data processing
and modeling involved in water treatment.

121

References
[1] W.J. Weber, Adsorption processes, Pure Appl. Chem. 37 (1974) 375392.
[2] Z. Xu, J. Cai, B. Pan, Mathematically modeling xed-bed adsorption in aqueous
systems, J. Zhejiang Univ. Sci. A Appl. Phys. Eng. 14 (2013) 155176.
[3] L. Lv, Y. Zhang, K. Wang, A.K. Ray, X.S. Zhao, Modeling of the adsorption breakthrough behaviors of Pb2+ in a xed bed of ETS-10 adsorbent, J. Colloid Interface
Sci. 325 (2008) 5763.
[4] K.H. Chu, Improved xed bed models for metal biosorption, Chem. Eng. J. 97 (2004)
233239.
[5] I.A. Aguayo-Villarreal, A. Bonilla-Petriciolet, V. Hernndez-Montoya, M.A. MontesMorn, 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.
[6] R. Tovar-Gmez, M.R. Moreno-Virgen, J.A. Dena-Aguilar, V. Hernndez-Montoya, A.
Bonilla-Petriciolet, M.A. Montes-Morn, Modeling of xed-bed adsorption of uoride on bone char using a hybrid neural network approach, Chem. Eng. J. 228
(2013) 10981109.
[7] 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) 557562.
[8] C.E.A. Padilha, C.A.A. Padilha, D.F.S. Souza, J.A. Oliveira, G.R. Macedo, E.S. Santos, Prediction of rhamnolipid breakthrough curves on activated carbon and Amberlite
XAD-2 using articial neural network and group method data handling models,
J. Mol. Liq. 206 (2015) 293299.
[9] H.C. Thomas, Heterogeneous ion exchange in a owing system, J. Am. Chem. Soc. 66
(1944) 16641666.
[10] G. Bohart, E.Q. Adams, Some aspects of the behavior of charcoal with respect to chlorine, J. Am. Chem. Soc. 42 (1920) 523544.
[11] Y.H. Yoon, J.H. Nelson, Application of gas adsorption kinetics, Am. Ind. Hyg. Assoc. J.
45 (1984) 509516.
[12] R.M. Clark, Evaluating the cost and performance of eld-scale granular activated
carbon systems, Environ. Sci. Technol. 21 (1987) 573580.
[13] G. Yan, T. Viraraghavan, M. Chen, A new model for heavy metal removal in a
biosorption column, Adsorpt. Sci. Technol. 19 (2001) 2543.
[14] M.S. Shafeeyan, W.M.A.W. Daud, A. Shamiri, A review of mathematical modeling of
xed-bed columns for carbon dioxide adsorption, Chem. Eng. Res. Des. 99 (2014)
961988.
[15] C. Faur, A. Cougnaud, G. Dreyfus, P. Le 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.
[16] E. Oguz, M. Ersoy, Removal of Cu2+ from aqueous solution by adsorption in a xed
bed column and neural network modeling, Chem. Eng. J. 164 (2010) 5662.
[17] I.A. Basheer, M. Hajmeer, Articial neural networks: fundamentals, computing,
design and application, J. Microbiol. Methods 43 (2000) 331.
[18] X. Du, Q. Yuan, J. Zhao, Y. Li, Comparison of general rate model with a new modelarticial neural network model in describing chromatographic kinetics of solanesol
adsorption in packed column by macroporous resins, J. Chromatogr. A 1145 (2007)
165174.
[19] C.K. Rojas-Mayorga, A. Bonilla-Petriciolet, I.A. Aguayo-Villarreal, V. HernndezMontoya, M.R. Moreno-Virgen, R. Tovar-Gmez, M.A. Montes-Morn, Optimization
of pyrolysis conditions and adsorption properties of bone char for uoride removal
from water, J. Anal. Appl. Pyrolysis 104 (2013) 1018.
[20] C.K. Rojas-Mayorga, J. Silvestre-Albero, I.A. Aguayo-Villarreal, D.I. Mendoza-Castillo,
A. Bonilla-Petriciolet, A new synthesis route for bone chars using CO2 atmosphere
and their application as uoride adsorbents, Microporous Mesoporous Mater. 209
(2014) 3844.
[21] Standard Methods for Examination of Water and Wastewater, 20th ed American
Public Health Association, 1998.
[22] B. Hassibi, D.G. Stork, G.J. Wolff, Optimal brain surgeon and general network
pruning, Proc. IEEE I.C.N.N. (1993) 293299.
[23] M. Figueiredo, A. Fernando, G. Martins, J. Freitas, F. Judas, H. Figueiredo, Effect of the
calcination temperature on the composition and microstructure of hydroxyapatite
derived from human and animal bone, Ceram. Int. 36 (2010) 23832393.
[24] S. Lurtwitayapont, T. Srisatit, Comparison of lead removal by various types of swine
bone adsorbents, Environ. Asia 3 (2010) 3238.
[25] F. Zeinali, A.A. Ghoreyshi, G.D. Najafpour, Adsorption of dichloromethane from
aqueous phase using granular activated carbon: isotherm and breakthrough curve
measurements, Middle-East J. Sci. Res. 5 (4) (2010) 191198.