Chemical Engineering Science 66 (2011) 3039–3044

Contents lists available at ScienceDirect

Chemical Engineering Science

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

Artificial Neural Network modeling of solubility of supercritical

carbon dioxide in 24 commonly used ionic liquids
Ali Eslamimanesh a, Farhad Gharagheizi b, Amir H. Mohammadi a,c,n, Dominique Richon a
a
MINES ParisTech, CEP/TEP—Centre Énergétique et Procédés, 35 Rue Saint Honoré, 77305 Fontainebleau, France
b
Saman Energy Giti Co., 3331619636 Tehran, Iran
c
Thermodynamics Research Unit, School of Chemical Engineering, University of KwaZulu-Natal, Howard College Campus, King George V Avenue, Durban 4041, South Africa

a r t i c l e i n f o abstract

Article history: Application of supercritical CO2 for separation of ionic liquids from their organic solvents or extraction
Received 19 August 2010 of various solutes from ionic liquid solvents have found great interest during recent years. Knowledge
Received in revised form of phase behaviors of the mixtures of supercritical CO2 þ ionic liquids is therefore drastic in order to
26 February 2011
efficiently design such separation processes. In this communication, Artificial Neural Network
Accepted 10 March 2011
procedure has been applied to represent the solubility of supercritical CO2 in 24 mostly used ionic
Available online 16 March 2011
liquids. An optimized Three-Layer Feed Forward Neural Network using critical properties of ionic
Keywords: liquids and operational temperature and pressure has been developed. Application of this model for
Supercritical carbon dioxide 1128 data points of 24 ionic liquids show squared correlation coefficients of 0.993 and average absolute
Ionic liquids
deviation of 3.6% from experimental values for calculated/estimated solubilities. The aforementioned
Artificial Neural Network
deviations show the prediction capability of the presented model.
Solubility
Model & 2011 Elsevier Ltd. All rights reserved.
Prediction

1. Introduction Cheek et al., 2008), and applications in membrane-based systems

for separation (Bennett and Leo, 2004). However, their applica-
Ionic liquids are organic salts that are liquid at room tempera- tions in supercritical processes have been more focused by the
tures (Carvalho et al., 2009a; Song et al., 2010). They are generally authors (Carvalho et al., 2009a). Blanchard and coworkers were
composed of a large organic cation and organic or inorganic the first (Blanchard and Brennecke, 2001; Blanchard et al., 1999,
anions (Song et al., 2010). The applications of ionic liquids have 2001) to state that the problem of the recovery of the compounds
generated number of discussions in the past decade. This is from the ionic liquid media can be solved through using super-
mainly because of their physico-chemical properties, which are critical CO2 that removes a desired solute from an ionic liquid
able to be adjusted through combination of cations and anions. solvent without contamination (Carvalho et al., 2009a). Carbon
This phenomenon allows the researchers to design specific dioxide as an inexpensive, non-explosive, available, and easy to
solvents for the development of more efficient processes and remove from the extracted products (Esmaeilzadeh et al., 2009;
products (Carvalho et al., 2009a; Song et al., 2010; Ren et al., Huang et al., 2001; Housaindokht et al., 2007; Yamini et al., 2003;
2010). No flammability, high thermal stability, a wide liquid Ismadji, 2008), is considered to be the most suitable fluid in
range, and electric conductivity are among the other elements supercritical processes e.g. 1: for extraction of fatty acid esters,
that make ionic liquids become interesting more and more vegetable oils, and antioxidants in food engineering industries
(Carvalho et al., 2009a; Song et al., 2010; Ren et al., 2010). and 2: for solving solid phenazopyridine, propranolol, and methyl
So far, several industrial applications of ionic liquids have been salicylate in pharmaceutical industries (Yamini et al., 2003;
proposed including application as solvents in organic synthesis Ismadji, 2008).
(Boon et al., 1986; Fry and Pienta, 1985), homogeneous and Due to the difficulties of experimental measurements and also
biphasic transfer catalysts (Zhao, 2006; Sheldon, 2001; Wong their time-consuming and costly nature, it is desirable to develop
et al., 2006), for electrical deposition (Zein El Abedin et al., 2007; predictive methods for estimating the phase behavior of these
kinds of systems. Furthermore, in order to efficiently design the
processes dealing with mixtures of CO2 in ionic liquids, one needs
n
Corresponding author at: MINES ParisTech, CEP/TEP—Centre Énergétique et to know the operational conditions to reach a desired CO2
Procédés, 35 Rue Saint Honoré, 77305 Fontainebleau, France.
Tel.: þ33 1 64 69 49 70; fax: þ 33 1 64 69 49 68.
solubility in the applied ionic liquid solvent. In recent years,
E-mail address: [email protected] several attempts have been done regarding this concept. The
(A.H. Mohammadi). Peng–Robinson (PR) (Peng and Robinson, 1976) equation of state

0009-2509/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ces.2011.03.016
3040 A. Eslamimanesh et al. / Chemical Engineering Science 66 (2011) 3039–3044

(EoS) was applied by Shariati and Peters (2004b) to model the

solubility of CHF3 in [C2mim][PF6] from 309.3 to 367.5 K and
1.6 to 51.6 MPa. Breure et al. (2005) used the concept of group
contribution (GC) method coupled with equation of state (EoS) to
model the phase behavior of binary systems of ionic liquids and
supercritical carbon dioxide. In addition, Yokozeki and Shiflett
(2006) used successfully a generic Redlich–Kwong (RK) (Redlich
and Kwong, 1949) EoS in order to understand the global phase
behavior of trifluoromethane þionic liquid [bmim][PF6] mix-
tures. They gained satisfactory results for vapor–liquid-equilibria
(VLE) and vapor–liquid–liquid-equilibria (VLLE) in such systems
at wide ranges of pressures.
The truncated perturbed chain polar statistical associating
fluid theory tPC-PSAFT (Karakatsani and Economou, 2006) has
been applied by Kroon et al. (2006) to calculate the solubility of Fig. 1. The schematic structure of a 3FFNN used in this study (Kalogirou, 2001).
CO2 in several ionic liquids at high pressures. Wang et al. (2006)
have evaluated the solubility of several gas samples in classical
ionic liquids by the Square Well Chain Fluid (SWCF) (Liu and Hu, post-optimization after generating the base-structure of the
1996, 1998; Hu et al., 1996) EoS. More theoretical and experi- 3FFNN. The mathematic formulation of this strategy can be
mental studies on phase behavior of the mixtures of supercritical written as follows:
carbon dioxide and ionic liquids can be found elsewhere (Carvalho
OutputðiÞ ¼ W2  tanhðW1  inputðiÞ þ b1 Þ þb2 ð1Þ
et al., 2009b, 2010; Lydersen, 1955; Joback and Reid, 1987; Yokozeki
et al., 2008; Schilderman et al., 2007; Kim et al., 2005; Oh and Lee, where i refers to the data point, the parameter W1 is the weight
2006; Raeissi and Peters, 2009; Costantini et al., 2005; Shiflett and relating the first layer to the second layer, W2 is the weight
Yokozeki, 2005, 2009; Shariati and Peters, 2004a; Machida et al., relating the second to the third or output layer, b1 denotes the
2009; Kurnia et al., 2009). bias to the second layer, and b2 stands for the bias of third or
However, because ionic liquid molecules are highly asym- output layer. W1, W2, b1, and b2 should be determined by a
metric neutral ion pairs with a large dipole moment as a result of process called ‘‘Training’’.
the charge distribution over the ion pair (Kroon et al., 2006) and also To generate a 3FFNN model, some parameters should be
the fact that CO2 molecules have a quadruple moment, (Kroon et al., defined regarding the proposed problem. The required para-
2006) the applied thermodynamic relations, thus, should be able to meters are the number of inputs (nip), number of outputs (nop),
take these strong interactions into account. Besides, the effects of and the number of objects (nobj) i.e. number of data points with
high pressures generally encountered in supercritical operations known input parameter (ip) and corresponding output parameter
have to be considered. All of the aforementioned thermodynamic (op). For instance, in a problem with 5 input parameters and one
models have been developed especially for a desired system mainly output parameter with 1140 data points, the nip, nop, and nobj are
because of these two reasons and not for a set of systems as we are 5, 1, and 1140, respectively.
interested in. Therefore, there is a need to develop general models We can determine the dimensions of each part in Eq. (1) using
capable of predictions of phase behaviors of such systems for n neurons. The input(i) is a row containing nip input parameters of
different kinds of ionic liquids that are considered now to be the ith object, so it is of dimension nip  1 and W1 is of dimension
alternative of conventional industrial solvents. n  nip. Therefore, the W1  input(i) should be of dimension
In this work, we focus on application of Artificial Neural n  nip. The b1 is of dimension n  nip. Thus, the results of the
Network for representing/predicting the solubility of supercritical second layer output are of dimension n  1. It is easy to figure out
CO2 in different types of ionic liquids in VLE region. the dimensions of W2 and B2. W2 should be of dimension nop  n,
and b2 should be of dimension nop  1. As a result, the output(i) is
of dimension nop  1.
2. Artificial Neural Network Algorithm There are unknown parameters for a desired number of
neurons n, which in a typical 3FFNN model are W1, W2, b1, and
Artificial Neural Networks (ANNs) are composed of simple b2. These parameters should be obtained using the ‘‘Training’’
elements working in a parallel computational strategy. These process. This process applies an optimization method to minimize
elements are called neuron and inspired by biological nervous an objective function defined between output and the experi-
systems (Hagan et al., 2002). ANNs have been applied for various mental values. One of the most widely used objective functions in
applications so far (Gharagheizi et al., 2011a, b; Chouai et al., 2002; 3FFNNs training process is the ‘‘mean square of errors’’ (MSE).
Piazza et al., 2006; Scalabrin et al., 2006; Chapoy et al., 2007; The parameters W1, W2, b1 and b2 are the most significant
Mohammadi and Richon, 2007, 2008a, b, 2010; Mohammadi et al., factors for selecting the optimization method. Firstly and perhaps
2008, 2010a, b). most importantly, the method should be accurate enough to deal
Several types of ANN strategies have been proposed in the with the specified problem. Another point to consider is that the
literature. One of the most-widely used is called ‘‘Three-Layer applied method should be robust when we have several tens or
Feed Forward Neural Networks (3FFNN) with Sigmoid (Hyper- hundreds of parameters in a simple 3FFNN. In this work, the
bolic) Transfer Function’’. This type is used to generate a non- Levenberg–Marquart optimization method is used as previous
linear correlation between input and output parameters. In other studies have shown that it is the most appropriate one regarding
words, this strategy is somehow used as a regression tool. The the mentioned characteristics (Gharagheizi et al., 2011a–e; Chouai
schematic structure of a FFNN is shown in Fig. 1 (Kalogirou, 2001). et al., 2002; Piazza et al., 2006; Scalabrin et al., 2006; Chapoy et al.,
This type of ANN normally contains three layers. The first one is 2007; Mohammadi and Richon, 2007, 2008a, b, 2010; Mohammadi
the input layer and the second one is the hidden layer of the et al., 2008, 2010a, b; Kalogirou, 2001; Gharagheizi, 2009a, b, c;
3FFNN. It consists of n neurons. The parameter n is one of the Gharagheizi and Sattari, 2009; Gharagheizi et al., 2009; Gharagheizi
main parameters of the 3FFNN method and should be obtained by and Mehrpooya, 2008; Sattari and Gharagheizi, 2008; Gharagheizi
 A. Eslamimanesh et al. / Chemical Engineering Science 66 (2011) 3039–3044 3041

et al., 2008; Gharagheizi and Fazeli, 2008; Gharagheizi and Alamdari, 4. Results and discussion
2008; Gharagheizi, 2007a, b, 2008, 2010; Vatani et al., 2007;
Mehrpooya et al., 2010; Mehrpooya and Gharagheizi, 2010). In the first step, the input and output parameters of the model
There are some empirical correlations (Hagan et al., 2002) for are defined. The input parameters include the critical temperature
determining n (number of neurons in hidden layer) of a 3FFNN, (Tc), the critical pressure (Pc), and the acentric factor (o) of ionic
however, n completely depends on the nature of the problem. liquids along with the temperature (T), and pressure (p). There-
Therefore, it should be post-optimized i.e. searching must be done fore, the nip is equal to 5.
in order to find its best value. Mapping the data in a smoother space should be paid much
Of particular interest is the fact that the main data set should attention in the calculations. Generally, the solubility of the
be divided into three main sub-data sets. This process is needed to supercritical carbon dioxide in the investigated ionic liquids
ensure the validity of the obtained 3FFNN as well as to show its ranges typically from 2  10  3 to 9  10  1 (mole fraction) in
prediction capability. It is required that the main data set is temperature range 304.25–636.3 K and pressure range 0.1–
divided into three sub-data sets called ‘‘Training set’’, ‘‘Validation 97.1 bar (Carvalho et al., 2009a; Song et al., 2010; Ren et al.,
set’’, and ‘‘Test set’’. In the present work, the ‘‘Training set’’ is used 2010; Boon et al., 1986; Fry and Pienta, 1985; Zhao, 2006; Sheldon,
to generate the 3FFNN model. The ‘‘Validation (Optimization) set’’ 2001; Wong et al., 2006; Zein El Abedin et al., 2007; Cheek et al.,
is applied to control the optimization process of ‘‘Training’’. The 2008; Bennett and Leo, 2004; Blanchard and Brennecke, 2001;
‘‘Test set’’ is used to test the prediction capability of the obtained Blanchard et al., 1999, 2001; Esmaeilzadeh et al., 2009; Huang
3FFNN. The effects of the allocation percent of each one of the et al., 2001; Housaindokht et al., 2007; Yamini et al., 2003; Ismadji,
three sub-data sets from the main data sets have been already 2008; Shariati and Peters, 2004b; Breure et al., 2005; Yokozeki and
studied (Gharagheizi, 2007b). In this work, ninety percent of the Shiflett, 2006; Karakatsani and Economou, 2006; Kroon et al.,
data of the main data set are randomly assigned to the ‘‘Training 2006; Wang et al., 2006; Liu and Hu ,1996, 1998; Hu et al., 1996;
set’’ and the remaining part is equally divided between the Carvalho et al., 2010; Yokozeki et al., 2008; Schilderman et al.,
‘‘Validation set’’ and the ‘‘Test set’’. 2007; Kim et al., 2005; Oh and Lee, 2006; Raeissi and Peters, 2009;
Costantini et al., 2005; Shiflett and Yokozeki, 2005, 2009; Shariati
and Peters, 2004b; Machida et al., 2009; Shin et al., 2008; Shin and
Lee, 2008). The natural logarithm of solubility was regarded as
3. Investigated compounds
output of the 3FFNN algorithm to decrease computational errors.
This issue was done to prevent obtaining negative output values
There are many ionic liquids considered nowadays to have
(trivial values) for the solubility. Furthermore, the natural loga-
great potential as being alternatives of conventional solvents in
rithm is a useful function to map a data set from several orders of
future industries. 24 known ionic liquids were investigated in this
magnitude to a smoother space. Transferring the solubility space
study to present a reliable model. Table 1 shows the physical
to a new space that ranges from  2.68 to  0.07 was done
properties values of these compounds accompanied with the
considering this mathematical issue. The nop is set to 1.
references. A group contribution method (Lydersen, 1955;
In the next step, the main data set was divided into three sub-
Joback and Reid, 1987) was applied to generate the physical
data sets. The ‘‘Training set’’ received ninety percent of the main
properties when the experimental values are not available.
data set i.e. 904 data points. Each one of the ‘‘Validation’’ and the
‘‘Test’’ sets received five percent each of the main data set. In
other words, each one received 112 data points. Consequently, the
Table 1
nobj is 1128. The process of division of these data to the sub-sets
The critical properties and acentric factors of compounds used in this work.
is performed randomly.
Compound Tc (K) Pc (MPa) o Refs.a Finally, several 3FFNNs modules were generated assuming
numbers 1 through 50 for n (number of neurons in hidden layer)
[C2-mim][Tf2N] 788.05 33.1 1.225 Shin et al. (2008) using the previously described procedure. The most convinced
[C4-mim][Tf2N] 831.39 26.9 1.289 Shin et al. (2008)
results were observed at n ¼19, a value which contributes from
[C5-mim][Tf2N] 1249.43 26.3 0.412 Carvalho et al. (2009a)
[THTDP][Tf2N] 1588.22 85.1 0.8801 Carvalho et al. (2010) post-optimization and good results in optimization and test steps.
[Dmim][Tf2N] 1345.1 18.7 0.5741 Ren et al. (2010) As previously stated, the Levenberg–Marquart (LM) (Hagan et al.,
[P14,6,6,6][Tf2N] 1604.75 8.513 0.7748 Song et al. (2010) 2002) method was used to train the ANN module. There are also
[C2-mim][TfO] 756.89 34.8 0.783 Shin and Lee (2008)
more accurate optimization methods other than the LM, however,
[C4-mim][TfO] 796.73 28.1 0.879 Shin and Lee (2008)
[C8-mim][TfO] 879.6 19.4 1.031 Shin and Lee (2008) they need much more convergence time. In other words, the more
[C6-mim][TfO] 837.5 23.1 0.963 Shin and Lee (2008) accurate optimization the more time needed for the algorithm to
[BMP][TfO] 933.39 25.859 0.6099 Song et al. (2010) converge to the global optimum. The LM is most-widely used
[C8-mim][BF4] 726.1 16 0.9954 Blanchard et al. (2001) algorithm for training due to being robust and accurate enough to
[bmim][BF4] 632.3 20.4 0.8489 Yokozeki et al. (2008)
deal with the considered type of study.
[N-bupy][BF4] 597.6 20.3 0.8307 Blanchard et al. (2001)
[C4-mim][DCA] 782.96 24.4 0.8419 Carvalho et al. (2009b) The statistical parameters of the model have been reported in
[emim][AC] 807.1 29.2 0.5889 Shiflett and Yokozeki (2009) Table 2. Table 3 indicates the ranges of calculated values of the
[emim][TFA] 824.67 28.86 0.6808 GMC solubilities by the presented ANN model. The comparison between
[emim][PF6] 663.5 19.5 0.6708 Shariati and Peters (2004a)
the calculated/predicted and the experimental values is depicted in
[THTDP][CL] 1225.8 79.5 0.7698 Carvalho et al. (2010)
[HEA] 710.66 46.36 0.8087 GMC
Fig. 2 for three sub-data sets used in this study. The calculated
[HEL] 821.61 43.54 1.2287 GMC solubility values accompanied with average absolute deviations from
[BHEAA] 836.38 34.85 1.2876 GMC experimental data and different types of sub-data sets are presented
[BHEAL] 950.45 33.21 1.666 GMC as Supplementary material. To check the prediction capability of the
[HHEMEA] 853.92 29.28 1.3308 GMC
presented model, the model has been applied to predict the solubility
[HHEMEL] 970.51 28.24 1.6719 GMC
of supercritical carbon dioxide in those ionic liquids existed in the
GCM: properties evaluated by group contribution method proposed by Lydersen ‘‘Test set’’ and also in another ionic liquid (HEL) excluding those
(1955) and Joback and Reid (1987). investigated in the three aforementioned data sets. As can be seen
a
References of experimental data. in Fig. 2 and Table 3 and Table 1 in Supplementary material, the
3042 A. Eslamimanesh et al. / Chemical Engineering Science 66 (2011) 3039–3044

predicted results have been coincided well with the existed experi- of the model over the ‘‘Training set’’, the ‘‘Validation set’’, the ‘‘Test
mental values. set’’, and the main data set were calculated as 0.993, 0.995, 0.995,
The squared correlation coefficient (R2), average absolute devia- 0.993, 3.7%, 3.4%, 3%, 3.6%, 0.99, 1.24, 0.99, 1.02, 0.06, 0.06, 0.05, and
tions (AAD), standard deviation errors, and root mean square errors 0.06, respectively. The mat file (MATLABs software file) of the
obtained 3FFNN model and the procedure of application of the
Table 2 program are freely available upon request from the authors.
Statistical parameters of the presented model. Another element to consider is that we have optimized the
parameters of the model using the characteristics of the investi-
Statistical parameter Value
gated ionic liquids. Although we have tested our model for
Training set prediction of the solubility of CO2 in an ionic liquid excluding
R2 0.993 those applying for developingþoptimizationþtesting the model,
Average absolute deviationa 3.7% extension of the proposed model to other systems containing
Standard deviation error 0.99 ionic liquids are not recommended. However, it may be used for a
Root mean square error 0.06
Nb 904
rough estimation of prediction of solubility of supercritical CO2 in
other ionic liquids with the chemical structures slightly/comple-
Validation set
tely different from those investigated.
R2 0.995
Average absolute deviation 3.4% A significant point should not be omitted from our discussions. In
Standard deviation error 1.24 this work, we have tried to use all of the solubility data of super-
Root mean square error 0.06 critical carbon dioxide in the studied ionic liquids which are available
N 112 in the literature as they are not so numerous. However, we are well
Test set aware that maybe not all of the data are fully trustable. This is due to
R2 0.995 the inaccuracy of the techniques used. For instance, the observation
Average absolute deviation 3%
technique for detecting the bubble point of mixtures inside a variable
Standard deviation error 0.99
Root mean square error 0.05 volume cell can much more inaccurate than the indirect technique
N 112 relying on pressure–volume dependence which displays a break
Training þ Validationþ Test set
point (Stefani et al., 2004).
R2 0.993
Average absolute deviation 3.6%
Standard deviation error 1.02 5. Conclusions
Root mean square error 0.06
N 1128
In this work, a Feed Forward Artificial Neural Network
a P
N
9Calc:ðiÞ=Pred:ðiÞExp:ðiÞ9 was successfully applied for calculation/estimation of the solubility
%AAD ¼ 100
N Exp:ðiÞ .
b i
Number of experimental data points. of supercritical carbon dioxide in various ionic liquids. For presenting

Table 3
The calculated and predicted solubility values (mole fraction) by the presented model along with absolute average deviations from experimental data.

Ionic liquid Temperature Pressure range N Calculated solubility range/ AAD% Refs.a
range (K) (bar) mole fraction

[C2-mim][Tf2N] 314.15–636.3 7.18–435.75 126 [2.121  10  1y8.12  10  1] 2.3 Ren et al. (2010), Yokozeki et al. (2008), and Shin et al. (2008)
[C4-mim][Tf2N] 313.1–363.26 10.28–561.53 95 [2.099  10  1y8.826  10  1] 2.2 Carvalho et al. (2009b), Oh and Lee (2006), Raeissi and Peters
(2009), and Shin et al. (2008)
[C5-mim][Tf2N] 312.85–363.29 10.44–598.05 54 [2.056  10  1y8.376  10  1] 2.8 Carvalho et al. (2009a) and Shin et al. (2008)
[THTDP][Tf2N] 313.05–363.53 3.55–395.36 90 [1.632  10  1y8.379  10  1] 1.9 Carvalho et al. (2010)
[Dmim][Tf2N] 323.15–343.15 14.39–20.03 14 [2.603  10  1y8.087  10  1] 2.3 Ren et al. (2010)
[P14,6,6,6][Tf2N] 304.25–375.35 7.2–222 76 [3.564  10  1y7.996  10  1] 1.9 Song et al. (2010)
[C2-mim][TfO] 314.05–344.55 10–378 44 [1.668  10  1y6.261  10  1] 1.8 Shin and Lee (2008)
[C4-mim][TfO] 314.05–344.55 10.2–375 52 [2.012  10  1y6.568  10  1] 2.9 Shin and Lee (2008)
[C6-mim][TfO] 314.05–344.55 16.5–363 56 [2.811  10  1y7.305  10  1] 1.7 Shin and Lee (2008)
[C8-mim][TfO] 314.05–344.55 8–340 52 [2.050  10  1y7.691  10  1] 1.7 Shin and Lee (2008)
[BMP][TfO] 304.25–375.35 6.9–736.4 160 [1.010  10  1y7.990  10  1] 4.5 Song et al. (2010)
[C8-mim][BF4] 313.15–333.15 15.61–93.73 21 [1.618  10  1y6.595  10  1] 7.9 Blanchard et al. (2001)
[bmim][BF4] 323.15–348.15 0.1015–20 18 [2.100  10  3y1.820  10  1] 11.5 Yokozeki et al. (2008)
[N-bupy][BF4] 313.15–333.15 15.47–95.8 21 [1.449  10  1y5.557  10  1] 4.5 Blanchard et al. (2001)
[C4-mim][DCA] 313.03–363.25 17.71–736.4 30 [2.020  10  6y6.210  10  1] 6.1 Carvalho et al. (2009b)
[emim][AC] 323.1–348.2 0.1–19.996 18 [5.100  10  3y4.354  10  1] 14 Shiflett and Yokozeki (2009)
[emim][TFA] 323.1–348.2 0.1–19.996 18 [6.400  10  3y4.741  10  1] 12.4 Shiflett and Yokozeki (2009)
[emim][PF6] 308.14–366.03 14.9–971 74 [1.003  10  1y6.180  10  1] 2.3 Shariati and Peters (2004a)
[THTDP][CL] 313.03–363.68 1.68–245.7 49 [1.223  10  1y8.920  10  2] 5.4 Carvalho et al. (2010)
[HEA] 313.15–328.15 1.24–15.42 12 [3.400  10  3y1.364  10  3] 6 Kurnia et al. (2009)
[BHEAA] 313.15–328.15 1.26–15.05 12 [8.100  10  3y8.530  10  2] 11.8 Kurnia et al. (2009)
[BHEAL] 313.15–328.15 1.21–15.98 12 [3.600  10  3y7.180  10  2] 6 Kurnia et al. (2009)
[HHEMEA] 313.15–328.15 1.34–15.16 12 [2.400  10  3y5.150  10  2] 4.3 Kurnia et al. (2009)
[HHEMEL] 313.15–328.15 1.58–15.62 12 [2.900  10  3y6.050  10–2] 2.3 Kurnia et al. (2009)

Total 1128 5

Predicted solubility range/

mole fraction
[HEL] 313.15–328.15 1.27–15.56 12 [3.400  10  3y8.920  10  2] 2.8 Kurnia et al. (2009)

a
References of experimental data.
 A. Eslamimanesh et al. / Chemical Engineering Science 66 (2011) 3039–3044 3043

dioxide in 1-alkyl-3-methylimidazoliumbis(trifluoromethylsulfonyl)imide
ionic liquids. J. Supercrit. Fluid 48, 99–107.

Carvalho, P.J., Alvarez, V.H., Marrucho, I.M., Aznar, M., Coutinho, J.A.P., 2009b.
High-pressure phase behavior of carbon dioxide in 1-butyl-3-methylimidazo-
lium bis(trifluoromethylsulfonyl) imide and 1-butyl-3-methylimidazolium
dicyanamide ionic liquids. J. Supercrit. Fluids 50, 105–111.

Carvalho, P.J., Alvarez, V.H., Marrucho, I.M., Aznar, M., Coutinho, J.A.P., 2010. High
carbon dioxide solubilities in trihexyltetradecylphosphonium-based ionic
liquids. J. Supercrit. Fluids 52, 258–265.
Chapoy, A., Mohammadi, A.H., Richon, D., 2007. Predicting the hydrate stability
zones of natural gases using artificial neural networks. Oil Gas Sci. Technol.—
Rev. IFP 62, 701–706.
Cheek, G.T., O’Grady, W.E., Zein El Abedin, S., Moustafa, E.M., Endres, F., 2008.
Studies on the electrodeposition of magnesium in ionic liquids. J. Electrochem.
Soc. 155, D91–D95.
Chouai, A., Laugier, S., Richon, D., 2002. Modeling of thermodynamic properties
using neural networks: application to refrigerants. Fluid Phase Equilib. 199,
53–62.
Costantini, M., Toussaint, V.A., Shariati, A., Peters, C.J., Kikic, I., 2005. High-pressure
phase behavior of systems with ionic liquids: part IV. Binary system carbon
dioxideþ 1-hexyl-3-methylimidazolium tetrafluoroborate. J. Chem. Eng. Data
50, 52–55.
Esmaeilzadeh, F., As’adi, H., Lashkarbolooki, M., 2009. Calculation of the solid
solubilities in supercritical carbon dioxide using a new Gex mixing rule.
J. Supercrit. Fluids 51, 148–158.
Fry, S.E., Pienta, N.J., 1985. Effects of molten salts on reactions. Nucleophilic
aromatic substitution by halide ions in molten dodecyltributylphosphonium
Fig. 2. Experimental data (mentioned in Table 3 and supplementary materials) salts. J. Am. Chem. Soc. 107, 6399–6400.
versus the presented model results (mole fraction). Gharagheizi, F., 2007a. QSPR analysis for intrinsic viscosity of polymer solutions by
means of GA-MLR and RBFNN. Comput. Mater. Sci. 40, 159–167.
Gharagheizi, F.A., 2007b. New Neural Network quantitative structure–property
a reliable regression model, 24 mostly used ionic liquids were relationship for prediction of y (lower critical solution temperature) of
polymer solutions. e-Polymers. No. 114.
investigated. The critical temperature (Tc), the critical pressure (Pc), Gharagheizi, F., 2008. QSPR studies for solubility parameter by means of genetic
and the acentric factor (o) of ionic liquids along with the temperature algorithm-based multivariate linear regression and generalized regression
(T), and pressure (P) were treated as input parameters. The results neural network. QSAR Comb. Sci. 27, 165–170.
Gharagheizi, F., 2009a. Prediction of standard enthalpy of formation of pure
demonstrated that the model is capable of representing/predicting compounds using molecular structure. Aust. J. Chem. 62, 376–381.
the solubility of supercritical carbon dioxide in the investigated classic Gharagheizi, F., 2009b. A new group contribution-based method for estimation
ionic liquids with an average absolute deviation of 3.6% from of lower flammability limit of pure compounds. J. Hazard Mater. 170,
595–604.
experimental values. The prediction capability of the model has been Gharagheizi, F., 2009c. New Neural Network group contribution model for
also proven through prediction of solubility of carbon dioxide in ionic estimation of lower flammability limit temperature of pure compounds. Ind.
liquids from a randomly selected test data set. As the procedure of the Eng. Chem. Res. 48, 7406–7416.
Gharagheizi, F., Alamdari, R.F.A., 2008. Molecular-based model for prediction of
presented model is simple and easy to use, application of similar
solubility of C60. Fuller. Nanotub. Car. N 16, 40–57.
methods using the ANN is of interest in comparison with the Gharagheizi, F., Fazeli, A., 2008. Prediction of Watson characterization factor of
presented thermodynamic models in the literature. hydrocarbon compounds from their molecular properties. QSAR Comb. Sci. 27,
758–767.
Gharagheizi, F., Mehrpooya, M., 2008. Prediction of some important physical
properties of sulfur compounds using QSPR models. Mol. Divers 12, 143–155.
Acknowledgment Gharagheizi, F., Sattari, M., 2009. Estimation of molecular diffusivity of pure
chemicals in water: a quantitative structure-property relationship study. SAR
QSAR Environ. Res. 20, 267–285.
Ali Eslamimanesh wishes to thank MINES ParisTech for pro- Gharagheizi, F., Alamdari, R.F., Angaji, M.T.A., 2008. New Neural Network—group
viding a Ph.D. scholarship. contribution method for estimation of flash point. Energy Fuels 22,
1628–1635.
Gharagheizi, F., Tirandazi, B., Barzin, R., 2009. Estimation of aniline point
temperature of pure hydrocarbons: a quantitative structure–property rela-
Appendix A. Supplementary materials tionship approach. Ind. Eng. Chem. Res. 48, 1678–1682.
Gharagheizi, F., Eslamimanesh, A., Mohammadi, A.H., Richon, D., 2011a. Repre-
Supplementary data associated with this article can be found sentation/prediction of solubilities of pure compounds in water using
artificial neural network—group contribution method. J. Chem. Eng. Data 56,
in the online version at doi:10.1016/j.ces.2011.03.016. 720–726.
Gharagheizi, F., Eslamimanesh, A., Mohammadi, A.H., Richon, D., 2011b. Artificial
neural network modeling of solubilities of 21 mostly-used industrial solid
compounds in supercritical carbon dioxide. Ind. Eng. Chem. Res. 50, 221–226.
References Gharagheizi, F., Eslamimanesh, A., Mohammadi, A.H., Richon, D., 2011c. Determi-
nation of parachor of various compounds using an artificial neural network-
group contribution method. Ind. Eng. Chem. Res., in press. doi:10.1021/
Bennett, M.D., Leo, D.J., 2004. Ionic liquids as stable solvents for ionic polymer ie102464t.
transducers. Sens. Actuators A 115, 79–90. Gharagheizi, F., Eslamimanesh, A., Mohammadi, A.H., Richon, D., 2011d. Use of
Blanchard, L.A., Brennecke, J., 2001. Recovery of organic products from ionic artificial neural network-group contribution method to determine surface
liquids using supercritical carbon dioxide. Ind. Eng. Chem. Res. 40, 287–292. tension of pure compounds. Chem. Eng. Data., in press. doi:10.1021/
Blanchard, L.A., Hancu, D., Beckman, E.J., Brennecke, J.F., 1999. Green processing je2001045.
using ionic liquids and CO2. Nature 399, 28–29. Gharagheizi, F., Eslamimanesh, A., Mohammadi, A.H., Richon, D., 2011e. Determi-
Blanchard, L.A., Gu, Z., Brennecke, J., 2001. High-pressure phase behavior of ionic nation of critical properties and acentric factors of pure compounds using
liquid/CO2 systems. J. Phys. Chem. B 105, 2437–2444. artificial neural network-group contribution algorithm. Chem. Eng. Data., in
Boon, J.A., Levisky, J.A., Pflug, J.L., Wilkes, J.S., 1986. Friedel-crafts reactions in press. doi:10.1021/je200019g.
ambient-temperature molten salts. J. Org. Chem. 51, 480–483. Gharagheizi, F.A., 2010. Chemical structure-based model for estimation of upper
Breure, B., Bottini, S.B., Brignole, E.A., Peters, C.J., 2005. Thermodynamic modeling flammability limit of pure compounds. Energy Fuels 27, 3867–3871.
of the phase behavior of binary systems of ionic liquids and carbon dioxide Hagan, M., Demuth, H.B., Beale, M.H., 2002. Neural Network Design. International
with the GC–EoS. In: Proceedings of AIChE Annual Meeting Conference. Thomson publishing.
pp. 841–845. Housaindokht, M.R., Haghighi, B., Bozorgmehr, M.R., 2007. A comparison among
Carvalho, P.J., Álvarez, V.H., Machado, J.J.B., Pauly, J., Daridon, J.L., Marrucho, I.M., three equations of state in predicting the solubility of some solids in super-
Aznar, M., Coutinho, J.A.P., 2009a. High pressure phase behavior of carbon critical carbon dioxide. Korean J. Chem. Eng. 24/1, 102–105.
3044 A. Eslamimanesh et al. / Chemical Engineering Science 66 (2011) 3039–3044

Hu, Y., Liu, H.L., Prausnitz, J.M., 1996. Equation of state for fluids containing Peng, D.Y., Robinson, D.B.A., 1976. New two-constant equation of state. Ind. Eng.
chainlike molecules. J. Chem. Phys. 104, 396–404. Chem. Fundam. 15/1, 59–64.
Huang, C., Tang, M., Tao, W., Chen, Y., 2001. Calculation of the solid solubilities in Piazza, L., Scalabrin, G., Marchi, P., Richon, D., 2006. Enhancement of the extended
supercritical carbon dioxide using a modified mixing model. Fluid Phase corresponding states techniques for thermodynamic modelling. I. Pure fluids.
Equilib. 179, 67–84. Int. J. Refrig. 29, 1182–1194.
Ismadji, S., 2008. Solubility of methyl salicylate in supercritical carbon dioxide at Raeissi, S., Peters, C.J., 2009. Carbon dioxide solubility in the homologous 1-alkyl-
several temperatures. J. Chem. Eng. Data 53, 2207–2210. 3-methylimidazolium bis (trifluoromethylsulfonyl) imide family. J. Chem. Eng.
Joback, K.K., Reid, R., 1987. Estimation of pure component properties from group Data 54, 382–386.
contribution. Chem. Eng. Commun. 57, 233–247. Redlich, O., Kwong, J.N.S., 1949. On the thermodynamics of solutions. V. An
Kalogirou, S.A., 2001. Artificial Neural Networks in renewable energy systems equation of state. Fugacities of gaseous solutions. Chem. Rev. 44, 233–244.
applications: a review. Renew. Sust. Energy Rev. 5, 373–401. Ren, W., Sensenich, B., Scurto, A.M., 2010. High-pressure phase equilibria of
Karakatsani, E.K., Economou, I.G., 2006. Perturbed chain-statistical associating {carbon dioxide (CO2)þ n-alkyl-imidazolium bis(trifluoromethylsulfonyl)a-
fluid theory extended to dipolar and quadrupolar molecular fluids. J. Phys. mide} ionic liquids. J. Chem. Thermodyn. 42, 305–311.
Chem. B 110, 9252–9261. Sattari, M., Gharagheizi, F., 2008. Prediction of molecular diffusivity of pure
Kim, Y.S., Choi, W.Y., Jang, J.H., Yoo, K.P., Lee, C.S., 2005. Solubility measurement components into air: a QSPR approach. Chemosphere 72, 1298–1302.
and prediction of carbon dioxide in ionic liquids. Fluid Phase Equilib. 228–229, Scalabrin, G., Marchi, P., Bettio, L., Richon, D., 2006. Enhancement of the extended
439–445. corresponding states techniques for thermodynamic modelling. II. Mixtures.
Kroon, M.C., Karakatsani, E.K., Economou, I.G., Witkamp, G.J., Peters, C.J., 2006. Int. J. Refrig. 29, 1195–1207.
Modeling of the carbon dioxide solubility in imidazolium-based ionic liquids Schilderman, A.M., Raeissi, S., Peters, C.J., 2007. Solubility of carbon dioxide in the
with the tPC-SAFT equation of state. J. Phys. Chem. B 110, 9262–9269. ionic liquid 1- ethyl-3-methylimidazolium bis(trifluoromethylsulfonyl) imide.
Kurnia, K.A., Harris, F., Wilfred, C.D., Abdul Mutalib, M.I., Murugesan, T., 2009. Fluid Phase Equilib. 260, 19–22.
Thermodynamic properties of CO2 absorption in hydroxyl ammonium ionic Shariati, A., Peters, C.J., 2004a. High-pressure phase behavior of systems with ionic
liquids at pressure of (100–1600) KPa. J. Chem. Thermodyn. 41, 1069–1073. liquids: II. The binary system carbon dioxide þ 1-ethyl-3-methylimidazolium
Liu, H.L., Hu, Y., 1996. Molecular thermodynamic theory for polymer systems part hexafluorophosphate. J. Supercrit. Fluid 29, 43–48.
II. Equation of state for chain fluids. Fluid Phase Equilib. 122, 75–97. Shariati, A., Peters, C.J., 2004b. High-pressure phase behavior of systems with ionic
Liu, H.L., Hu, Y., 1998. Equation of state for systems containing chainlike liquids: III. The binary system carbon dioxide þ 1-ethyl-3-methylimidazolium
molecules. Ind. Eng. Chem. Res. 37, 3058–3066. hexafluorophosphate. J. Supercrit. Fluids 30, 139–144.
Lydersen, A.L., 1955. Estimation of Critical Properties of Organic Compounds. Sheldon, R., 2001. Catalytic reactions in ionic liquids. Chem. Commun., 2399–2407.
Shiflett, M.B., Yokozeki, A., 2005. Solubilities and diffusivities of carbon dioxide
Report 3. Engineering Experimental Station, College of Engineering, University
in ionic liquids: [bmim][PF6] and [bmim][BF4]. Ind. Eng. Chem. Res. 44,
of Wisconsin, Madison, WI.
4453–4464.
Machida, H., Taguchi, R., Sato, Y., Florusse, L.J., Peters, C.J., Smith, R.L., 2009.
Shiflett, M.B., Yokozeki, A., 2009. Phase behavior of carbon dioxide in ionic liquids:
Measurement and correlation of supercritical CO2 and ionic liquid systems for
[emim][Acetate], [emim] [Trifluoroacetate] and [emim][Acetate] þ [emim]
design of advanced unit operation. Front. Chem. Eng. China 3/1, 12–19.
[Trifluoroacetate] mixtures. J. Chem. Eng. Data 54, 108–114.
Mehrpooya, M., Gharagheizi, F.A., 2010. Molecular approach for prediction of
Shin, E.K., Lee, B.C., 2008. High-pressure phase behavior of carbon dioxide
sulfur compounds solubility parameters, phosphorus sulfur and silicon and
with ionic liquids: 1-alkyl-3-methylimidazolium trifluoromethonesulfonate.
related elements. Phosphorus Sulfur 185, 204–210.
J. Chem. Eng. Data 53, 2728–2734.
Mehrpooya, M., Mohammadi, A.H., Richon, D., 2010. Extension of an Artificial
Shin, E.K., Lee, B.C., Lim, J.S., 2008. High-pressure solubilities of carbon dioxide in
Neural Network algorithm for estimating sulfur content of sour gases at
ionic liquids: 1-alkyl-3-methylimidazolium bis (trifluoromethylsulfonyl)
elevated temperatures and pressures. Ind. Eng. Chem. Res. 49, 439–442.
imide. J. Supercrit. Fluid 45, 282–292.
Mohammadi, A.H., Richon, D., 2007. Use of artificial neural networks for estimat-
Song, H.N., Lee, B.C., Lim, J.S., 2010. Measurement of CO2 solubility in ionic liquids:
ing water content of natural gases. Ind. Eng. Chem. Res. 46, 1431–1438.
[bmp][tfo] and [p14,6,6,6][tf2n] by measuring bubble-point pressure. J. Chem.
Mohammadi, A.H., Richon, D., 2008a. A Mathematical model based on artificial
Eng. Data 55, 891–896.
neural network technique for estimating liquid water  hydrate equilibrium Stefani, V.D., Baba-Ahmed, A., Richon, D., 2004. A review of experimental methods
of water  hydrocarbon system. Ind. Eng. Chem. Res. 47, 4966–4970. for solid solubility determination in cryogenic systems. Cryogenics 44/9,
Mohammadi, A.H., Richon, D., 2008b. Estimating sulfur content of hydrogen 631–641.
sulfide at elevated temperatures and pressures using an artificial neural Vatani, A., Mehrpooya, M., Gharagheizi, F., 2007. Prediction of standard enthalpy of
network algorithm. Ind. Eng. Chem. Res. 47, 8499–8504. formation by a QSPR model. Int. J. Mol. Sci. 8, 407–432.
Mohammadi, A.H., Richon, D., 2010. Hydrate phase equilibria for hydrogen þwater Wang, T., Peng, C., Liu, H., Hu, Y., 2006. Description of the pVT behavior of ionic
and hydrogen þtetrahydrofuran þwater systems: predictions of dissociation liquids and the solubility of gases in ionic liquids using an equation of state.
conditions using an artificial neural network algorithm. Chem. Eng. Sci. 65, Fluid Phase Equilib. 250, 150–157.
3352–3355. Wong, H., Han, S., Livingston, A.G., 2006. The effect of ionic liquids on product yield
Mohammadi, A.H., Afzal, W., Richon, D., 2008. Determination of critical properties and catalyst stability. Chem. Eng. Sci. 61, 1338–1341.
and acentric factors of petroleum fractions using artificial neural networks. Yamini, Y., Arab, J., Asghari-khiavi, M., 2003. Solubilities of phenazopyridine,
Ind. Eng. Chem. Res. 47, 3225–3232. propranolol, and methimazole in supercritical carbon dioxide. J. Pharm.
Mohammadi, A.H., Martı́nez-López, J.F., Richon, D., 2010a. Determining phase Biomed. Anal. 32, 181–187.
diagrams of tetrahydrofuran þmethane, carbon dioxide or nitrogen clathrate Yokozeki, A., Shiflett, M.B., 2006. Global phase behaviors of trifluoromethane in
hydrates using an artificial neural network algorithm. Chem. Eng. Sci. 65, ionic liquid [bmim][pf6]. AIChE J. 52/11, 3952–3957.
6059–6063. Yokozeki, A., Shiflett, M.B., Junk, C.P., Grieco, L.M., Foo, T., 2008. Physical and
Mohammadi, A.H., Belandria, V., Richon, D., 2010b. Use of an artificial neural chemical absorptions of carbon dioxide in room-temperature ionic liquids.
network algorithm to predict hydrate dissociation conditions for hydrogen þ J. Phys. Chem. B 112, 16654–16663.
water and hydrogen þ tetra-n-butyl ammonium bromideþ water systems. Zein El Abedin, S., Po€lleth, M., Meiss, S.A., Janek, J., Endres, F., 2007. Ionic liquids as
Chem. Eng. Sci 65, 4302–4305. green electrolytes for the electrodeposition of nonmaterials. Green Chem. 9,
Oh, D.J., Lee, B.C., 2006. High-pressure phase behavior of carbon dioxide in 549–553.
ionic liquid 1-butyl-3-methylimidazolium bis (trifluoromethylsulfonyl) imide. Zhao, H., 2006. Innovative applications of ionic liquids as green engineering
Korean J. Chem. Eng. 23/5, 800–805. liquids. Chem. Eng. Commun. 193, 1660–1677.