Chemometrics and Intelligent Laboratory Systems 104 (2010) 195–204

Contents lists available at ScienceDirect

Chemometrics and Intelligent Laboratory Systems

j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / c h e m o l a b

A neural network-based method for estimation of binary gas diffusivity

R. Eslamloueyan ⁎, M.H. Khademi
School of Chemical and Petroleum Engineering, Shiraz University, Mollasadra Avenue, Shiraz, Iran

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

Article history: In this study, a feedforward three-layer neural network is developed to predict binary diffusion coefficient
Received 20 June 2010 (DAB) of gases at atmospheric pressure over a wide range of temperatures based on the critical temperature
Received in revised form 16 August 2010 (Tc), critical volume (Vc) and molecular weight (M) of each component in the binary mixture. The accuracy of
Accepted 16 August 2010
the method is evaluated through a test data set not used in the training stage of the network. Furthermore,
Available online 20 August 2010
the performance of the neural network model is compared with that of well known correlations suggested in
Keywords:
the literature. The results of this comparison show that our developed method outperforms other
Neural network correlations, with respect to accuracy as well as extrapolation capabilities.
Binary diffusion coefficient © 2010 Elsevier B.V. All rights reserved.
Gas diffusivity

1. Introduction (ii) The Chapman–Enskog equation [1]:


Estimation of diffusion coefficients in the liquid and gas phases is 0:00263T 3 = 2 1 =MA + 1=MB 1 = 2
DAB = ð2Þ
crucial in the design and simulation of processes involving mass Pσ 2AB 2
transfer such as liquid–liquid extraction, leaching, gas absorption,
reactors, packed and tray towers. For instance, binary diffusion where P is the gas pressure in atm.
coefficients of gases are needed in the design of gas phase reactors. (iii) The Gilliland equation [2]:
Some other fields where gas diffusivity plays an important role are
global changes, atmospheric chemistry, combustion science, studies 0:0043T 3 = 2 ð1 =MA + 1=MB Þ1 = 2
DAB =   ð3Þ
of indoor air pollution and atmosphere–biosphere interactions. 1=3 1=3 2
P VA + VB
Diffusion is also a major factor in peak broadening in chromatography.
Therefore, development of accurate and reliable correlations for
where VA and VB are molar volumes in cm3 at the boiling
determination of diffusion coefficients has drawn the attention of
points, which can be obtained experimentally, or they can be
many researchers. Availability of well known theories for the gas
estimated through group contribution methods.
phase such as kinetic theory of gases has motivated scientists to
(iv) The Arnold equation [3]:
propose several theoretical and semi-empirical equations for estima-
tion of binary diffusivity coefficients of gases. Some of these 0:0083T 3 = 2 ð1 =MA + 1=MB Þ1 = 2
correlations have been presented in the following: DAB =   ð4Þ
1=3 1=3 2
P VA + VB ð1 + SAB = T Þ
(i) The Stefan–Maxwell (SM) equation [1]:

1 = 2 where SAB is Sutherland's constant, which can be estimated by
a 8RT 1 1
DAB = + ð1Þ various ways. The above equation, which introduces a second
nσ 2AB π MA MB
temperature dependent term in the denominator to account for
molecular “softness”, shows a dependence varying from T 3/2 to
where a is a constant taking various values (1/3π, 1/8, 1/2π and
T 5/2.
3/32) based on different researcher studies, n is the number of
(v) The Hirschfelder–Bird–Spotz (HBS) equation [4]:
gas phase molecules per cm3, σAB is the collision diameter
between gas molecules A and B, R is the gas constant, T is the 3=2 1=2
0:00186T ð1=MA + 1 =MB Þ
absolute temperature and MA and MB are the molecular masses DAB = ð5Þ
Pσ 2AB ΩAB
of solute A and carrier gas B, respectively.

The term ΩAB is the collision integral depending in a

⁎ Corresponding author. Fax: + 98 711 6287294. complicated way on temperature and the interaction energy
E-mail address: eslamlo@shirazu.ac.ir (R. Eslamloueyan). of the colliding molecules, AB. Hirschfelder et al. [5] followed

0169-7439/$ – see front matter © 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.chemolab.2010.08.009
196 R. Eslamloueyan, M.H. Khademi / Chemometrics and Intelligent Laboratory Systems 104 (2010) 195–204
the Chapman–Enskog kinetic approach combined with the
Lennard–Jones intermolecular potential function. ΩAB values as
a function of the reduced temperature T ⁎ = kT/AB, where k is
the Boltzmann constant, can be obtained from relevant
tables [5,6]. The main disadvantage of the HBS equation is the
difficulty encountered in evaluating σAB and ΩAB. There are
correlations for estimation of these parameters from viscosity
measurements.
(vi) Chen and Othmer [7] provided the most explicit approximation
of the HBS equation using the critical values of temperature TC,
and volume, VC:
0:43ðT =100Þ1:81 ð1=MA + 1 =MB Þ1 = 2
DAB =  0:1405
2 ð6Þ [12] developed a neural network consisting of two consecutive
P TCA TCB =104 ðVCA =100Þ0:4 + ðVCB =100Þ0:4
Both TC and VC values can be estimated in various ways [1,5].
(vii) Fuller–Schettler and Giddings (FSG) [8] developed a successful
equation in which atomic and structural volume increments
and other parameters were obtained by a least-squares fit to
over 300 measured data. In the Fuller et al. [8] method:
0:00143T 1:75 ð1 =MA + 1=MB Þ1 = 2
DAB = pffiffiffi h i ð7Þ
1=3 1=3 2
2P ð∑υÞA + ð∑υÞB
Σv is determined by summing the atomic contributions.
(viii) Huang et al. [9] investigated the effects of pressure and
temperature on the gas diffusivity. Based on their experimental
data they modified the Arnold3 equation as follows:
5:06T 1:75 ð1 =MA + 1=MB Þ1 = 2
DAB =   ð8Þ
1=3 1=3 2
P 1:286 VA + VB
They compared their correlation (Eq. (8)) with most of the other
techniques. In this comparison, they used 134 experimental diffusion
coefficient values obtained from literature. The results of their work
are summarized in Table 1. To achieve higher precision in estimating
gaseous diffusion coefficients, more complicated methods derived
from HBS equation should be used. These methods, which are based
on detailed gas dynamics, are commonly recommended for unusual
systems (e.g. large molecules, high temperatures, etc.) which are
beyond the valid range of Gilliland and Arnold equations. The Fuller–
Schettler–Giddings (FSG) [8] equation provides the best practical
combination of simplicity and accuracy.
For binary gas mixtures at low pressure, DAB is inversely
proportional to the pressure, increases with increasing temperature,
and is almost independent of composition for a given gas-pair. For
most binary mixtures, diffusivity data obtained from experiments and
correlations are quite limited in range and accuracy. Fuller and
Giddings [10] compared the existing theoretical or empirical
equations for predicting gaseous diffusion coefficients. Using exper-
imental data for 38 binary gas systems, they examined the prediction
accuracy of various proposed correlations, and tabulated the percent-
age errors for each method. Based on their work, the average absolute
Table 1
Overall average relative errors of various methods used for the prediction of binary
gaseous diffusion coefficients for 134 experimental test data [9].
Method Accuracy (%)
Gilliland [2] 6.64
Arnold [3] 11.75
Hirschfelder–Bird–Spotz [4] 18.99
Chen–Othmer [7] 10.85
Fuller–Schettler–Giddings [8] 3.40
Huang–Young–Huang–Kuo [9] 3.51
percentage errors varied from 4.2 to 20%, depending on the estimation
method, and the proposed correlation by Fuller et al. [8] was the most
accurate. They also showed the difficulties encountered for theoretical
estimation of diffusion coefficients. To overcome these difficulties
artificial neural networks can be used as a powerful multivariable
function.
Neural networks have been used extensively to predict the
physical properties in various fields of chemical engineering over
the last two decades. Eslamloueyan and Khademi [11] proposed a
feedforward three-layer neural network to predict the thermal
conductivity of pure gases at atmospheric pressure over a wide
range of temperatures based on their critical temperatures, critical
pressures and molecular weights. Also, Eslamloueyan and Khademi
multilayer perceptrons (MLP) to estimate the conductivities of binary
gaseous mixtures at atmospheric pressure. Kauffman and Jurs [13]
predicted the surface tension, viscosity, and thermal conductivity for
common organic solvents by using a quantitative structure. They used
multiple linear regression methods and computational neural net-
works to train and evaluate their models based on statistical indices
and overall root-mean square error. Boozarjomehri et al. [14]
characterized the basic properties of pure substances and petroleum
fractions by a neural network.
Although different theoretical and semi-empirical correlations
have been suggested for estimation of diffusivity coefficients by
researchers, there are still considerable differences among their
predictions. In the correlations presented by other researchers,
some molecular parameters are used that have not been determined
experimentally for all substances, and they are usually estimated by
auxiliary equations whose accuracy depends on the operating
conditions and the type of substances. On the other hand, calculation
of these parameters is usually cumbersome and in some cases
different equations have been suggested which do not necessarily
lead to the same results. Furthermore, the existence of various
methods for estimating the gas diffusivity can add more difficulties in
the selection of the best prediction scheme. The objective of our work
is to develop a neural network model for prediction of binary diffusion
coefficient of gases at atmospheric pressure over a wide range of
temperatures and substances. The network inputs are the gas
temperature, critical temperature, critical volume and molecular
weight of each compound. In the next section, the proposed method
as well as the data used in its development and validation is
presented. Section 3, contains the results and discussion about the
proposed method.
2. Methodology
The inherent ability of artificial neural networks to learn and
recognize nonlinear and complex relationships can be used to predict
binary diffusion coefficient of gases. The development steps of the
proposed neuromorphic model for prediction of gas diffusivity have
been explained in the following sections. A brief overview of the
artificial neural network used in this study has been presented by
Eslamloueyan and Khademi [11,12].
2.1. Data acquisition and analysis
One of the most important stages in the development of the
suggested neuromorphic model is the availability of reliable exper-
imental data for the model training. These data include experimental
published data on critical temperature, critical volume and molecular
weight of the pure components as well as binary diffusion coefficients
at different temperatures. In this study, 467 experimental data were
collected from different literatures. These data were divided into two
parts; training data set and test data set. More than two-thirds of the
total data were chosen for training to set up the ANN and the other
 R. Eslamloueyan, M.H. Khademi / Chemometrics and Intelligent Laboratory Systems 104 (2010) 195–204
remains as the test data set. The training data set were selected so that
it could cover all the temperature range of the experimental data. A set
of 336 experimental data points for binary diffusion coefficient of both
organic and inorganic gases that has been used in this study is presented in
Table 2. In addition, some of the binary systems have not been used for the
training step of the network (e.g. H2O–H2, n-C4H10–H2, 3-methylheptane-
He, 2,3-Dimethylhexane-He, 2,2,4-trimethylpentane-He, C2H4F2–He,
Fluorobenzene–He, CHCl3–He, 1-Chlorobutane-He, 1-Bromobutane-He,
1-Iodopropane-He, CO–He, iso-C4H10–N2, 1-Pentanol-N2, 1-Heptanol-N2,
CO2–Ar, Xe–Ar, Benzene–O2, C2H6–CH4, n-C4H10–CH4, iso-C4H10–C2H6,
H2O–H2, n-C4H10–H2, 3-methylheptane-He, 2,3-Dimethylhexane-He,
2,2,4-trimethylpentane-He, Fluorobenzene–He, CHCl3–He, 1-Chlorobu-
Table 2
List of binary system A–B used in the development of the neuromorphic model.
No. Binary system A–B (no. of data points) References
1 CH4–He (1) [10]
CO–D2 (1), n-C6H14–H2 (1), n-C7H16–H2 (1), n-C10H22–H2 (1),
2 [15]
CO–H2 (1), O2–H2 (1), SO2–H2 (1), NH3–H2 (1), Ar–H2 (1), Xe–H2
(1), cyclohexane–H2 (1), Benzene–H2 (1), Acetone–H2 (1), NH3–
He (1), H2O–He (1), O2–H2O (1), N2–H2O (3), C2H4–H2O (1), Ar–
Ne (1), Kr–Ne (1), Ar–N2 (1), SO2–N2 (1), n-C6H14–N2 (1), n-
C8H18–N2 (1), 2,2,4-trimethylpentane-N2 (1), n-C10H22–N2 (1),
benzene–N2 (1), C3H8–N2O (1), Ar–NH3 (1), CO2–O2 (1), CCl4–O2
(1), Cyclohexane–O2 (1), n-C6H14–O2 (1), n-C8H18–O2 (1), 2,2,4-
trimethylpantane-O2 (1), Kr–Ar (1), SO2–Ar (1), N2–CO (1), Ar–
CO (1), CO2–CO (1), C3H8–CO2 (1), N2O–CO2 (1), SO2–CO2 (1),
Ar–CH4 (1), H2O–CH4 (1), n-C6H14–C2H6 (1)
3 CH4–H2 (1), C3H8–H2 (1), N2–H2 (1), N2–He (5), O2–He (3), Ar– [16]
He (3), Kr–He (3), C2H6–He (5), C2H4–He (5), C3H6–He (4),
C2H6–N2 (6), C3H8–N2 (1), C2H4–N2 (4), C3H6–N2 (4), CO2–CH4
(1)
4 C2H6–H2 (1), C3H8–H2 (1), iso-C4H10–H2 (2), neo-C5H12–H2 (2), [17]
n-C4H10–N2 (1), CH3Cl–CH4 (3), C2H5Cl–CH4 (3), C3H8–CH4 (3),
Iso-C4H10–CH4 (3), neo-C5H12–CH4 (3), C3H8–C2H6 (2), n-C4H10–
C2H6 (3), neo-C5H12–C2H6 (3), n-C4H10–C3H8 (2), neo-C5H12–
C3H8 (2), neo-C5H12-n-C4H10 (3), neo-C5H12-iso-C4H10 (2),
C2H5Cl–CH3Cl (2)
5 n-C5H12–H2 (3), n-C6H14–H2 (3), CH3OH–H2 (1), C2H5OH–H2 [18]
(3), 1-Butanol-H2 (3), 2-Butanol-H2 (4), cyclohexane–H2 (3),
Benzene–H2 (3), Toluene–H2 (3)
6 CO2–H2 (1), N2–H2 (1), Ar–He (2) [19]
7 N2–H2 (1), O2–N2 (1) [20]
8 N2–He (1), Ar–He (1) [21]
9 N2–He (4), O2–He (1), Ar–He (2), CH3OH–He (1), C2H5OH–He [22]
(4), 1-Propanol-He (4), 2-Propanol-He (4), 1-Butanol-He (4), 1-
Pentanol-He (4), 1-Hexanol-He (4), Benzene–He (5), CO2–He
(4)
10 N2–He (1) [23]
11 N2–He (6), CO2–He (4), CO2–H2O (2), CO2–N2 (5), CHCl3–N2 (3), [24]
2-Propanol-N2 (1), Acetone–N2 (2), Methyl acetate–N2 (2),
Methyl formate–N2 (1), Ethyl formate–N2 (2), cyclohexane–N2
(2), benzene–N2 (3), 2-Propanol-CO2 (3), Methyl acetate-CO2
(2), Ethyl formate-CO2 (2), Cyclohexane–CO2 (3), Benzene-CO2
(3), CCl4-CO2 (3), CHCl3-CO2 (2)
12 CH4–He (1) [25]
13 n-C4H10–He (3), n-C5H12–He (3), n-C6H14–He (4), Ether–He (3), [26]
Acetone–He (3), Benzene–He (3), n-C4H10–Ar (3), n-C5H12–Ar
(3), n-C6H14–Ar (4), Benzene–Ar (3), Acetone–Ar (3), Ether–Ar
(3)
14 n-C8H18–He (1), 2,4-Dimethylhexane-He (1), 3-Ethylhexane-He [27]
(1), 3-Ethyl-2-methyl-pentane-He (1)
15 CH2F2–He (1), C6F6–He (1), C2H4Cl2–He (1), 1-Chloropropane- [28]
He (1), 2-Chlorobutane-He (1), 1-Chloropentane-He (1),
Chlorobenzene-He (1), Bromoethane-He (1), 2-Bromobutane-
He (1), 2-Iodopropane-He (1)
16 O2–N2 (1), [29]
17 n-C4H10–N2 (1), C2H4–N2 (1) [30]
18 n-C5H12-N2 (1), CH3OH–N2 (1), 2-Propanol-N2 (1), Acetone–N2 [31]
(1), Methyl acetate–N2 (1), Ethyl acetate–N2 (1), benzene–N2
(1)
19 C2H4–N2 (1) [32]
20 CHCl3–N2 (1),1-Propanol-N2 (1), 1-Butanol-N2 (1), 1-Hexanol- [33]
N2 (1), 1-Octanol-N2 (1), Acetone–N2 (1), Methyl isobutyl
ketone-N2 (1), Chlorobenzene–N2 (1),
197
tane-He, 1-Bromobutane-He, 1-Iodopropane-He, CO–He, iso-C4H10–N2,
1-Pentanol-N2, 1-Heptanol-N2, CO2–Ar, Xe–Ar, Benzene–O2, C2H6–CH4,
C2H6–CH4, n-C4H10–CH4, iso-C4H10–C2H6). These data were used in the
test data set to investigate the extrapolation capability of the developed
ANN model. Table 3 summarizes the overall range of experimental data
points employed in the development of the proposed multilayer
perceptron (MLP) network. We have not used the diffusivity data given
in some publications that are based on the estimation correlations, and all
data points employed in this study are experimental data. Superscripts A
and B stand for heavy and light components, respectively.
After identifying and collecting the data set, the next step is the
selection of input variables, which are the model's independent
variables. The available correlations for prediction of binary diffusion
coefficient at constant pressure are essentially based on the
assumption that gas diffusivity at atmospheric pressure can be
described as a function of temperature, critical temperature, critical
volume and molecular weight of each compound as follows:
 
A B A B A B
DAB = f T; Tc ; Tc ; Vc ; Vc ; M ; M ð9Þ
According to Eq. (9), we chose temperature, critical temperature,
critical volume and molecular weight of each compound, respectively,
as the inputs of the proposed MLP network for prediction of binary
diffusion coefficient of gases. Table 4 summarizes the input data set
used in the development of the neuromorphic model for various
components.
2.2. Neural network training
Once the input data are determined, the design of the artificial
neural network architecture is begun. In this work, a network with
only one hidden layer has been selected as the starting network
structure. The neural network toolbox of MATLAB programming
language has been used to design the feedforward MLP network. In
order to find the number of neurons in the hidden layer a constructive
approach has been used. Based on the constructive approach, a small
number of neurons are used in the hidden layer and if the error of the
trained network does not meet the desired tolerance the number of
neurons in the hidden layer is increased by one and training cycle and
performance evaluation is repeated [35]. This procedure is continued
until the trained network performs satisfactorily (i.e., its training,
validation and testing error are lower than the target goal).
Applying the constructive approach to design the neural network
model for the estimation of gases diffusivity led to a three-layer
network with seven neurons in the input layer, ten neurons in the
hidden layer and one neuron at the output layer. The type of neurons’
transfer functions used in this study is tansig function which is
defined as the following,
expðxÞ− expð−xÞ
f ðxÞ = ð10Þ
expðxÞ+ expð−xÞ
The MLP network has been trained by Levenberg Marquardt (LM)
algorithm [36–38]. The learning rate of LM method can be adjusted by
Table 3
Range of data of binary gases properties.
Property Minimum Maximum
T/K 77.2 523
TAc /K 126.2 652.5
VAc /cm3 mol−1 55.95 624
MA 16.043 186.056
TBc /K 5.19 647.14
VBc /cm3 mol−1 41.7 262.7
MB 2.016 58.123
DAB/cm2 s−1 0.0368 1.077
198 R. Eslamloueyan, M.H. Khademi / Chemometrics and Intelligent Laboratory Systems 104 (2010) 195–204

Table 4 an adaptive parameter which is shown by μ. The algorithm starts with

Inputs of the ANN model for all compound types. a small value of μ (0.001). If the performance function does not
No Component type Tc/K Vc/cm3 mol−1 M decreased in a step that step is repeated by increasing the adaptive
parameter through its multiplying to a coefficient greater than 1 (e.g.
1 CH4 190.56 98.6 16.043
2 CH3Cl 416.2 143 50.488 10). On the other hand, if the performance function decreases in a step
3 C2H5Cl 460.3 199 64.514 , then the adaptive parameter (μ) is multiplied to a factor less than 1
4 N2 126.2 90.1 28.014 (e.g. 0.1), so that the LM algorithm will tend to Gauss–Newton
5 n-C4H10 425.12 255 58.123
method which have a faster convergence rate. The performance
6 H2 33.25 64.2 2.016
7 C2H6 305.32 145.5 30.07 function used for LM algorithm is mean square errors (MSE). Some
8 C3H8 369.83 200 44.097 other training parameters are, maximum number of epochs to
9 i-C4H10 407.85 262.7 58.123 train = 500, performance goal = 10−5 and minimum performance
10 Neo-C5H12 433.75 303.2 72.15 gradient = 10−10.
11 H2O 647.14 55.95 18.015
12 CO 132.85 93.1 28.01
13 D2 38.35 60.2 4.028 2.3. Selection of optimal configuration
14 He 5.19 57.3 4.003
15 Ar 150.86 74.57 39.948 The performance of a trained network can be measured to some
16 CO2 304.12 94.07 44.01
extent by the errors on the training, validation and test data sets, but it
17 O2 154.58 73.37 31.999
18 N2O 309.6 97 44.013 is often useful to investigate the network response in more detail.
19 SO2 430.8 122 64.065 Regression analysis has been applied to assess the network capability
20 Xe 289.74 118 131.29 for binary diffusion coefficient prediction. The coefficient of determi-
21 NH3 405.4 72.47 17.031 nation, R2, has been used as a measure to evaluate how the trained
22 Acetone 508.1 209 58.079
network estimation is correlated to the experimental data.
23 n-Hexane 507.6 368 86.177
24 Cyclohexane 553.5 308 84.161 Also, different neural network topologies have been compared
25 Benzene 562.05 256 78.114 using their mean relative errors (MRE) and mean square errors (MSE).
26 n-Heptane 540.2 428 182.59 The MRE and MSE are defined with the following equations:
27 n-Decane 617.7 624 142.285
28 C2H4 282.34 131.1 28.054
1 N jDAB −DAB j
exp Pred
29 n-Octane 568.7 492 114.231
30 CCl4 556.3 276 153.822
MRE = ∑ exp ð11Þ
N i=1 DAB
31 2,2,4-Trimetylepentane 543.9 469.7 114.231
32 Ne 44.4 41.7 20.18
33 Kr 209.4 91.2 83.8 1 N  exp 
Pred 2
34 CH3OH 512.64 118 32.042 MSE = ∑ DAB −DAB ð12Þ
N i=1
35 1-Butanol 563.05 275 74.123
36 2-Butanol 536.05 269 74.123
37 C5H12 469.7 311 72.15 N is the number of data points, and Dexp
AB and DAB
Pred
are the
38 Toluene 591.75 316 92.141 experimental and predicted values of binary diffusion coefficient,
39 C2H5OH 513.92 167 46.0169
respectively.
40 3-methyleheptane 563.6 471.1 114.231
41 2,4-dimethylehexane 553.5 472 114.231
42 2,3-dimethylehexane 563.4 468.2 114.231 3. Results and discussion
43 3-ethylehexane 565.4 460.5 114.231
44 3-ethyle-2-methylepentane 567 445.3 114.231
The results of the iterative constructive method for the determi-
45 2,2,4-trimethylepentane 543.9 469.7 114.231
46 1-propanol 536.78 219 60.096 nation of the optimum number of neurons in the hidden layer of the
47 2-propanol 508.3 220 60.096 proposed MLP network has been presented in Table 5. In this table,
48 1-pentanol 588.15 326 88.15 MRE, MSE and R values of the trained networks have been shown for
49 1-hexanol 611.4 381 102.177 different number of hidden layer neurons. The configuration with
50 Ether 400 190 46.068
minimum error measures (i.e., MRE and MSE), and appropriate R-
51 CH2F2 351.26 121 52.024
52 C2H4F2 386.41 181 66.051 value, is selected as the best network architecture. According to
53 Fluorobenzene 560.09 269 96.103 Table 5, the best neural network configuration has one hidden layer
54 C6F6 516.73 335 186.056 with ten neurons. The parameters (weight and bias values) of the best
55 CHCl3 536.5 240 119.377
selected MLP network have been given in Table 6.
56 1,1-dicholoroethane 523 236 98.959
57 1-choloropropane 503.1 254 78.541
Fig. 1 illustrates the correlation between the simulation results of
58 1-cholorobutane 542 312 92.56 the developed neural network and the experimental training data
59 2-cholorobutane 520.6 312 92.568 points. The perfect fit (output equal to targets) is indicated by the
60 1-choloropentane 552 364.5 106.595 solid line. The close proximity of the best linear fit to the perfect fit, as
61 Cholorobenzene 632.4 308 112.558
observed in Fig. 1, shows a good correlation among the network
62 Boromoethane 503.8 214.9 108.966
63 1-boromobutane 569.5 321.5 137.019 predictions and the experimental data.
64 2-boromobutane 558.7 315.5 137.019
65 1-Iodopropane 589.4 291.5 169.99
66 2-Iodopropane 574.6 285.5 169.99 Table 5
67 1-heptanol 631.9 435 116.203 MRE, MSE and R values for different neural network configurations.
68 1-Octanol 652.5 490 130.23
69 Methyl acetate 506.8 228 74.079 No. of neurons MRE MSE R-value
70 Ethyl formate 508.5 229 74.079
5 5.66 1.4 × 10−3 0.9938
71 Ethyl acetate 523.2 286 88.106
8 5.80 7.7 × 10−4 0.9930
72 Methyl isobutyle ketone 574.6 340.6 100.161
10 4.47 8.4 × 10−4 0.9949
73 Cholorobenzene 632.4 308 112.558
12 5.52 9.3 × 10−4 0.9945
74 C3H6 364.9 184.6 42.081
15 4.94 8.3 × 10−4 0.9950
Source: Data from Reid et al. [34]. 20 5.83 1.3 × 10−3 0.9925
 R. Eslamloueyan, M.H. Khademi / Chemometrics and Intelligent Laboratory Systems 104 (2010) 195–204 199

Table 6
Parameters (weight and bias) of the ANN.

Hidden layer Output layer

Weights Biases Weights Bias

Neuron T TAc TBc VAc VBc M A

M B
DAB

1 −0.0093 0.0988 −0.4495 −0.052 0.2689 0.0031 −0.2705 −3.400 1.3096 −0.096
2 0.0089 0.0294 0.0013 −0.0224 0.0006 0.0029 −0.0147 4.7031 0.16
3 −0.0029 0.0009 0.0016 0.0063 0.0157 0.0021 0.0736 −2.693 −0.5135
4 0.004 0.0069 0.0926 0.0065 −0.0586 0.0025 0.1483 −1.939 −0.5212
5 −0.018 −0.0172 −0.0059 0.0835 −0.1139 −0.0598 0.9301 −2.462 0.0526
6 −0.1351 −0.003 0.0185 0.266 0.2725 0.26 −0.0983 2.2892 −0.5185
7 0.0673 −0.072 0.5697 0.0119 −0.2986 0.2631 −0.3251 −4.228 0.0574
8 0.0107 −0.0015 0.0645 −0.0093 −0.0691 −0.0018 0.2452 −0.261 1.1812
9 0.0008 0.002 0.0019 0.0028 0.0106 −0.0004 0.0599 −1.983 1.2859
10 −0.0004 0 0.0001 0.0001 0.0001 0.0002 0 −0.190 −1.3211

Also, the performance of the developed network has been and Othmer [7] or SFG [8] methods) that are based on critical properties
evaluated through using a test data set consisting of 131 data points, and a limited set of atomic and structural parameters that are available
not previously used for the network training. The results of the for most components. These methods require no supplementary data, so
evaluation indicate that the MRE and MSE for the proposed MLP they are applicable to a wide range of substances. The second category
network are 4.47% and 8.4 × 10−4, respectively. Fig. 2(a) indicates includes methods such as those proposed by Chapman–Enskog [1],
correlation between the network predictions and experimental Gilliland [2], Arnold [3], HBS [4] and Huang et al. [9] which relate diffusion
test data for binary diffusion coefficient of gases. The perfect fit to various molecular properties. The correlations of this category require
(outputs equal to targets) is indicated by the solid line. The slope and extensive data such as the knowledge of dipole moment and
the y-intercept of the best linear regression relating targets to the
network outputs are respectively 0.9924 and 0.0033, which nearly
overlaps the perfect linear fit. The correlation coefficient (R-value)
between the network predictions and the experimental test data is
0.9949. This shows a very good agreement among the simulated and
test data. As it could be seen in Fig. 2(a), there are two groups of data
points: (1) group one with lower diffusion coefficients (b0.5 cm2 s−1)
which fits very well to experiment, and (2) the group two with DAB
above 0.5 cm2 s−1 which exhibits significantly larger deviations from
data points. This observed trend could be related to higher diffusivity
rate for the substances having lower molecular weights which might
have led to higher inaccuracy in diffusivity measurement. It is obvious
that the second group has a lower molecular weight in comparison
with first group. The same trend is also seen in Arnold [3] and FSG [8]
prediction methods (see Fig. 2(b)).
For the sake of the ease of comparison, we have classified different
methods for estimating DAB into two broad categories. The first category
includes our developed MLP network and those procedures (like Chen

Fig. 1. Correlation of the predictions of the developed ANN model versus the Fig. 2. (a) Correlation of the predictions of the developed ANN model and
experimental training data. (b) correlation of Arnold [3] and FSG [8] methods versus the experimental test data.
200 R. Eslamloueyan, M.H. Khademi / Chemometrics and Intelligent Laboratory Systems 104 (2010) 195–204

Table 7
Comparison of the proposed ANN model with experimental data and other correlations for test data set.

No Binary system A–B Ref. T/K DAB/cm2 s−1 REa

Exp. ANN ANN Chapman –Enskog Gilliland Arnold HBS Chen and Othmer FSG Huang et al.

1 N2–H2 [44] 293 0.7975 0.7763 2.6526 31.652 5.3379 11.375 20.757 3.6959 4.1542 1.4847
[15] 297.2 0.779 0.7962 2.2121 28.519 0.6457 7.0957 16.870 1.164 0.5966 3.0311
b
2 H2O–H2 [15] 307.1 0.915 0.9841 7.5476 22.882 4.1907 15.451 37.566 10.147 5.4935 7.1650
3 C2H6–H2 [16] 298 0.54 0.502 7.0309 14.087 1.5019 3.1947 12.849 4.3604 0.2702 2.0747
4 n-C4H10-H2b [15] 287.9 0.361 0.3662 1.4303 4.8201 1.9248 9.6021 18.200 4.9309 3.1993 6.5402
[16] 298 0.40 0.397 0.7459 9.5405 2.2915 5.0883 21.202 0.7986 1.0697 1.2564
[17] 378.2 0.6155 0.6623 7.5983 15.948 3.6396 3.1619 20.213 0.8395 2.4347 5.9167
[17] 438.4 0.801 0.8747 9.2066 19.394 4.1140 1.8698 19.793 1.2363 2.9151 9.7740
5 CH3OH–H2 [18] 373 1.02 0.9344 8.39 30.980 9.2091 18.855 33.715 15.301 8.0739 11.047
6 C2H5OH–H2 [18] 373 0.782 0.839 7.288 22.272 5.7645 8.524 21.966 6.7871 4.5862 7.6724
423 0.946 0.9625 1.7475 22.405 2.9188 5.9845 19.245 3.2447 1.7049 7.8290
7 1-Butanol-H2 [18] 393 0.711 0.6746 5.1172 25.613 13.929 11.410 26.624 13.761 12.853 16.765
453 0.9097 0.8463 6.9668 28.116 13.739 11.588 25.994 12.831 12.661 19.493
8 2-Butanol-H2 [18] 423 0.785 0.7763 1.1137 24.108 11.332 7.5952 22.858 9.1268 10.224 15.816
9 n-C5H12-H2 [18] 353 0.4895 0.4826 1.4095 12.915 3.2653 4.4448 13.212 0.6884 2.0557 3.9090
393 0.583 0.5824 0.1078 14.106 1.9951 5.4621 11.672 1.2657 0.7696 5.2250
10 Neo-C5H12-H2 [17] 378.3 0.5443 0.5457 0.2603 12.130 1.8016 7.9616 3.6395 3.7372 0.5737 4.1285
11 n-C6H14-H2 [18] 373 0.474 0.4886 3.0764 9.6097 0.8272 9.1381 10.578 2.3265 2.0879 1.2142
423 0.5923 0.5856 1.1322 12.641 0.5587 8.331 10.29 2.8272 1.8161 4.5274
12 Cyclohexane-H2 [18] 393 0.596 0.5883 1.2872 15.782 3.2718 1.881 11.728 2.8998 2.0623 6.4596
13 Benzene-H2 [18] 393 0.65 0.7138 9.8164 16.196 5.0123 0.8493 16.167 2.1978 3.8246 8.1428
483 0.894 0.9439 5.5861 16.982 0.9251 4.2836 11.99 3.2805 0.3137 9.0043
14 Toluene-H2 [18] 373 0.5834 0.5195 10.945 21.425 12.540 6.8676 23.466 11.873 11.446 14.310
453 0.744 0.7227 2.8589 17.537 3.6428 2.0125 14.884 1.7698 2.4379 10.070
15 N2–He [26] 310.2 0.784 0.7634 2.6334 42.412 4.0088 17.588 20.579 7.7428 2.8085 1.5161
[15] 317 0.794 0.7957 0.214 41.258 1.5518 15.793 18.702 10.644 0.3208 0.4586
[22] 324 0.798 0.8239 3.246 39.606 1.7713 13.28 16.120 14.529 3.0439 3.2841
[26] 330.2 0.867 0.846 2.4203 42.809 3.1688 17.762 20.327 9.0938 1.958 2.1941
[26] 346.2 0.948 0.9 5.0639 43.848 3.7970 18.978 21.179 8.6942 2.5941 3.9713
[21] 370 1.03 0.9918 3.7107 42.898 0.5304 17.228 18.994 12.834 0.7133 2.3473
16 CO2–He [26] 324.3 0.668 0.6795 1.7257 30.554 0.6597 4.4955 11.432 16.119 1.9184 2.1323
353.1 0.794 0.7832 1.363 33.622 1.7177 7.8508 14.041 13.957 0.4887 2.3785
17 O2–He [23] 323 0.809 0.859 6.1758 36.512 4.6626 9.456 14.720 18.285 5.9713 6.3004
18 Ar–He [21] 298 0.7335 0.7453 1.6125 38.999 0.5953 13.471 18.248 11.127 0.6476 3.0142
334 0.889 0.8809 0.9056 40.279 0.1344 14.485 18.435 12.712 1.3864 0.8534
19 Kr–He [21] 322 0.7372 0.7467 1.2914 38.266 7.8980 13.013 20.746 8.742 6.7463 6.3841
20 C2H6–He [29] 322.6 0.556 0.5473 1.5604 28.914 4.6695 1.6903 10.253 15.336 3.4774 3.1476
385.3 0.745 0.746 0.1313 30.752 2.9186 2.49 9.7344 18.713 1.7046 5.6524
447.3 0.941 0.9322 0.9359 31.424 0.2064 2.1662 8.293 23.127 1.0414 6.5677
21 C2H4–He [29] 296.8 0.525 0.5592 6.5149 30.426 7.7937 4.3084 13.865 12.000 6.6407 4.3491
348.1 0.674 0.6881 2.0866 31.165 5.0641 3.6975 12.204 16.423 3.877 5.3657
399.9 0.860 0.861 0.1195 33.574 5.1527 5.8538 13.158 17.287 3.9668 8.6768
22 C3H6–He [29] 365.5 0.584 0.5231 10.424 27.270 4.0150 1.0963 7.2763 17.646 2.8148 5.4796
428 0.750 0.7604 1.3881 28.237 1.4790 1.3095 5.9084 21.904 0.2471 6.7365
23 n-C4H10-He [30] 423 0.634 0.5972 5.8027 28.264 4.9696 0.6241 12.493 15.996 3.7813 9.7761
24 n-C5H12-He [30] 372.6 0.422 0.4811 14.000 18.933 1.9211 11.808 0.7239 22.337 3.1955 0.1156
25 n-C6H14-He [28] 417 0.574 0.4814 16.130 34.910 17.040 9.3622 18.607 0.7677 16.004 20.955
26 3-methylheptane-Heb [31] 373 0.3334 0.3383 1.4729 15.961 0.8828 15.271 1.6297 19.426 2.1444 1.1596
27 2,3-Dimethylhexane-Heb [31] 373 0.342 0.3392 0.8132 17.835 1.6539 12.742 0.552 16.831 0.4241 3.6450
28 2,2,4trimethylpentane-Heb [31] 373 0.3455 0.3603 4.2917 18.790 2.6502 11.717 1.2858 16.014 1.4329 4.6211
29 C2H5OH–He [23] 423 0.821 0.7888 3.9172 33.095 4.3897 7.8764 15.356 8.6935 3.1942 9.2255
463 0.925 0.9047 2.1948 31.994 0.6021 5.437 12.543 13.614 0.6407 7.7373
30 1-Propanol-He [23] 443 0.711 0.744 4.6425 26.839 0.5793 1.2907 8.0418 17.496 1.837 5.6043
483 0.829 0.855 3.142 28.564 0.3517 0.1995 8.7714 17.839 1.6065 7.8316
31 2-Propanol-He [23] 423 0.677 0.6805 0.5184 28.449 2.5731 1.2024 10.482 14.108 1.3549 7.5008
503 0.882 0.9041 2.5046 28.784 1.2622 0.1058 8.0327 19.839 2.5284 7.9340
32 1-Butanol-He [23] 463 0.689 0.6867 0.3379 27.393 1.7551 1.0351 8.4327 15.382 0.5265 8.8075
503 0.792 0.7853 0.8512 28.476 1.1944 0.3609 8.4194 16.619 0.0410 10.167
33 1-Pentanol-He [23] 443 0.536 0.54 0.7445 19.383 5.0549 11.652 0.3513 23.896 6.3685 1.4040
523 0.729 0.7431 1.9378 23.966 3.2821 7.0835 2.4039 23.025 4.5735 7.0084
34 1-Hexanol-He [23] 463 0.531 0.5296 0.2549 19.267 4.7301 12.364 0.9004 23.499 6.0397 2.7878
503 0.631 0.6517 3.2849 23.070 1.8859 7.9659 2.3587 20.744 3.1599 7.3669
35 Ether–He [30] 423 0.607 0.7234 19.174 14.255 29.318 20.405 10.849 42.967 30.935 22.777
36 Acetone–He [30] 372.6 0.638 0.568 10.972 35.768 14.335 12.277 24.045 1.131 13.264 16.047
37 Benzene-He [30] 423 0.614 0.6256 1.8842 26.667 5.0315 1.508 8.329 13.982 3.844 9.8349
473 0.778 0.7434 4.4514 31.566 8.8662 4.1693 12.666 10.116 7.7266 15.858
38 C2H4F2–Heb [32] 429.6 0.754 0.7899 4.7598 28.772 9.2051 0.1031 7.6918 20.335 8.0698 14.130
39 Fluorobenzene-Heb [32] 429.7 0.566 0.6018 6.3331 20.671 1.2613 9.9903 0.5342 23.486 0.0266 6.6229
40 CHCl3–Heb [32] 429.1 0.624 0.6843 9.6603 24.809 0.6307 4.4529 4.0656 18.749 1.889 4.8004
41 1-Chlorobutane-Heb [32] 429.2 0.555 0.5271 5.0344 24.346 0.4398 5.2249 4.996 16.878 0.8051 5.8186
42 1-Bromobutane-Heb [32] 426.6 0.545 0.5398 0.963 25.190 0.7979 3.6815 6.726 14.279 0.4425 6.0148
43 1-Iodopropane-Heb [32] 430 0.579 0.5698 1.5932 25.743 0.7493 2.8298 7.2941 14.063 2.009 4.7382
44 CO–Heb [15] 295.6 0.702 0.6628 5.5827 40.908 0.2830 15.943 20.797 7.9242 0.9639 3.5470
45 CO2–He [23] 353 0.8 0.7828 2.1476 34.147 2.5031 8.5835 14.727 13.044 1.284 3.1519
 R. Eslamloueyan, M.H. Khademi / Chemometrics and Intelligent Laboratory Systems 104 (2010) 195–204 201

Table 7 (continued)
No Binary system A–B Ref. T/K DAB/cm2 s−1 REa

Exp. ANN ANN Chapman –Enskog Gilliland Arnold HBS Chen and Othmer FSG Huang et al.

46 C2H6–N2 [29] 345.7 0.193 0.1957 1.4199 2.4047 2.6298 0.8987 0.6838 1.5886 1.4123 2.7711
407.6 0.256 0.259 1.1768 5.8005 2.0670 2.6007 1.5362 3.1906 0.8424 6.1542
47 C2H4–N2 [15] 298 0.163 0.1615 0.9222 2.1725 5.8186 2.6404 4.0018 1.0717 4.641 2.3988
[29] 322.8 0.189 0.1865 1.3223 4.882 6.5787 2.6488 4.3581 1.3983 5.4105 5.1020
387.6 0.260 0.2565 1.358 9.0241 6.4654 1.255 3.7158 0.1891 5.2958 9.2345
449.8 0.335 0.3302 1.4317 11.731 5.8081 0.0267 2.9984 1.4138 4.6303 11.935
48 C3H6–N2 [29] 344.6 0.164 0.1629 0.6579 3.1642 5.8565 2.9552 2.2211 3.2292 4.6793 5.9182
428.9 0.243 0.2327 4.2258 9.2525 6.8143 2.2362 2.6976 2.9477 5.6491 11.833
49 n-C4H10-N2 [30] 298 0.0954 0.1041 9.069 10.872 2.9339 2.4131 6.7884 1.4487 4.221 6.6716
[17] 378.2 0.1596 0.1543 3.3006 5.2457 6.6298 4.2426 12.782 6.6517 5.4623 8.8363
b
50 iso-C4H10-N2 [15] 298 0.090 0.1025 13.836 16.090 9.1100 8.2751 6.8456 6.6204 10.474 13.071
51 CHCl3–N2 [26] 383.2 0.143 0.1443 0.8768 0.927 0.0111 1.414 0.3961 2.9683 1.2617 2.6724
52 CCl4-N2 [37] 373 0.12 0.1225 2.1221 6.6834 1.9854 2.7974 6.1032 0.2082 3.2607 0.0794
[26] 403.2 0.134 0.14 4.4831 7.372 4.6610 6.49 9.2812 3.3188 5.9697 0.5655
53 2-Propanol-N2 [26] 362.9 0.159 0.1534 3.5281 5.5622 6.5252 10.952 10.339 9.5124 5.3564 7.7870
54 1-Pentanol-N2b [37] 418 0.159 0.1536 3.3856 5.9872 6.0177 7.6634 8.7209 10.325 4.8425 10.505
55 1-Heptanol-N2b [37] 453 0.145 0.1513 4.3741 0.4355 0.5226 0.3701 2.2396 5.7141 1.7795 6.1826
56 Acetone–N2 [26] 363.3 0.154 0.1581 2.6721 0.2477 0.1295 4.1671 9.9203 3.5773 1.1193 1.5048
57 Methyl acetate–N2 [26] 383.1 0.192 0.1601 16.624 18.938 18.171 20.769 23.438 21.171 17.149 20.362
58 Ethyl formate–N2 [26] 383.2 0.158 0.1598 1.1331 1.6307 0.5179 3.6884 3.6737 4.4064 0.7259 3.1872
59 Cyclohexane–N2 [26] 383.2 0.134 0.1384 3.2781 0.8791 0.3932 1.5924 2.3891 5.1352 0.8523 3.0658
60 Benzene–N2 [26] 364.2 0.129 0.1391 7.8585 5.9007 1.4833 0.6364 0.5161 0.8776 2.7522 0.0239
403.2 0.163 0.1625 0.2916 2.3726 4.0351 3.6048 4.2327 5.6937 2.8351 7.7903
61 O2–N2 [15] 316 0.230 0.2406 4.6022 9.9081 0.8462 0.8592 0.5331 4.0821 0.3936 1.2585
62 n-C4H10-Ar [30] 423 0.17 0.1708 0.4828 0.8455 1.5604 2.6639 9.4397 0.8149 0.3295 6.5393
63 n-C5H12-Ar [30] 372.6 0.115 0.1323 15.079 6.7861 2.0825 4.2821 0.8066 0.2236 3.359 0.0426
64 n-C6H14-Ar [30] 473 0.175 0.175 0.003 9.5302 8.4723 4.9301 10.097 11.139 7.3278 15.495
65 Benzene–Ar [30] 372.6 0.142 0.1369 3.6027 7.274 12.472 12.502 35.692 14.278 11.378 14.222
66 Acetone–Ar [30] 423 0.213 0.1777 16.552 13.998 12.850 14.419 22.198 14.898 11.761 17.258
67 Ether–Ar [30] 298 0.0849 0.1143 34.628 39.521 38.823 27.672 25.601 28.932 40.56 43.864
68 CO2–Arb [15] 276.2 0.133 0.1251 5.9342 5.2808 0.5327 3.3332 1.299 3.1285 1.7898 6.1807
69 Xe–Arb [15] 329.9 0.137 0.1286 6.1364 1.4352 2.5158 3.0279 1.2832 4.3428 3.7977 3.5713
70 CO2–N2 [26] 313.7 0.201 0.2038 1.3774 12.562 11.757 13.751 10.998 9.4664 10.654 9.7192
342.6 0.238 0.2379 0.0281 15.719 13.048 14.223 12.063 10.319 11.961 12.979
71 CO2–H2O [15] 307.4 0.202 0.1552 23.154 14.414 8.8797 21.802 30.277 4.3914 10.241 11.960
[26] 408.5 0.311 0.3229 3.8396 13.842 16.316 9.6728 19.887 3.8966 17.771 11.400
72 N2–H2O [26] 408.2 0.464 0.4486 3.3149 16.344 3.6982 19.047 26.707 13.406 2.494 7.7513
73 Benzene–O2b [15] 311.3 0.101 0.1058 4.7373 7.9076 3.1946 6.0845 10.400 5.9293 1.9842 0.7687
74 2-Propanol-CO2 [26] 383.5 0.12 0.1241 3.4027 16.389 6.9790 8.4742 6.9244 0.1095 8.3167 4.0882
75 Ethyl formate–CO2 [26] 348.3 0.105 0.1015 3.3317 8.6748 1.6847 16.743 15.172 9.5223 0.4553 2.0111
76 Cyclohexan–CO2 [26] 403.4 0.114 0.1192 4.5528 7.9303 0.4899 11.623 5.9405 8.9749 0.7543 4.3957
77 Benzene–CO2 [26] 378 0.116 0.1126 2.9703 5.2616 6.3605 17.170 16.507 12.04 5.1896 8.5612
408.2 0.13 0.1286 1.1142 5.4039 4.4147 13.916 13.601 9.7968 3.2195 8.4376
78 CCl4–CO2 [26] 384.3 0.093 0.0993 6.7538 18.346 5.0046 5.5647 0.4704 0.9756 6.3175 2.1139
79 CHCl3–CO2 [26] 383.3 0.12 0.1098 8.5369 0.1927 9.2215 20.077 17.052 15.313 8.0864 11.663
80 C2H5Cl–CH3Cl [17] 418.5 0.1216 0.119 2.1214 13.034 5.3040 9.0871 11.560 4.0743 6.6207 0.2456
81 C2H6–CHb4 [17] 298 0.1615 0.1556 3.6697 6.7518 3.3913 2.6262 7.1964 0.2314 2.1833 0.1167
378 0.252 0.2484 1.4119 2.2627 6.1314 2.3126 7.3987 1.6681 4.9577 8.3375
437.8 0.3274 0.3492 6.6628 6.2312 6.5748 1.4794 6.8797 1.2664 5.4066 12.059
82 C3H8–CH4 [17] 378.2 0.2013 0.2244 11.491 2.2402 3.5468 1.3516 4.6646 1.7793 2.3407 5.8261
83 n-C4H10-CH4b [17] 298 0.1079 0.1034 4.2109 17.106 3.1223 0.3646 16.340 0.8314 4.4118 6.8668
378.3 0.1704 0.1725 1.2198 6.0629 0.8621 0.1754 15.494 1.6668 0.3775 3.2113
436.3 0.2243 0.2271 1.2464 0.2008 3.3303 0.8763 15.595 3.2908 2.1215 8.9273
84 Iso-C4H10-C2H6b [17] 298 0.0681 0.0664 2.494 29.835 10.011 2.3747 4.9382 3.5634 11.387 14.005
436.9 0.1476 0.1414 4.234 6.3411 0.8513 4.3482 8.1445 4.4948 0.3884 6.6239
85 CH3Cl–CH4 [17] 418.2 0.2795 0.2849 1.9495 4.3259 4.9588 6.0645 13.731 4.8154 3.7704 9.5080
86 C3H8–C2H6 [17] 378.2 0.1316 0.1212 7.8929 12.084 0.2783 3.1961 2.3819 0.8172 1.5322 2.0914
a Pred Exp Exp
RE = |DAB − DAB | / (DAB ) 100.
b
Binary A–B system not used in the training of the proposed ANN.

intermolecular forces, liquid molar volume at normal boiling point and
Sutherland's constant. Since the all needed parameters of the second
category correlations are only available for a limited number of common
substances, the range of applicability of the correlations in this category is
seriously restricted. The methods of category one, are usually easier to
apply and are applicable to a wider range of compounds than those of
category two.
Before comparing the accuracy of the developed network model
with other correlations, a brief explanation should be given about how
the correlation parameters have been determined. To use Chapman–
Enskog [1] and HBS [4] equations, some relationships have to be used
for determining the interaction value σAB from σA and σB. ΩAB is a
function of kT/AB. Also, the value of the characteristic Lennard–Jones
energy and length must be determined. The simple equations σ = 1.18
V1/3, /k = 1.21 Tb, AB = (A B)1/2 and σAB = (σA + σB) / 2 have been
employed to predict these parameters. The accurate relation of
Neufield et al. [39] has been used to determine ΩAB. The dimensionless
parameter ΩAB is more complex, but its value is usually near 1. For
binary systems that one or both components of the gas mixture are
polar, , σ and ΩAB are estimated by Brokaw's method [40]. Note that
202 R. Eslamloueyan, M.H. Khademi / Chemometrics and Intelligent Laboratory Systems 104 (2010) 195–204

Fig. 3. Mean relative errors (MRE) of the developed ANN model and various other methods used for prediction of DAB in comparison to the experimental test data.

in the Brokaw's method [40], the polarity effect is related exclusively
to the dipole moment; this may not always be a satisfactory
assumption [41]. In Gilliland [2] and Huang et al.'s [9] methods, the
molar volumes for each component are found by summing atomic
diffusion volumes proposed by Fuller et al. [42]. To use Arnold [3]
equation for prediction of DAB, a relation must be chosen to obtain the
liquid molar volume at normal boiling point. In this study, the Tyn and
Calus [43] method is used to predict this parameter. Also, the value of
Sutherland's constant in Eq. (4) is determined by Arnold's [3] method
as a function of normal boiling point and liquid molar volume at
normal boiling point. All critical properties have been taken from the
book written by Reid et al. [34].
To have a comparison between the neural network model developed
in this work with others’ correlations, we have estimated the diffusivity
of several components in the test data set by using the developed
network as well as seven different correlations proposed by Chapman–
Enskog [1], Gilliland [2], Arnold [3], Hirschfelder, Bird and Spotz (HBS)
[4], Chen and Othmer [7], Fuller, Schettler and Giddings (FSG) [8] and
Huang et al. [9]. Table 7 indicates the results of this comparison and the
relative error of each method separately. Although a few input data are
required for the proposed ANN method, its accuracy is more than the
accuracy of all seven well known correlations suggested by others. It
should be mentioned that some of the binary systems in Table 7 have
not been used for the training step of the network. The degree of
accuracy of the ANN model for these systems proves the extrapolation
capability of the developed model in this study.
Fig. 3 shows the mean relative errors of the proposed ANN model
and various methods used in prediction of DAB for all test data points.
As can be seen from the figure, the proposed neural network model is
better than the other methods. The mean relative error of the ANN
method for all test data is 4.47%. The Gilliland [2] method, a
Table 8
Comparison between the predicted diffusivity by the ANN model in high pressure and
the experimental data.
2 −1 2 −1
No. Binary system A–Ba T/K P/atm Dexp
AB /cm s DPred
AB /cm s RE

1 N2–He 323 9.97 0.0820 0.0823 0.36

29.9 0.0272 0.0274 0.73
49.8 0.0166 0.0165 0.60
59.8 0.0140 0.0137 2.14
2 Ar–He 298 9.97 0.0742 0.0748 0.80
29.9 0.0249 0.0249 0.00
49.8 0.0152 0.0150 1.31
59.8 0.0127 0.0125 1.57
323 9.97 0.0851 0.0847 0.47
29.9 0.0288 0.0282 2.08
49.8 0.0175 0.0170 2.85
59.8 0.0145 0.0141 2.75
3 CH4–He 248 9.97 0.0501 0.0525 4.79
29.9 0.0169 0.0175 3.55
49.8 0.0103 0.0105 1.94
59.8 0.00872 0.0088 0.91
273 9.97 0.0588 0.0596 1.36
29.9 0.0198 0.0199 0.50
49.8 0.0119 0.0119 0.00
59.8 0.0101 0.0099 1.98
4 CO2–He 298 9.97 0.0616 0.0600 2.59
29.9 0.0206 0.0200 2.91
323 9.97 0.0701 0.0677 3.42
29.9 0.0236 0.0226 4.23
Fig. 4. The predicted binary diffusion coefficients by the ANN model at various
a
temperatures. Experimental data from Hu and Kobayashi [47].
 R. Eslamloueyan, M.H. Khademi / Chemometrics and Intelligent Laboratory Systems 104 (2010) 195–204
modification of the Arnold [3] method (with an average error of
7.25%), does not consider the Sutherland's constant and has the lower
average error of Arnold's correlation, that is 7.09%. The HBS [4]
method with an average error of 11.64% and the Chapman-Enkog [1]
method with an average error of 18.7% are among time-consuming
techniques and the least accurate. It should be noted that the
calculation of Lennard–Jones parameters in the Chapman–Enskog
[1] and HBS [4] equations by other methods can improve the quality of
their relative errors. Therefore, the accuracy of the estimation of
binary gas diffusion coefficients by Chapman–Enskog [1] and HBS [4]
equations are dependent on the correlation used for the determina-
tion of Lennard–Jones parameters. This is definitely not an advantage
for these methods, and can cause the user of these correlations to get
into trouble. Although the average relative errors of FSG [8] and
Huang et al. [9] methods (i.e. 4.53% and 4.79%, respectively) are about
1.3% and 7.1% greater than that of the ANN model, they show good
results which are close to the ANN model developed in this work.
Since the FSG [8], Huang et al. [9] and Gilliland [2] methods estimate
the needed parameters for theirs correlation based on the structural
formula of components, they get in trouble when these methods are
used to determine the binary diffusivity of isomers (e.g. 3-methyl-
heptane-He, 2,3,-dimethylhexane-He, and 2,2,4-trimethylpentane-He
in Table 7). Indeed, isomers have the same group contributions and
this fact leads to the same values for the correlation parameters. It
should be mentioned, the proposed ANN method has an average
relative error of 2.19% for these three mixtures while the FSG [8],
Huang et al. [9] and Gilliland [2] methods only deviate on the average
by 1.33%, 3.14% and 1.73%, respectively. A comparison of the ANN
model with the method of Chen and Othmer [7] indicates that both
methods require the same input variables, but the mean relative error
of Chen and Othmer [7] method is two times greater than that of the
ANN model.
As Table 7 shows , the ANN predictions of gas diffusivities for some
of the binary systems (i.e. H2O–H2, CH3OH–H2, Benzene–H2 at 393 K,
Toluene–H2 at 373 K, C3H6–He at 365.5 K, n-C5H12–He, n-C6H14–He,
Ether–He, Acetone–He, CHCl3–He, n-C4H10–N2 at 298 K, iso-C4H10–N2,
Methyl acetate-N2, n-C5H12–Ar, Acetone–Ar, Ether–Ar, CO2–H2O at
307 K, C3H8–CH4) does not follow the experimental data very well.
The same problem is also observed when we use Arnold [3] and FSG
[8] methods for estimation of diffusion coefficients of some binary
systems such as CH3OH–H2, n-C6H14–He, Ether–He, Acetone–He, iso-
C4H10–N2, Methyl acetate–N2, Acetone–Ar, Ether–Ar, CO2–H2O.
Marrero and Mason [45] indicated that, in most cases, DAB is
proportional to temperature as TS, where s is in the range of 1.5 to 2.
Most correlations are consistent to this range of s value; for instance,
in the FSG [8] and Huang et al. [9] methods, DAB ∝ T1.75, in the Chen
and Othmer [7] method, DAB ∝ T1.81, in the Gilliland [2] method,
DAB ∝ T1.5, and in the Slattery and Bird [46] method DAB ∝ T2.34. By
plotting the predictions of the ANN model for diffusion coefficients in
terms of temperature the proportionality of the proposed ANN
method with temperature can be obtained. Fig. 4 represents the
estimation of the ANN model for diffusivity coefficients of five binary
gases (C6H14–H2, Ar–He, C2H6–N2, C6H6–CO2, and 1-Butanol–H2) at
several temperatures. According to this figure, the correlation of gas
diffusivity with temperature for the ANN model follows a straight line
with an average slope (m) of 1.568. This slope agrees very well with
the proposed value of 3/2 by Chapman–Enskog [1] and Gilliland [2]
and differs from the values of Chen and Othmer [7] (i.e. 1.81) and
Slattery and Bird [46] (2.334).
It should be noted that the abovementioned correlations consider
the effect of pressure on the gas diffusivity DAB as proportional to the
inverse of pressure.
On the other hand, the proposed neural network model does not
directly take into account the effect of pressure on the gas diffusivity.
The lack of sufficient training experimental data points at high
pressures is the main reason that the authors decided to exclude the
203
pressure from the network input parameters. Although this issue
might be considered as a weakness of the ANN model, it can be
overcome through using a simple relation between the diffusivities at
atmospheric and high pressures: (P.DAB)1 = (P.DAB)2. For this pur-
pose, the predictions of the ANN model at high pressures through
using this formula were compared to that of experimental diffusivity
at high pressures. Table 8 presents the comparison between the
predicted diffusivity by the ANN model in high pressure and the
corresponding experimental data. As can be seen from Table 8, the
ANN model predictions for high pressure diffusivity coefficients are in
good agreement with experimental data.
In general, the ANN model, in spite of its simplicity and ease of
application, have the lowest average error and can be applied to a
large number of binary systems, polar and non-polar, as well as a wide
range of temperatures.
4. Conclusions
An artificial neural network (ANN) has been developed for the
prediction of diffusion coefficients of binary gaseous mixtures at
atmospheric pressure as a function of temperature. Temperature,
critical temperature, critical volume and molecular weight of each
component are used as the input variables of the proposed ANN
model. A set of 336 experimental data points for binary diffusion
coefficient of both organic and inorganic gases has been used for the
network training stage. The best architecture of the feedforward
network, obtained by trial and error, consists of one hidden layer with
ten neurons. The mean square error (MSE) of the designed network
for training data points is near 0.001. The performance of the
proposed ANN model has also been examined through its application
to a test data set consisting of 131 experimental data of diffusivity for
various binary systems over a wide range of temperatures. The results
of this evaluation indicate that the developed ANN model is able to
predict binary gases diffusion coefficient with a considerably lower
relative error (4.47%) than that of other alternative correlations. The
results of applying the trained neural network model to the test data
indicate that the method has a very good interpolation and
extrapolation capabilities with respect to not only the temperature
ranges but also the molecular types. Other advantage of the ANN
method, in comparison to some suggested cumbersome correlations,
is that its input variables are usually available for most substances.
References
[1] J.C. Giddings, Dynamics of Chromatography, Marcel Dekker, New York, 1965.
[2] E.R. Gilliland, Diffusion coefficients in gaseous systems, Ind. Eng. Chem. 26 (6)
(1934) 681–685.
[3] J.H. Arnold, Studies in diffusion: I—estimation of diffusivities in gaseous systems,
Ind. Eng. Chem. 22 (10) (1930) 1091–1095.
[4] J.C. Hirschfelder, R.B. Bird, E.L. Spotz, The transport properties of gases and
gaseous mixtures, J. Chem. Phys. 44 (1) (1949) 205–231.
[5] J.O. Hirschfelder, C.F. Curtis, R.B. Bird, Molecular Theory of Gases and Liquids,
Wiley, New York, 1954.
[6] R.B. Bird, W.E. Stewart, E.N. Lightfoot, Transport Phenomena, Wiley, New York,
1960.
[7] D.F. Othmer, H.T. Chen, Correlating diffusion in binary gas systems, I&EC Process
Design and Development 1 (4) (1962) 249–254.
[8] E.N. Fuller, P.D. Schettler, J.C. Giddings, A new method for prediction of binary gas-
phase diffusion coefficients, Ind. Eng. Chem. 58 (5) (1966) 19–27.
[9] T.C. Huang, F.J.F. Young, C.J. Huang, C.H. Kuo, Measurements of diffusion
coefficients by the method of gas chromatography, J. Chromatogr. 70 (1972)
13–24.
[10] E.N. Fuller, J.C. Giddings, A comparison of methods for predicting gaseous
diffusion coefficients, J. Gas Chromatogr. 3 (7) (1965) 222–227.
[11] R. Eslamloueyan, M.H. Khademi, Estimation of thermal conductivity of pure gases
by using artificial neural networks, Int. J. Thermal Sci. 48 (2009) 1094–1101.
[12] R. Eslamloueyan, M.H. Khademi, Using artificial neural networks for estimation of
thermal conductivity of binary gaseous mixtures, J. Chem. Eng. Data 54 (2009)
922–932.
[13] G.W. Kauffman, P.C. Jurs, Prediction of surface tension, viscosity, and thermal
conductivity for common organic solvents using quantitative structure–property
relationships, J. Chem. Inf. Comput. Sci. 41 (2001) 408–418.
204 R. Eslamloueyan, M.H. Khademi / Chemometrics and Intelligent Laboratory Systems 104 (2010) 195–204
[14] R. Boozarjomehri, F. Abdolahi, M.A. Moosavian, Characterization of basic
properties for pure substances and petroleum fractions by neural network,
Fluid Phase Equilib. 231 (2005) 188–196.
[15] E.L. Cussler, Diffusion Mass Transfer in Fluid Systems, 3th edition, Cambridge
University Press, United States of America, 2009 (New York).
[16] G. Karaiskakis, D. Gavril, Determination of diffusion coefficients by gas
chromatography, J. Chromatogr. A 1037 (2004) 147–189.
[17] S. Gotoh, M. Manner, J.P. Sorensen, W.E. Stewart, Binary diffusion coefficients of
low-density gases I. Measurements by modified Loschmidt method, J. Chem. Eng.
Data 19 (2) (1974) 169–171.
[18] T.C. Huang, S.J. Sheng, F.J.F. Young, Determination of gas-vapor binary diffusion
coefficient by gas chromatography, J. Chin. Chem. Soc. (Taipei) 15 (1968) 127–139.
[19] J.C. Giddings, S.L. Seager, Method for rapid determination of diffusion coefficients,
Ind. Eng. Chem. Fundam. 1 (4) (1962) 277–283.
[20] J. Bohemen, J.H. Purnell, Diffusional band spreading in gas-chromatographic
columns. Part I. The elution of unsorbed gases, J. Chem. Soc. (1961) 360–367.
[21] J.C. Liner, S. Weissman, Determination of the temperature dependence of gaseous
diffusion coefficients using gas chromatographic apparatus, J. Chem. Phys. 56
(1972) 2288–2290.
[22] S.L. Seager, L.R. Geertson, J.C. Giddings, Temperature dependence of gas and vapor
diffusion coefficients, J. Chem. Eng. Data 8 (2) (1963) 168–169.
[23] E. Grushka, V. Maynard, Measurements of gaseous diffusion coefficients by gas
chromatography, J. Chem. Ed. 49 (1972) 565–567.
[24] I. Nagata, T. Hasegawa, Gaseous diffusion coefficients, J. Chem. Eng. Jpn. 3 (2)
(1970) 143–145.
[25] F. Yang, S. Hawkes, F.T. Lindstrom, Determination of precise and reliable gas diffusion
coefficients by gas chromatography, J. Am. Chem. Soc. 98 (17) (1976) 5101–5107.
[26] G.L. Hargrove, D.T. Sawyer, Determination of gaseous inter-diffusion coefficients
for solute vapor-carrier gas pairs, Anal. Chem. 39 (2) (1967) 244–246.
[27] E. Grushka, V. Maynard, Measurement of diffusion coefficients of octane isomers
by the chromatographic broadening method, J. Phys. Chem. 77 (11) (1973)
1437–1442.
[28] E.N. Fuller, K. Ensley, J.C. Giddings, Diffusion of halogenated hydrocarbons in
helium. The effect of structure on collision cross sections, J. Phys. Chem. 73 (1969)
3679–3685.
[29] H.J. Arnikar, T.S. Rao, K.H. Karmarkar, The use of an electrodeless discharge as a
detector in gas chromatography, J. Chromatogr. 26 (1967) 30–34.
[30] J.C. Giddings, K.L. Mallik, An expanding role for gas chromatography, measure-
ment of chemical and physical parameters, Ind. Eng. Chem. 59 (4) (1967) 19–27.
[31] H.J. Arnikar, H.M. Ghule, Electrodeless discharge as detector in the rapid
determination of binary diffusion coefficient of gases, Int. J. Electron. 26 (2)
(1969) 159–162.
[32] J.H. Knox, L. McLaren, The spreading of air peaks in capillary and packed gas
chromatographic columns, Anal. Chem. 35 (4) (1963) 449–454.
[33] S.M. Ashraf, R. Srivastava, A. Hussain, Determination of binary gas-phase diffusion
coefficients using chromatography, J. Chem. Eng. Data 31 (1) (1986) 100–102.
[34] R.C. Reid, J.M. Prausnitz, B.E. Poling, The properties of gases and liquids, 4th ed.,
McGraw-Hill, New York, 1987.
[35] S. Haykin, Neural Networks: A Comprehensive Foundation, 2nd ed., Prentice-Hall,
Englewood Cliffs, NJ, 1999.
[36] K. Levenberg, A method for the solution of certain problems in least squares, SIAM
J. Numer. Anal. 16 (1944) 588–604.
[37] D. Marquardt, An algorithm for least-squares estimation of nonlinear parameters,
SIAM J. Appl. Math. 11 (1963) 431–441.
[38] M.T. Hagan, M. Menhaj, Training feedforward networks with the Marquardt
algorithm, IEEE Trans. Neural Netw. 5 (1994) 989–993.
[39] P.D. Neufeld, A.R. Janzen, R.A. Aziz, Empirical equations to calculate 16 of the
transport collision integrals Ω(l, s)* for the Lennard–Jones (12–6) potential, J.
Chem. Phys. 57 (1972) 1100–1102.
[40] R.S. Brokaw, Predicting transport properties of dilute gases, Ind. Eng. Chem.
Process Des. Dev. 8 (1969) 240–253.
[41] J.J. Byrne, D. Maguire, J.K.A. Clarke, Gas-phase interdiffusion coefficients for some
polar organic compounds, J. Phys. Chem. 71 (1967) 3051–3052.
[42] E.N. Fuller, K. Ensley, J.C. Giddings, Diffusion of halogenated hydrocarbons in
helium, the effect of structure on collision cross sections, J. Phys. Chem. 73 (11)
(1969) 3679–3685.
[43] M.T. Tyn, W.F. Calus, Estimating liquid molal volume, Processing 21 (4) (1975)
16–17.
[44] J.C. Giddings, S.L. Seager, Rapid determination of gaseous diffusion coefficients by
means of gas chromatography apparatus, J. Chem. Phys. 33 (1960) 1579–1580.
[45] T.R. Marrero, E.A. Mason, Gaseous diffusion coefficients, J. Phys. Chem. Ref. Data 1
(1972) 3.
[46] J.C. Slattery, R.B. Bird, Calculation of the diffusion coefficient of dilute gases and of
the self-diffusion coefficient of dense gases, Am. Inst. Chem. Eng. 4 (2) (1958)
137–141.
[47] A.T.C. Hu, R. Kobayashi, Measurements of gaseous diffusion coefficients for dilute
and moderately dense gases by perturbation chromatography, J. Chem. Eng. Data
15 (2) (1970) 328–335.