Anda di halaman 1dari 11

Article

pubs.acs.org/IECR

Analysis of Mass Transfer Performance of Monoethanolamine-Based


CO2 Absorption in a Packed Column Using Articial Neural Networks
Kaiyun Fu, Guangying Chen, Zhiwu Liang,*, Teerawat Sema, Raphael Idem,,
and Paitoon Tontiwachwuthikul,

Joint International Center for CO2 Capture and Storage (iCCS), College of Chemistry and Chemical Engineering, Hunan
University, Changsha, 410082, P.R. China

International Test Centre for CO2 Capture (ITC), Faculty of Engineering and Applied Science, University of Regina, Regina,
Saskatchewan, S4S 0A2, Canada

ABSTRACT: Exact and reliable estimation of mass transfer performance is very important for the design, simulation, and
optimization of CO2 absorption in a packed column. In this study, two types of articial neural networks (ANNs), namely back-
propagation neural network and radial basis function network, were applied to predict the mass-transfer performance of CO2
absorption into aqueous monoethanolamine (MEA) in packed columns (containing Berl saddles, Pall rings, IMTP random
packing, and 4A Gempack, Sulzer DX structured packing, respectively) from input variables. These variables were inert gas ow
rate, liquid ow rate, solution concentration, liquid CO2 loading, CO2 mole fraction, temperature, and total packing area, which
were considered to predict the targeted output mass transfer variables. The predicted results from ANN were validated against
experimental data as well as compared with results from well-known correlations in terms of the volumetric mass ux, CO2 mole
fraction, and temperature proles along the height of the packed column. The comparisons between the predicted and
experimental results showed that the proposed ANN models performed very well in predicting mass transfer performance of CO2
absorption into aqueous MEA in a packed column.

1. INTRODUCTION desired amount of CO2 from the ue gas to the amine solution in
One of the most serious environmental concerns facing mankind a packed column, accurate prediction of the mass transfer
currently is that of an increased dependence on the combustion eciency is critical. To maximize the design condence and to
of fossil fuels (i.e., coal, petroleum, and natural gas) for energy provide the best possible scale-up data for the construction of
supply, which results in an increased emission of carbon dioxide packed column-based CO2 capture plants, a deep understanding
(CO2). Thus, a question has arisen as to how to reduce CO2 of the fundamental concepts of designing and modeling is
emission in an economical and ecient way. CO2 capture from essentially needed.9 Presently, CO2 absorption into aqueous
fossil fuel-red power plants, which accounts for approximately amines in a packed column is regarded as one of the most
80% of CO2 emissions worldwide, is a major option to mitigate intricate systems because of the existence of complex interactions
CO2 emissions to the atmosphere .1 among hydromechanics, mass transfer, heat transfer, thermody-
Presently, an amine-based CO2 absorption technology in a namics, and kinetics of the chemical reaction.
packed column is considered to be the most commercially Up to now, a number of empirical or semiempirical
mature and cost-eective method for separating CO2 from fossil correlations have been proposed to predict mass transfer in
fuel-red power plants.2 Of all the amines, monoethanolamine packed columns. However, these correlations vary in accuracy,
(MEA), the most studied chemical solvent for CO2 capture, is limitation, and sometimes system specic applicability. Wang et
considered to be the bench-mark absorbent which has also been al.10 presented an extensive review of numerous mass transfer
employed industrially for over 50 years.3,4 This technology is correlations and concluded that the limited experimental data
especially suitable for absorbing large volumes of low-pressure and the lack of understanding of the complex transport
ue gas, and retrotting to existing power plants without phenomena occurring in the packed columns make it dicult
signicant changes in the plant conguration. The characteristics
to use these correlations with a wide application range and a high
of ue gas (i.e., large volume and low pressure) require large and
level of condence. To improve reliability and accuracy of the
highly ecient separation equipment to achieve a certain capture
eciency, which directly determines the installed equipment and predictive models used for design, simulation, and optimization
operational costs. of CO2 absorption in packed column, it becomes essential to
Over the past decades, various separators, such as packed, tray, explore advanced predictive correlations for mass transfer
and spray column, and membrane contactors have been performance in packed column.
intensively studied and have received signicant attention for
CO2 absorption.58 Among them, the packed column has been Received: October 1, 2013
widely known to possess high hydrodynamic and mass transfer Revised: January 14, 2014
performance characteristics. To accurately estimate the proper Accepted: February 24, 2014
size and height of a packed column required for the transfer of the Published: February 24, 2014

2014 American Chemical Society 4413 dx.doi.org/10.1021/ie403259g | Ind. Eng. Chem. Res. 2014, 53, 44134423
Industrial & Engineering Chemistry Research Article

With the advent of the ever-accelerated pace of computer rules to respond to a new set of input data within the range of the
technology, computer-aided numerical simulation has enjoyed training examples to predict the required output.25 In general,
great popularity and applicability in many aspects of science and ANN is used to settle engineering problems that are dicult for
engineering. In the eld of chemical engineering, a number of conventional process simulation tools or mechanistic correlation,
process simulation tools, such as articial networks (ANNs), as it possesses a high self-mapping ability and does not require
Aspen Plus, Aspen HYSYS, Promax, and ProTreat, have played prior knowledge of the relationships of the process parameters
important roles in modeling process units and/or optimizing and accurate mathematical models (or correlations). Moreover,
process-congurations with the aims of quality improvement and the intricate relationships among tting parameters increase the
cost reduction.1114 For the simulation of CO2 absorption with freedom of the ANN framework, making them more eective to
chemical reaction in a packed column, several main approaches reect the complexity of the systems, allowing higher general-
have been used to estimate the mass transfer performance in ization ability, and making them more accurate than common
terms of the volumetric mass ux (NCO2av), mass transfer mathematical correlations. For these reasons, it is conventional
coecient, and CO2 concentration and temperature distribu- and eective to use the ANN method to analyze and settle
tions along the column. Ji et al.15 formulated an absorption complex engineering problems.
There are two kinds of ANNs that are very powerful and
model in terms of NCO2av to describe the system of CO2
practical to approximate complex nonlinear systems: back-
absorption into sodium hydroxide solution (NaOH) in a packed propagation neural network (BPNN) and radial basis function
column by using a nonlinear least-squares method. Aboudheir et network (RBFNN). The emphasis of this paper is to apply
al.16 developed a rigorous computer model for the simulation of BPNN and RBFNN to the simulation of CO2 absorption into
the absorption of CO2 in aqueous 2-amino-2-methyl-1-propanol aqueous MEA in a packed column through the view of
(AMP) solutions in a packed absorption column that took the engineering. The high accuracy and outstanding performance
eects of heat transfer into account. Liu et al.17 introduced a of BPNN and RBFNN were validated by comparing the
complex model for simulating the chemical absorption of CO2 in predicted results with experimental data and literature
aqueous NaOH in a randomly packed column by combining the correlations.
quasi-single-liquid-phase computational uid dynamics (CFD)
model, the turbulent mass transfer model with the c 2c 2. THEORETICAL BASIS OF CO2MEAH2O SYSTEM
equations (which considered the inuence of concentration
2.1. Chemical Reactions. According to the generally
eld in calculating the turbulent diusivity) for its closure, and
accepted zwitterion mechanism, which was originally proposed
the heat balance equation. In addition, Hanley and Chen18
by Caplow26 and reintroduced by Danckwerts27 for primary
compared correlation predictions with experimental data in
amine reactions with CO2, MEA interacts with CO2 in a 2:1 ratio;
terms of height equivalent to a theoretical plate (HETP), CO2
that is, one molecule of MEA absorbs CO2 to form a zwitterion,
mole fraction and temperature distributions along the column for while another one acts as a base for the deprotonation of a
a number of dierent chemical systems, packing types performed zwitterion.
with Aspen Rate Based Distillation v7.2.2 While these methods
can predict the chemical absorption of CO2 in a packed column
eectively, they suer numerous drawbacks in presenting a The reaction of CO2 with MEA (RNH2 denotes as MEA):
feasible model, such as the requirements of (i) a high-level k1, k 1
understanding of intricate relationships among the parameters, RNH 2 + H+ RNH+3 (1)
(ii) huge databases within software, (iii) explicitly predened
rules or equations, and (iv) the simulation performing in an ideal k 2 , k 2
CO2 + RNH 2 RNH+2 COO (2)
way. For example, Aspen plus has been widely used in the
modeling of CO2 capture by aqueous MEA solution in a packed Dissociation of dissolved carbon dioxide to carbonic acid:
column and achieved good prediction results. This is because
MEA has been well studied for CO2 capture, and the k 3 , k 3
CO2 + H 2O HCO3 + H+ (3)
thermodynamic and kinetics data are adequate for the MEA
CO2H2O system within an amine package. However, as to new Formation of bicarbonate:
and data-limited solvents and packings, the fundamental data and
the underlying relationships in the chemical mass transfer k4 , k 4
CO2 + OH HCO3 (4)
process are either unknown or dicult to determine, thereby
making Aspen plus modeling powerless. Dissociation of bicarbonate:
For more than ve decades, especially within the past 15 years,
k5 , k 5
articial neural networks (ANNs) have found widespread HCO3 H+ + CO32 (5)
application in diverse elds such as pattern recognition,
identication, classication, and controlling systems owing to Ionization of water:
their powerful ability to deal with the uncertainties of noisy data
or linear or nonlinear relationships.1924 An ANN attempts to k6 , k 6
H 2O OH + H+ (6)
mimic human brain activities, mapping from a set of given data
(input data) to an associated set of known data (target output), where ki is the second order forward rate coecient for reaction
comparing the output of the network to the given data, and then (i), and ki is the backward rate coecient for the reaction (i).
adjusting the internal weights and biases, with the goal of Reactions 2 and 4 are considered to be reversible with nite
minimizing the error between the network output and the target reaction rates, whereas reactions 1, 3, 5, and 6 are considered to
output. Once the learning (or training) process is satisfactorily be reversible, instantaneous, and at equilibrium, since they
completed, the neural network hence generates a set of certain involve only proton transfer.28
4414 dx.doi.org/10.1021/ie403259g | Ind. Eng. Chem. Res. 2014, 53, 44134423
Industrial & Engineering Chemistry Research Article

2.2. Kinetics Model Based on Zwitterion Mechanism. 3. MODELING AND OPTIMIZATION METHOD FOR
From the zwitterion mechanism, the general reaction rate of CO2 MASS TRANSFER WITH CHEMICAL REACTION
with MEA is given as follows: The design and scale-up of a packed column for various packings
k 2 +
+
k b[RNH3 ] require accurate assessments of mass transfer models such as the
[CO2 ][RNH 2] k2
[RNH 2 COO ] k b[B] volumetric mass ux, the mass transfer coecients of the gas and
rCO2 = 1 k 2 liquid phases, and the eective interfacial area. Generally,
+ k k [B]
k2 2 b researchers assess the reliability of the model on the basis of
(7) the deviation of the predictions and experimental results.
3.1. Basic Theory of Mass Transfer. The mass ux per unit
where rCO2 is the overall CO2 absorption rate, B is a base which
volume of CO2 (NCO2av) transferring from the gas phase to a
could be amine, H2O, or OH. kb and kb are the forward and
backward second order reaction-rate constant for base B, liquid bulk at a steady state can be expressed as5
respectively. * )
NCO2a v = K Ga v P(yCO yCO (12)
In the case of the deprotonation of the zwitterions being very 2 2

fast, such as in the CO2MEA reaction, k2/ where NCO2av is the mass ux per unit volume of CO2, KGav is the
k2[RNH+2 COO]kb[RNH3+]/kb[B] is usually considered gas-phase volumetric overall mass transfer coecient, P is the
to be negligible, and hence, the rate of reaction can be reduced to
total pressure, yCO2 and y*CO2 represent the mole fraction of CO2
[CO2 ][RNH 2] in the gas phase and in equilibrium with the bulk concentration,
rCO2 = 1 k 2 respectively.
+
k2 k 2 k b[B] (8) The term y*CO2 can be calculated by
Moreover, a formation of the zwitterion has been shown to be PyCO
much slower than the reverse rate to MEA and CO2, and is * =
yCO 2

considered to be the rate determining step, which can be dened


2 HeCO2 (13)
as where yCO2 is the CO2 mole fraction, and HeCO2 represent
k 2/(k 2 k b[B]) 1/k 2 Henrys law constant of CO2 in solution.
(9) The relationships among the overall volume mass transfer
Hence, the reaction rate in eq 7 can be simplied to the following coecient, the individual-phase coecients, eective interfacial
expression, where the reaction rate appears to be rst order with area, and enhancement factor can be given as follows:5
respect to concentration of amine and CO2: 1 1 HeCO2
= +
rCO2 = k 2[CO2 ][RNH 2] (10) K Ga v k Ga v EkLoa v (14)

Hikita et al.29 proposed the following correlation for the k2, which where av is the eective interfacial area, kG is the gas-phase mass
depends on the temperature: transfer coecient, koL is the physical liquid-phase mass transfer
coecient, and E is the enhancement factor.
2152 Substituting eq 14 into eq 12 and rearranging yields:
k 2 = 10.99
T (11)
a v EkLoP(HeCO2yCO yCO )
2 2
2.3. Physical and Thermodynamic Properties. The NCO2a v =
(1 + E HeCO2kLo/k Ga v )
physicochemical properties of the uids (i.e., solubility, density, (15)
viscosity, surface tension, and diusivity) are necessary for Considering an element of column with height (dZ), the mass
analyzing the CO2 absorption process because of their great balance can also be expressed as follows:
inuence on the mass transfer performance, especially on the
eective interface area and the liquid lm coecient. dYCO2
The density and viscosity of MEA solution were taken from NCO2a v = G I
dZ (16)
the work of Weiland et al.30 who took CO2 loading into account,
and developed empirical equations to describe the density and where G1 and YCO2 represent molar ow rate of inert gas and
viscosity of the MEA solution as a function of temperature, mole ratio of CO2 in the gas phase, respectively.
concentration, and CO2 loading. However, when the concen- 3.2. Numerical Model of Mass-Transfer Correlations.
tration is higher than 40 wt % MEA, the viscosity correlation of Up to the present, extensive empirical or semiempirical
Weiland et al.30 becomes inapplicable, and therefore we need to correlations have emerged for various random and structured
seek a more suitable correlation. Hsu and Li31 developed a packings. These correlations are based on a huge number of
RedlichKister equation-based correlation for the estimation of experimental data and theoretical assumptions and, therefore,
the viscosity, which can be used for MEA concentrations higher their application scope, precision, and stability sometimes vary
than 40 wt %. The predicted results of viscosity showed good considerably. Among these correlations, the mass-transfer
agreement with experimental data, with an average deviation at correlations proposed by Onda et al.36 and Henriques de Brito
about 1.0%. The physical solubility and diusivity of CO2 can be et al.37 are two widely used models for random and structured
calculated via the N2O analogy because of the similarity of the packings, respectively.5 The Ondas correlations, which were
molecular weight, chemical conguration, volume, and electronic developed based on the experiment data of Raschig rings, Berl
structure between CO2 and N2O.3234 In addition, the surface saddles, Pall ring, Spheres packing, and Rode packing, have been
tension of the aqueous MEA solution was obtained from the highly recommend by many publications such as Perrys
correlation developed by Vazquez et al.35 Handbook5 to apply to various random packings. These
4415 dx.doi.org/10.1021/ie403259g | Ind. Eng. Chem. Res. 2014, 53, 44134423
Industrial & Engineering Chemistry Research Article

correlations were present based on the assumption that the al.,38 which deviates by less than 3% from the most accurate
wetted surface on packing pieces is equal to the eective gas solution, was therefore adopted here:
liquid interface. The eective surface area (ae), the physical
liquid-side mass transfer coecient (koL), and the gas phase mass E = 1 + {1/[1/(E 1)1.35 + 1/(E1 1)1.35 ]}(1/1.35)
transfer coecient (kG) in packed columns are shown in eqs (23)
2224.
c 0.75 L L2a t
0.1 0.05 E = 1 + [(DamineCamine)/ABDCO2CCO2] (24)
ae
= 1 exp 1.45
at L a tL L 2 g
E1 = Ha /tanh( Ha ) (25)
0.1
L 2
Ha = (k 2DamineCamine)/(kLo)2 (26)
L La t
(17)
where E is the enhancement, E is the enhancement factor for
0.33
L 0.67
0.5 instantaneous reaction, E1 is the enhancement factor of a pseudo-
kLo = 0.0051 L L (a tdP)0.4 rst-order reaction, AB is the stoichiometry of reactants for
L g aeDL L DL absorption reaction, CCO2 is the CO2 concentration in liquid
(18) phase, Camine is the concentration of amine solution in the bulk
liquid, Damine and DCO2 are the diusivity of amine and CO2 in
D G 0.7 u 0.33 liquid, respectively, and Ha is the Hatta number.
k G = 5.23 L G (a td p)2.0
RT a t 3. 4. Articial Neural Networks Modeling. As is well-
G G G
D (19)
known, articial neural networks (ANNs) have the powerful
where ae and at are the eective and total surface area of packing, ability to recognize underlying linear and nonlinear relationships
respectively, c and L are the critical surface tension of packing among input and output data.39 The basic topology structure of
material and the surface tension of liquid, respectively, L and G an ANN is composed of a number of simple and highly
are the liquid and gas phases ow rate, L and G are the viscosity interconnected processing elements called neuron units which
of liquid phase and gas phase, respectively, L and G are the are organized within multilayers (i.e., one input layer, one output
liquid density and gas density, respectively, g is the gravitational layer, and at least one hidden layer). Each neuron unit is
constant, DL and DG are the diusion of the liquid and gas phase, completely interconnected with neuron units in neighboring
respectively, dp denotes nominal size of packing. layers by means of direct communication linked with associated
The correlations developed by Henriques de Brito et al.37 were weights. ANNs work by receiving limited representative samples
proposed for predicting ae, koL, and kG for the Sulzer structured data, learning complex relationships among these data in certain
packing Mellapak 125.Y, 250.Y, and 500.Y. under industrial-scale algorithms, adjusting the weights of the internal connections to
operating conditions. In their model, the eective mass transfer minimize errors between the network output and target output,
area of the packing can be considerably higher than the dened and responding to unseen input data to predict targeted output
geometric area of the packing owing to liquid ow instabilities, within the range covered by the training examples. This type of
leading to waves, lm detachment, and droplet formation procedure is highly amendable to predicting mass transfer
between the sheets. performance of CO2 absorption in an amine solution. The two
ANN approaches BPNN and RBFNN were considered for mass
ae uL 0.3 transfer performance prediction application.
= 0.465 L 3.4.1. Back-Propagation Neural Network (BPNN). BPNN, as
at a tL (20) one of the most classic ANNs,40 is based on the steepest gradient
descent method with attempt to minimize the error (normally
(DL k 2CA,L)0.5 presented in term of root-mean-square of percentage error,
kLo = RMSPE) between the input and output values. In a BPNN, a
Ha (21)
portion of input/output pattern pairs selected from a sample data
set are used to train the network, and then the network iteratively
DG G uGdh G
0.8 0.3

k G = 0.0338P adjusts its connection weights and bias values according to the
dh G cos 45 G DG (22) feedback errors, which are calculated between the network
output and actual output. After being well trained, the BPNN can
where uL and uG are the supercial velocity of liquid and gas then be used to predict the unseen target parameters. Figure 1
ows, respectively, CA,L is the concentration of solute A in the shows the general architecture of a three-layer BPNN with
bulk liquid, Ha is the Hatta number, P is the pressure, and dh is multiple input nodes and a single output node.
the hydraulic diameter of packing. The training procedures of BPNN algorithm can be
3.3. Enhancement Factor. In comparison with pure summarized as follows:
physical mass transfer, chemical reaction-accompanied mass (1) Collecting all required data, and selecting a portion of
transfer would greatly enhance the absorption rate beyond pure input/output pattern pairs from the sample data set.
physical diusion because the reacting species is consumed (2) Normalizing all the data (scaled data between 0 and 1) in
within liquid-lm very quickly thereby preserving a very great the program in order to decrease the sample distribution range.
driving force. The enhancement factor (E) refers to the increase This is one of the most suitable methods for data preprocessing,
in mass transfer due to the chemical reaction relative to purely especially when the values of these input/output variables dier
physical absorption. The explicit equation presented by Wellek et by more than 1 order of magnitude.
4416 dx.doi.org/10.1021/ie403259g | Ind. Eng. Chem. Res. 2014, 53, 44134423
Industrial & Engineering Chemistry Research Article

ek = (dk Ok )f k (32)

f k = 1 (33)
where ek is the vector of errors for each output neuron unit, dk is
the kth desired output value, f k is the derivative of the activation
function (i.e., purelin function) for output nodes.
Then, the hidden layer errors become:
n
ej = f h ( wkjek )
k=1 (34)

f h = hj(1 hj) (35)


where ej is the vector of errors for each hidden layer neuron unit,
f h is the derivative of the activation function (i.e., logistic
function) for hidden nodes.
Figure 1. The general architecture of a three-layer BPNN with multiple (6) Adjusting the weights and biases in the output layer:
inputs and single output, where I1, I2, ..., In and Ok represent the input
vector and the output vector, respectively, x1, x2, ..., xn are the normalized wkj(t + 1) = wkj(t ) + rek + (wkj(t ) wkj(t 1)) (36)
input vector, w1j, w2j, ..., wnj are weight factors associated with the inputs
to the node, and bj is the threshold connecting the hidden and output b k(t + 1) = b k(t ) + rek (37)
layers.
where r is the learning rate, is the momentum factor, and t is the
I Imin calculation period.
xi = i
Imax Imin (27) These calculation steps will be repeated constantly to train the
th th
model until the output layer errors are acceptable for each
where xi is the i normalized input variable; Ii is the i original pattern and neuron. On the other hand, although overtraining
input variable; Imax and Imin are the maximum and minimum the BPNN model makes training data t well, it will reduce the
among all original input variables, respectively. generalization ability of the model since the random noise from
(3) Transferring the net input signal to the jth node in the new data (for testing) will be quite dierent from that of training
hidden layer by summing the weight and the bias, and calculating data. A common solution is to set a maximum number of
the output of the jth node in the hidden layer. The logistic sigmoid iterations to stop the training process early even if setting the
activation function (range between 0 and 1), which is the most training goal is not achieved.
popular function in BPNN because of its powerful ability to Although the BP algorithm is successful in training a feed
simulate the nonlinear relations, was used between the input- forward network, it is subjected to some drawbacks of slowly
hidden layers. learning convergent rate and easily converging to local minimum.
n In addition, the selections of step size, learning rate, momentum,
nj = wijxi + bj and the number of hidden neurons largely depend on trial-and-
i=1 (28) error and experience. Hornik et al.41 argued that a BP network
1 with only one hidden layer and sucient neurons can
hj = fh (n j) = approximate any nonlinear function to an arbitrary degree of
1 + exp( n j) (29) accuracy.
where subscripts i and j represent the input and hidden layer 3.4.2. Radial Basis Function Neural Network (RBFNN).
nodes, respectively, x is the normalized input, wij is weight linked RBFNN enjoys much praise in recent years due to its remarkable
the ith input node to the jth hidden layer node, and b is the bias properties of (i) the powerful abilities for generating multivariate
between the input and hidden layers, h is the vector of hidden- nonlinear mapping functions, (ii) structure determination by
layer neurons, f h(nj) is the logistic sigmoid activation function. self-adaptation, and (iii) the independence of output on initial
(4) Transferring the net hidden layer signal to the kth node in weight.39,42 RBFNN is a feed-forward network with three layers:
the output layer by summing the weights and the biases, and one input layer, one output layer, and only one hidden layer.
calculating the output of the kth node in the output layer using the Figure 2 shows the structure of a basic RBFNN with multi-input
purelin transfer function. and single output.
The inputoutput mapping undergoes two steps: (i) the m-
nk = wkjxj + b k dimensional input vector I1, I2, ..., Im collects the input
j (30) information and nonlinearly transforms from the input layer to
the hidden layer, and (ii) the hidden layer consisted of n hidden
Ok = fk (nk ) = nk (31) nodes linearly transforms information from hidden layer to
output layer. The connections between the input and hidden
where subscript k represents the output layer, wkj is the weight layers and the connection between the hidden layer and output
linked the jth hidden layer node to the kth output layer, and bk is layer are called centers and weights, respectively.
the threshold linked the hidden to output layers. Ok is the output
of output-layer neurons. f k (nk) is a linear activation function. K
(5) Calculating the errors between the calculated and the Ok = w k k (I ) + b
desired outputs: k=1 (38)

4417 dx.doi.org/10.1021/ie403259g | Ind. Eng. Chem. Res. 2014, 53, 44134423


Industrial & Engineering Chemistry Research Article

values greatly dier from the targeted values, one has to readjust
many times the sample data, the conguration parameters, or the
network topological structure, until the error between model
output and target value is acceptable.
Step 4: After being well trained, the RBFNN model can then
be used to predict the target parameters.
In RBFNN, it is very important to determine the number of
neuron in the hidden layer because it aects the complexity and
the approximating capability of the network. Too many neurons
in the hidden layer may lead to poor generalization or overtting,
while insucient neurons cannot learn the data adequately. In
addition, the position of the centers in the hidden layer also
aects the network performance signicantly, while the non-
linear transfer functions used by RBFNN have little inuence.
3.5. Data collection. The building of a predictive model
requires a series of representative data sets that should factually
and eectively reect the nonlinearities, complexities, and
Figure 2. The general architecture of a three-layer RBFNN with
multiple inputs and single output, where I1, I2, ..., Im and Ok denote the intricacies of the targeted system behavior. In this paper, ve
input and output of the network, k(I) is the radial basis function, w = databases of MEA-based mass transfer results were collected
(w1, w2, ..., wK) is the connecting weights between the hidden neuron and from the work of Tontiwachwuthikul et al.,43 and deMon-
the output layer, and b is the bias value. tigny.44,45 The detailed parameters of the packed columns and
the operational variables are presented in Table 1.
Tontiwachwuthikul et al.43 carried out a number of experi-
where I1, I2, ..., Im and Ok denote the input and output of the
ments for the absorption of air-diluted CO2 into aqueous MEA
network, m is the number of input variables, w = w1, w2, ..., wK) is
solution. The absorber column was packed with 12.7 mm
the connecting weights between the hidden neuron and the
ceramic Berl saddles with a total packing height of 6.55 and 0.1 m
output layer, n is the number of hidden neurons, k(I) is the
in diameter. DeMontigny44 reported the absorption data of CO2
radial basis function, b is the bias value.
absorption into ultrahighly concentrated MEA solution over the
The activation function usually chosen for the hidden node is a
range of 2.98 kmol/m3 to 9.02 kmol/m3 in a pilot plant
Gaussian function:
absorption column containing 16 mm Pall rings, IMTP#15
k(I ) = exp( || I k ||2 /k2) (39) random packing, and 4A Gempack structured packing. DeMon-
tigny45 also conducted CO2 absorption by MEA in the
where k denotes the center vector of the kth hidden neuron, I absorption column by using the Sulzer DX type packing. The
k is the Euclidean distance between I and k, and k is the experiments were performed in a small scale pilot-plant
radius or width of the kth hidden neuron. absorption column of 2.35 m packing height and 0.028 m in
Like BPNN, the forecast by means of RBFNN rst determines diameter. All these experiments were conducted in a counter-
the structure of the RBF network and the inputoutput data set current mode in order to provide a larger mass transfer force for
based on the practical problems, so as to achieve the the gas and liquid phases contacting with each other. Once the
corresponding mapping relationships. Within the Matlab steady-state conditions were reached, the CO2 concentrations in
program, the training procedures of the RBF algorithm can be the gas phase and solution temperatures were collected along the
summarized as follow: column height for dierent gas/liquid ratios, inlet CO2, and
Step 1: Design the inputoutput pattern, mainly including (i) solution concentration.
the selection of key parameters, (ii) the design of input and
output variables, and (iii) the preprocessing of sample data.
Step 2: Determine the center and width of the radial basis 4. RESULTS AND DISCUSSION
function and the weights information between hidden layer and For the CO2 absorption with chemical reaction in a packed
output layer. column with a special packing, 14 factors could have potential
Step 3: Call the newrbe function to create the RBFNN, and call impact on the hydromechanics and mass transfer performance.
the sim function to test the established network. If the output These factors are inert gas ow rate (GI, kmol/(m2h)), liquid

Table 1. The Parameters of the Columns and the Operational Variables

Tontiwachwuthikul et al.43 deMontigny44 deMontigny45


diameter of the column (mm) 100 100 28
hight of packing (m) 6.15 2.40 2.35
packing Berl saddles Pall ring, IMTP, 4A Gempack Sulzer DX
gas phase CO2 concentration (%) 11.519.2 5.319.9 13.9214.54
inert gas load (kmol/m3h) 39.9653.28 31.7972.58 30.730.9
liquid ow rate (m3/m2h) 9.513.5 6.9423.96 5.412.8
solvent concentration (kmol/m3) 2.03.8 2.989.02 1.03.0
CO2 loading (mol CO2/mol amine) 00.237 0.1050.228 0.0300.195
feed temperature (K) 19.020.0 22.528.3 17.621.7
pressure (atm) 1 1 1

4418 dx.doi.org/10.1021/ie403259g | Ind. Eng. Chem. Res. 2014, 53, 44134423


Industrial & Engineering Chemistry Research Article

ow rate (L, m3/(m2h)), solution concentration (C, kmol/m3), BPNN and RBFNN as well as those predicted values obtained
CO2 loading in solution (, mol/mol), CO2 mole fraction (yCO2), from the correlation proposed by Onda et al.36 for the MEA
temperature (T, K), reaction kinetics (k2, kmol/(m2hkPa)), CO2 absorption in a column packed with Berl Saddles. In this
liquid and gas phase density (L and G, kg/m3), liquid and gas simulation, the AAD, RMSPE, and R for BPNN and RBFNN
viscosity (L and G, cp), diusivity of CO2 in liquid and gas were found to be 1.8% and 0.7%; 2.7% and 1.7%; 0.996 and
phases (DL and DG, m2/h), and liquid surface tension (L, mN/ 0.999, respectively, which are favorably acceptable for mass
m). In the above 14 parameters, only 6 parameters (G, L, C, , transfer study. By contrast, the AAD, RMSPE, and R for the
yCO2 and T) are independent, and the remaining eight can be correlations of Onda were 25.9%, 32.3%, and 0.878, respectively.
Therefore, these two ANN models were found to greatly
obtained by a combination of the other parameters. outperform the correlations of Onda et al.36 especially in low
By means of BPNN and RBFNN models, this paper focuses on NCO2av value conditions, in which the relative error for BPNN
the prediction of MEA-based CO2 absorption in packed columns and RBFNN were much smaller than that of the Ondas
containing Berl saddles, Pall rings, IMTP#15 random packing, correlations. In addition, the RBFNN performed slightly better
and 4A Gempack, Sulzer DX structured packing, respectively. than the BPNN, probably due to two reasons: (i) the drawback of
The predicted results were validated with experimental results as BPNN easily converges to a local minimum of the error surface
well as compared with numerical correlations from available and thus the global minimum is not found; and (ii) to achieve the
literature, so as to verify the eectiveness and accuracy of the predicted goal under the same accuracy requirement, RBFNN
BPNN and RBFNN models. The simulated data in this work generally possesses more neurons in the hidden layer than
were derived from original experimental results; that is, multiple BPNN, thereby increasing the freedom of the network, which
sample points in each experimental run, along with the value for helps to reect the complexity of the system. Similar results were
the packed height in which , yCO2, and T varied greatly while L, observed in the cases of the Pall ring, IMTP, 4A Gempack, and
GI, and C were assumed to be constant. Note that those data DX packing in that the predicted results obtained from the
containing extremely low and yCO2 were left out of the model present work were found to (i) t favorably with the
prediction to avoid as far as possible misuse of experimental data experimental results and (ii) be better than those predicted
with high relative error due to almost inevitably operating and from the correlations proposed by Onda et al.36 as can be seen in
systematic errors. The BPNN and RBFNN were developed Figure 4ad. Additionally, the RBFNN was observed to perform
within the Matlab R2010a program with the neural network slightly better than BPNN. The evaluation parameters are
toolbox. presented in Table 2.
The mass ux per unit volume per unit interfacial volume According to the parity charts as presented in Figures 3 and 4,
(NCO2av, kmol/(m3h)) is convenient and eective in assessing it can be found that the BPNN and RBFNN have the capacity to
predict NCO2av with adequate accuracies for several packings
the mass transfer performance of any solvent and packing in a
packed column, and is also a key parameter for designing the (Berl saddles, Pall ring, IMTP, 4A Gempack, and DX). However,
packed columns. The design methodology and steps can be it should be acknowledged that the comparisons presented in
found in the work of Liang et al.9 who gave a detailed review of Figures 3 and 4 are unfair. This is because the correlations
the design and modeling for postcombustion CO2 capture proposed by Onda et al. and Henriques de Brito et al. were
systems using reactive solvents. So, the prophase of this work is developed on the basis of the data of multiple packings, while the
ANN models performing in this work were well trained for each
to predict the NCO2av of MEACO2 absorption. The predicting
type of packing. To further validate the outstanding performance
results were evaluated by average absolute deviation (AAD), of BPNN and RBFNN, we attempt to develop individually two
root-mean-square of percentage error (RMSPE), and the models for random packings (by combining all data of three
correlation coecient (R). These evaluation parameters display random packing) and structured packings (by combining all data
how closely the model prediction matched the experimental of two structured packing) by using BPNNs and RBFNNs. As is
results. well-known, the total packing area at is an important parameter in
Figure 3 shows a parity plot of the experimental values of mass transfer prediction because it will aect the hydromechanics
NCO2av compared with those predicted values obtained from and mass transfer performance greatly. In the modeling of the
CO2 absorption into aqueous MEA in packed columns for
dierent packings, besides the six key parameters mentions above
(i.e., G, L, C, , yCO2 and T), at should be considered as another
input value for the model designing since it varies from packing
to packing. Figure 5 shows the comparisons of predicted NCO2av
values from BPNN and RBFNN against experimental values for
random and structured packings.
As can be seen in Figure 5, the predicted results from BPNN
and RBFNN are found to be very close to the experimental values
with AAD of 2.9% and 1.7%, respectively for random packings;
and 4.8% and 3.6%, respectively for structured packings.
Needless to say, the approaches of the ANNs are well suited
for application in the estimation of mass ux. In the present work,
the AAD values for structured packings were found to be higher
than that of random packings. This could be because of the less
Figure 3. Comparisons between predicted and experimental NCO2av for experimental data (22 runs: 14 runs for 4A Gempack, 8 runs for
CO2 into MEA in column containing Berl saddles. DX packing) for structured packings to be used for building a
4419 dx.doi.org/10.1021/ie403259g | Ind. Eng. Chem. Res. 2014, 53, 44134423
Industrial & Engineering Chemistry Research Article

Figure 4. Comparisons between predicted and experimental NCO2av for CO2 into MEA in column containing (A) Pall ring, (B) IMTP, (C) 4A
Gempack, and (D) DX packing.

Table 2. Statistical Analysis of Correlations (corr) on the Mass Flux


number of data sets AAD (%) RMSPE (%) R
data source packing total training testing corr BPNN RBFNN corr BPNN RBFNN corr BPNN RBFNN
Tontiwachwuthikul Berl saddles 160 120 (75%) 40 (25%) 25.9 1.8 0.7 32.3 2.7 1.7 0.878 0.996 0.999
et al.43
deMontigny44 Pall ring 196 147 (75%) 49 (25%) 18.2 2.4 0.1 25.7 4.5 0.1 0.966 0.999 1.000
IMTP 764 573 (75%) 191 (25%) 24.1 2.6 0.2 29.5 4.0 0.6 0.916 0.998 1.000
4A 208 156 (75%) 52 (25%) 22.1 5.3 3.4 26.7 10.8 4.2 0.892 0.989 0.998
Gempack
deMontigny45 DX 68 51 (75%) 17 (25%) 25.4 6.4 4.9 37.1 8.5 7.3 0.854 0.981 0.988

comprehensive neural network to nd the relationship among values, while dyCO2/dZ and dT/dZ are considered as
the key parameters, while there are 52 runs (10 runs for Berl output values.
saddles, 7 runs for Pall ring, and 35 runs for IMTP) for random
packings that can be used. (ii) Assume a CO2 mole fraction in the outlet as yout
CO2.

In most cases, the CO2 mole fraction and temperature proles (iii) Estimate the rst set of y1CO2 and T1 at a certain height of Z1,
along the height of the column can reect the prediction results which should be very close to the column top. In this case,
more directly and eectively. To simulate the CO2 concentration the distance Z1 between Z1 and the column top (Ztop)
and temperature proles, only inlet gas/liquid ow conditions can be considered as dierential height dZ.
are generally known, not the outlet conditions. Thus, the (iv) Calculate corresponding CO2 loading a1 at a column
problem of two-point boundary value would then occur. To
height of Z1 based on mass balance.
obtain an output, the shooting method is recommended.16,43,46
In this work, the algorithm of the formulated ANNs model to (v) Put the values of the (y1CO2 + yout 1
CO2)/2, (a + ain)/2, (T +
1

calculate the CO2 concentration and temperature proles was Tin)/2; and the other four factors into the established
done as follows: model, then obtain predicted values of (y1CO2/Z1)pre
(i) Build an ANN model based on the experimental data in and (T1/Z1)pre, where y1CO2 = y1CO2 yout CO2 and T =
1

which at, G, L, C, , yCO2, and T are considered as input T1 Tin


4420 dx.doi.org/10.1021/ie403259g | Ind. Eng. Chem. Res. 2014, 53, 44134423
Industrial & Engineering Chemistry Research Article

Figure 5. Comparisons between predicted and experimental NCO2av for


CO2 into MEA in a column containing (A) random packings, (B) Figure 6. Concentration (A) and temperature (B) proles along the
structured packings. height of the column for CO2 absorption into MEA in a column packed
with IMTP random packing (dots are experimental data obtained from
(vi) Compare the values of the predicted (y1CO2)pre and (T1)pre deMontigny44 (run B-023), lines are the predicted results obtained in
the present work).
with those of the estimated (y1CO2)est and (T1)est, if the
dierences for both values are acceptable, then proceed;
otherwise, go back to Step (ii).
(vii) Iteratively calculate until (ynCO2)pre yinCO2.
(viii) End the algorithm.
Figures 6 and 7 show comparisons of the CO2 mole fraction
and temperature proles between the predicted results obtained
from BPNN and RBFNN and the experimental results (runs B-
023 and C-003) obtained from the work of deMontigny.44 The
AAD values of the CO2 mole fraction and temperature proles
between the simulation results from BPNN and RBFNN; and
experimental results for both random and structured packings
were found to be in the acceptable range of less than 10% as
shown in Figures 6 and 7. This observation clearly shows that the
proposed two ANNs models can be eectively used for
predicting CO2 concentration and temperature proles along
the height of the column for CO2 absorption, the column packed
with randomly packing (Berl saddles, Pall rings, and IMTP) and
structured packing (4A Gempack and DX packing).

5. CONCLUSIONS
In this work, we employed two articial neural networks, BPNN
and RBFNN, to the simulation of the CO2 absorption into
aqueous MEA in a packed column. The predicted results of mass
ux obtained from this work via BPNN and RBFNN modeling
were compared with the numerical correlations reported in the
literature. It was clearly observed that the ANN modeling Figure 7. Concentration (A) and temperature (B) proles along the
(BPNN and RBFNN) provided better predicted results. It was height of the column for CO2 absorption into MEA in a column packed
also found that the RBFNN modeling performed slightly better with 4A Gempack structured packing (dots are experimental data
than BPNN modeling. In addition, the CO2 concentration and obtained from deMontigny44 (run C-003), lines are the predicted results
temperature proles along the height of the column were obtained in the present work).

4421 dx.doi.org/10.1021/ie403259g | Ind. Eng. Chem. Res. 2014, 53, 44134423


Industrial & Engineering Chemistry Research Article

successfully predicted using ANN with a satisfactory AAD of less t = calculation period, s
than 10%. T = temperature, K

AUTHOR INFORMATION
Corresponding Author
wij = weights connecting the ith input node to the jth hidden
layer node for BPNN
wkj = weights connecting the jth hidden node to the kth output
*Tel.: +86-13618481627. Fax: +86-731-88573033. E-mail: layer node for BPNN
zwliang@hnu.edu.cn. wk = connecting weights for RBFNN
xi = the ith normalized input variable for BPNN
Notes yA* = mole fraction of solute A in equilibrium with CA,L
The authors declare no competing nancial interest.

yA,G = mole fraction of solute A in the bulk gas-phase


yA,i = mole fraction of solute A at the interface
ACKNOWLEDGMENTS YA = mole ratio of component A in gas bulk
The nancial supports from the National Natural Science Z = height, m
Foundation of China (No. 21276068, No. U1362112 and No. Greek Symbols
21376067); National Key Technology R&D Program (No. k() = radial basis function
2012BAC26B01, No.2014BAC18B04); Innovative Research f h() = logistic sigmoid activation function from input layer to
Team Development Plan-Ministry of Education of China (No. hidden layer for BPNN
IRT1238); Specialized Research Fund for the Doctoral Program f h = local slope of the node activation function for hidden
of Higher Education (No. 20130161110025); Innovative nodes for BPNN
Research Program for Graduate Student of Hunan Province, f k() = purelin transfer function from input layer to hidden
China (No. CX2013B158); and Chinas State Project 985 in layer for BPNN
Hunan UniversityNovel Technology Research & Develop- f k = local slope of the node activation function for output
ment for CO2 Capture are gratefully acknowledged.

nodes for BPNN


i = viscosity of phase i, cp
NOMENCLATURE k = center vector of the kth hidden neuron for RBFNN
= liquid CO2 loading, mol CO2/mol amine i = density of phase i, kg/m3
av = eective interfacial area, m2/m3 = momentum factor
at = total packing surface area, m2/m3 c = critical surface tension of the packing material, mN/m
b = bias value L = surface tension of liquid phase, mN/m
BPNN = back-propagation neural network k = radius or width of the kth hidden neuron for RBFNN
C = solution concentration, kmol/m3
CA,i = concentration of component A at the interface, kmol/m3
CA,L = concentration of component A in the bulk liquid, kmol/
REFERENCES
(1) Quadrelli, R.; Peterson, S. The energy-climate challenge: Recent
m3 trends in CO2 emissions from fuel combustion. Energy Policy 2007, 35,
dh = hydraulic diameter, m 5938.
dk = desired output value for BPNN (2) Rao, A. B.; Rubin, E. S. A technical, economic, and environmental
dp = nominal size of packing, m assessment of amine-based CO2 capture technology for power plant
greenhouse gas control. Environ. Sci. Technol. 2002, 36, 4467.
Di = diusivity of phase i, m2/h (3) Kohl, A. L.; Nielsen, R. B. Gas Purication, 5th ed.; Gulf Publishing
ek = vector of errors for BPNN Company: Houston, TX, 1997.
E = enhancement factor (4) Rochelle, G. T. Amine scrubbing for CO2 capture. Science 2009,
E1 = enhancement factor for a pseudo-rst-order reaction 325, 1652.
Ei = enhancement factor for instantaneous reactions (5) Perry, R. H.; Green, D. Perrys Chemical Engineers Handbook, 7th
GI = inert gas ow rate, kmol/(m2h) ed.; McGraw-Hill Book Company: New York, 1997.
hi = the ith input variable for BPNN (6) Strigle, R. F. Random Packings and Packed Towers, 2nd ed.; Gulf
Ha = Hatta number Publishing Company: Houston, TX, 1987.
He = Henrys law constant, m3kPa/kmol (7) Kuntz, J.; Aroonwilas, A. Performance of spray column for CO2
Ii = the ith input variable capture application. Ind. Eng. Chem. Res. 2008, 47, 145.
(8) deMontigny, D.; Tontiwachwuthikul, P.; Chakma, A. Using
kG = gas side mass transfer coecient, kmol/(m2hkPa) polypropylene and polytetrafluoroethylene membranes in a membrane
KGav = gas-phase volumetric overall mass transfer coecient, contactor for CO2 absorption. J. Membr. Sci. 2006, 277, 99.
kmol/(m3hkPa) (9) Liang, Z.; Sanpasertparnich, T.; Tontiwachwuthikul, P.; Gelowitz,
koL = physical liquid-phase mass transfer coecient, kmol/(m2 D.; Idem, R. Design, modeling and simulation of post-combustion CO2
hkPa) capture systems using reactive solvents. Carbon Manage. 2011, 2, 265.
ki = second order forward rate coecient for reaction i, m3/ (10) Wang, G. Q.; Yuan, X. G.; Yu, K. T. Review of mass-transfer
(kmolh) correlations for packed columns. Ind. Eng. Chem. Res. 2005, 44, 8715.
ki = second order reverse rate coecient for reaction i, m3/ (11) Piche, S.; Grandjean, B. P. A.; Iliuta, I.; Larachi, F. Interfacial mass
(kmolh) transfer in randomly packed towers: A confident correlation for
L = liquid ow rate, m3/(m2h) environmental applications. Environ. Sci. Technol. 2001, 35, 4817.
(12) Freguia, S.; Rochelle, G. T. Modeling of CO2 capture by aqueous
NCO2av = CO2 absorption rate per unit area, kmol/(m3h) monoethanolamine. AIChE J. 2003, 49, 1676.
Ok = output vector for BPNN (13) Martinez, I.; Murillo, R.; Grasa, G. Integration of a Ca looping
P = total system pressure, kPa system for CO2 capture in existing power plants. AIChE J. 2011, 57,
PCO2 = partial pressure of CO2, kPa 2599.
r = learning rate (14) Mores, P.; Scenna, N.; Mussati, S. CO2 capture using
RBFNN = radial basis function network monoethanolamine (MEA) aqueous solution: Modeling and optimiza-

4422 dx.doi.org/10.1021/ie403259g | Ind. Eng. Chem. Res. 2014, 53, 44134423


Industrial & Engineering Chemistry Research Article

tion of the solvent regeneration and CO2 desorption process. Energy with structured packings and with ceramic rings. Ind. Eng. Chem. Res.
2012, 45, 1042. 1994, 33, 647.
(15) Ji, X.; Kritpiphat, W.; Aboudheir, A.; Tontiwachwuthikul, P. Mass (38) Wellek, R. M.; Brunson, R. J.; Law, F. H. Enhancement factors for
transfer parameter estimation using the optimization technique: Case gas-absorption with second-order irreversible chemical reaction. Can. J.
study in CO2 absorption with chemical reaction. Can. J. Chem. Eng. Chem. Eng. 1978, 56, 181.
1999, 77, 69. (39) Hagan, M.; Beale, M.; Demuth, H. Neural Network Design; PWS
(16) Aboudheir, A.; Tontiwachwuthikul, P.; Idem, R. Rigorous model Publishing Company: Boston, MA. 1996.
for predicting the behavior of CO2 absorption into AMP in packed-bed (40) Rumelhart, D. E.; Hinton, G. E.; Williams, R. J. Learning
absorption columns. Ind. Eng. Chem. Res. 2006, 45, 2553. representations by back-propagating errors. Nature 1986, 323, 533.
(17) Liu, G. B.; Yu, K. T.; Yuan, X. G.; Liu, C. J. New model for (41) Hornik, K.; Stinchcombe, M.; White, H. Multilayer feedforward
turbulent mass transfer and its application to the simulations of a pilot- networks are universal approximators. Neural Network 1989, 2, 359.
scale randomly packed column for CO2NaOH chemical absorption. (42) Park, J.; Sandberg, I. W. Universal approximation using radial-
Ind. Eng. Chem. Res. 2006, 45, 3220. basis-function networks. Neural Comput. 1991, 3, 246.
(18) Hanley, B.; Chen, C. C. New mass-transfer correlations for packed (43) Tontiwachwuthikul, P.; Meisen, A.; Lim, C. J. CO2 absorption by
towers. AIChE J. 2012, 58, 132. NaOH, monoethanolamine and 2-amino-2-methyl-1-propanol solu-
(19) Lek, S.; Delacoste, M.; Baran, P.; Dimopoulos, I.; Lauga, J.; tions in a packed column. Chem. Eng. Sci. 1992, 47, 381.
Aulagnier, S. Application of neural networks to modeling nonlinear (44) deMontigny, D. Carbon dioxide absorption studies using ultra-
relationships in ecology. Ecol. Model 1996, 90, 39. highly concentrated monoethanolamine solutions. M.A.Sc. Thesis,
(20) Chen, J.; Wong, D. S. H.; Jang, S. S.; Yang, S. L. Product and University of Regina, Regina, Saskatchewan, Canada, 1998.
process development using artificial neural-network model and (45) deMontigny, D. Comparing packed columns and membrane
information analysis. AIChE J. 1998, 44, 876. absorption for CO2 capture. Ph.D. Thesis, University of Regina, Regina,
(21) Piche, S.; Iliuta, I.; Grandjean, B. P. A.; Larachi, F. A unified Saskatchewan, Canada, 2004.
approach to the hydraulics and mass transfer in randomly packed towers. (46) Pandya, J. Adiabatic gas absorption and stripping with chemical
Chem. Eng. Sci. 2001, 56, 6003. reaction in packed column. Chem. Eng. Commun. 1983, 19, 343.
(22) Byvatov, E.; Fechner, U.; Sadowski, J.; Schneider, G. Comparison
of support vector machine and artificial neural network systems for
drug/nondrug classification. J. Chem. Inf. Model 2003, 43, 1882.
(23) Huesken, D.; Lange, J.; Mickanin, C.; Weiler, J.; Asselbergs, F.;
Warner, J.; Meloon, B.; Engel, S.; Rosenberg, A.; Cohen, D.; Labow, M.;
Reinhardt, M.; Natt, F.; Hall, J. Design of a genome-wide siRNA library
using an artificial neural network. Nat. Biotechnol. 2005, 23, 995.
(24) Staudenmayer, J.; Pober, D.; Crouter, S.; Bassett, D.; Freedson, P.
An artificial neural network to estimate physical activity energy
expenditure and identify physical activity type from an accelerometer.
J. Appl. Physiol. 2009, 107, 1300.
(25) Demuth, H.; Beale, M.; Hagan, M. Neural Network Toolbox Users
Guide. The MathWorks, Inc.: Natick, MA, 1998.
(26) Caplow, M. Kinetics of carbamate formation and breakdown. J.
Am. Chem. Soc. 1968, 90, 6795.
(27) Danckwerts, P. V. The reaction of CO2 with ethanolamines.
Chem. Eng. Sci. 1979, 34, 443.
(28) Hagewiesche, D. P.; Ashour, S. S.; Al-Ghawas, H. A.; Sandall, O.
C. Absorption of carbon dioxide into aqueous blends of monoethanol-
amine and N-methyldiethanolamine. Chem. Eng. Sci. 1995, 50, 1071.
(29) Hikita, H.; Asai, S.; Ishikawa, H.; Honda, M. The kinetics of
reaction of carbon dioxide with monoethanolamine, diethanolamine,
and triethanolamine by rapid mixing method. J. Chem. Eng. 1977, 13, 7.
(30) Weiland, R.; Dingman, J.; Cronin, D.; Browning, G. Density and
viscosity of some partially carbonated aqueous alkanolamine solutions
and their blends. J. Chem. Eng. Data 1998, 43, 378.
(31) Hsu, C.; Li, M. Viscosities of aqueous blended amines. J. Chem.
Eng. Data 1997, 42, 714.
(32) Versteeg, G.; Van Swaaij, W. Solubility and diffusivity of acid gases
(CO2, N2O) in aqueous alkanolamine solutions. J. Chem. Eng. Data
1988, 33, 29.
(33) Wang, Y. W.; Xu, S.; Otto, F. D.; Mather, A. E. Solubility of N2O in
alkanolamines and in mixed solvents. Chem. Eng. J. 1992, 48, 31.
(34) Tsai, T.; Ko, J.; Wang, H.; Lin, C.; Li, M. Solubility of nitrous
oxide in alkanolamine aqueous solutions. J. Chem. Eng. Data 2000, 45,
341.
(35) Vazquez, G.; Alvarez, E.; Navaza, J. M.; Rendo, R.; Romero, E.
Surface tension of binary mixtures of water + monoethanolamine and
water + 2-amino-2-methyl-1-propanol and tertiary mixtures of these
amines with water from 25 to 50 C. J. Chem. Eng. Data 1997, 42, 57.
(36) Onda, K.; Takeuchi, H.; Okumoto, Y. Mass transfer coefficients
between gas and liquid phase in packed columns. J. Chem. Eng. Jpn. 1968,
1, 56.
(37) Henriques de Brito, M.; von Stockar, U.; Bangerter, A. M.; Bomio,
P.; Laso, M. Effective mass transfer area in a pilot plant column equipped

4423 dx.doi.org/10.1021/ie403259g | Ind. Eng. Chem. Res. 2014, 53, 44134423

Anda mungkin juga menyukai