discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/223185229
CITATIONS READS
19 229
4 authors, including:
Asghar Alizadehdakhel
Islamic Azad University of Rasht
13 PUBLICATIONS 173 CITATIONS
SEE PROFILE
All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Asghar Alizadehdakhel
letting you access and read them immediately. Retrieved on: 16 May 2016
International Communications in Heat and Mass Transfer 36 (2009) 850–856
CFD and artificial neural network modeling of two-phase flow pressure drop☆
Asghar Alizadehdakhel a, Masoud Rahimi a,⁎, Jafar Sanjari a, Ammar Abdulaziz Alsairafi b
a
CFD Research Center, Chemical Engineering Department, Razi University, Taghe Bostan, Kermanshah, Iran
b
Faculty of Mechanical Engineering, College of Engineering and Petroleum, Kuwait University, Kuwait
a r t i c l e i n f o a b s t r a c t
Available online 1 July 2009 A large number of experiments in a 2 cm diameter and 6 m length tube were carried out in order to study the
two-phase flow regimes and pressure drops in it. The two-phase flow in the experimental tube was modeled
Keywords: using commercial CFD code, Fluent 6.2. An Artificial Neural Network (ANN) with three inputs including gas
Two-phase flow and liquid velocities and tube slope was designed and trained to predict average pressure drop across the
Pressure drop tube. The comparison between CFD and ANN predictions of pressure drops with experimental measurements
CFD
shows that the CFD results are more accurate than the ANN evaluations for new conditions.
ANN
© 2009 Elsevier Ltd. All rights reserved.
Modeling
1. Introduction of these studies, the calculation began with prediction of the flow
pattern and then proposing an associated method of calculating the
Two-phase flow of gas and liquid in pipes frequently occurs in the liquid holdup. Consequently, the liquid holdup is used to determine
chemical and petroleum industries as well as natural gas transfer line. the two-phase friction factor. Comparative studies proved that these
While pressure losses in single-phase flow in pipes are accurately models are inconsistently performed as flow conditions change.
modeled with familiar expressions such as the Bernoulli and Navier– Therefore, selecting of the most appropriate flow correlation is quite
Stokes equations, accurate predictions of pressure loss in two-phase important in this category of research.
flow have been proved to be more challenging because of added Artificial neural networks are analytical tools that imitate the
complexities. The lower density and viscosity of the gas phase causes neural aspect of the human brain, whereby learning is based on
it to flow at a higher velocity than the liquid phase, a phenomenon experience and repetition rather than the application of rule-based
known as slippage. Further complicating matters are the variety of principles and formulas. An ANN consists of a layered network of
physical phase distributions that are characterized by flow regimes or neurons (nodes), with each neuron connected to a large number of
flow patterns. Detailed discussions of these patterns are given in the others. The input signal to the network is passed among the neurons,
literature [1]. with each neuron calculating its own output using weighting
In horizontal flows as gravity acts normally to flow direction, the associated with connections. Learning is achieved by the adjustment
separation of flow occurs. The respective flow regimes are bubble flow, of the weights associated with inter-neuron connections. ANNs
stratified, stratified-wavy, plug, slug and annular flow. provide capabilities such as learning, self-organization, generalization
In the vertical flows, the flow regimes can be listed as bubble, plug, and training; and are excellent for pattern recognition and trend
slug or churn, annular and wispy-annular flow. Many investigations prediction for processes that are non-linear, poorly-understood, and/
have been carried out to find the friction factor and consequently the or too complex for accurate first-principle mathematical modeling.
pressure drop across horizontal [2] and vertical [3] two-phase as well They seem ideal for applications to multiphase flow systems [8].
as multiphase flows [4] in pipes. Cai et al. [9] applied the Kohonen self-organizing neural network to
Similar flow regimes are defined for two-phase flow in inclined identify flow regimes in horizontal air–water flow. Osman and Aggour
pipes. However, for inclined flows, the most prominent feature is a [10] developed a new model for identifying different flow regimes and
preponderance of a stratified flow pattern in the downward inclined predicting the holdup of horizontal gas–liquid flow. His model was
pipe and a preponderance of an intermittent flow pattern (plug or based on three-layer back-propagation neural networks. Shippen and
slug) in upward inclined pipe [5,6]. Scott [11] presented another neural network model for prediction of
Many empirical correlations and mechanistic models have been liquid holdup in a two-phase horizontal flow. Data obtained from five
proposed to predict liquid holdup and pressure loss [7]. Almost in all independent studies were used in their modeling.
The shape of gas/liquid interface has a great importance in the
☆ Communicated by W.J. Minkowycz.
two-phase flow pressure drop in pipelines. Nigmatulin and Bonetto
⁎ Corresponding author. [12] studied the shape of a Taylor bubble in a vertical downward slug
E-mail address: masoudrahimi@yahoo.com (M. Rahimi). flow in a pipe. In their research, the phase distribution in the Taylor
0735-1933/$ – see front matter © 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.icheatmasstransfer.2009.05.005
A. Alizadehdakhel et al. / International Communications in Heat and Mass Transfer 36 (2009) 850–856 851
For horizontal and vertical setups, turning first to Fig. 2(a), the
figure shows that the average pressure drop in the horizontal setup
increases by increasing of GVN while the slope of the curves increases
by increasing the LVN. In addition, variation of average pressure drop
with gas velocity has a similar trend for different liquid velocities.
Fig. 2(b) reveals that the pressure drops of vertical flow at LVN = 0 are
almost simillar to those of the horizontal flow. Opening the liquid flow
causes different pressure drop trends as compared with those of
horizontal setup. Here, in low gas flow rates the liquid weight is the
major cause of pressure drop and the pressure drop decreases by
increasing of the gas flow rate. By more increase in the gas flow rate,
the major cause of pressure drop shifts to the friction between the
fluids and wall and also the friction between the two phases.
Therefore, in high gas flow rates, the pressure drop increases by
increasing the gas flow.
Pressure drops at the two-phase inclined flows are shown in Fig. 3.
The figure indicates that the way that the pressure changes in
downward two-phase flow has a similar pattern with those of
horizontal flow and at steeper setups lower pressure drops were
obtained. In addition, lower differences between pressure drops
happened by increasing of the gas flow rate. This can be explained by
the fact that the major cause of difference in pressure drops at
different tube slopes is the differences in their liquid flow shapes.
Fig. 2. Average pressure drop for the vertical and horizontal setups: a) horizontal and
b) vertical.
where Vsg and Vsl are gas and liquid superficial velocities, respectively.
ρl is the liquid density, g is the gravitational acceleration and σ is the Fig. 3. Average pressure drop for upward and downward two-phase flows, Pa m− 1:
surface tension. a) downward and b) upward.
A. Alizadehdakhel et al. / International Communications in Heat and Mass Transfer 36 (2009) 850–856 853
However, by increasing the downward slope, the liquid waves become fluids and the VOF model was employed. The VOF method relies on
smaller and the pressure drop decreases. Fig. 3(b) presents an the fact that two or more phases are not interpenetrating and for each
increasing trends in pressure drop by increasing the tube slope in additional phase the volume fraction of the phase must be added in
upward inclined layouts. The figure also reveals that the trend of the computation. In the VOF model, the sum of the volume fractions of
pressure changes in the upward flow becomes close to that of vertical all phases in each control volume is equal to one. In the modeling, the
setup by increasing of the slope. Another point in this figure is that in incompressible Navier–Stokes equations for velocity, u, and pressure,
contrast with downward flows, although the differences among p, are solved simultaneously. These equations are as follows [20]:
pressure drops of different slopes decrease by increasing of the
gasflow rate, a considerable differences remain among the pressure Continuity: j:u = 0 ð2Þ
drops in upward flows with different inclination angles. It can be
explaind by the fact that in the upward flows, even for high gas flow Au 1 1
rates, the gravitational force helps the formation of liquid waves, but Momentum: + j:ðuuÞ = − ½jp − j:ð2μSÞ + FSF ð3Þ
At ρ ρ
in the downward flows the gravitational force causes the vawe to
disapear. where, FSF is the continuum surface force (CSF) vector and S is the
deformation tensor given as follows:
3. CFD modeling
1 T
S= ju + ½ju : ð4Þ
3.1. Multiphase flow modeling 2
The commercial CFD package, FLUENT6.2 was used to model the In Eq. (3), the density (ρ) and viscosity (μ) of the fluid depend on
gas–liquid flow regimes and pressure drops in the tube. As far as the the volume fractions of each phase and they are calculated by the
interface between gas/liquid phases is important in predicting the following equations:
flow regime and pressure drop in the tube, applying the free surface
modeling is necessary. This is one of the challenging research areas for αρair μ air + ð1 − α Þρwater μ water
ρ = αρair + ð1 − α Þρwater and μ =
providing an efficient solver. Volume of Fluid (VOF) [18] and Levelset αρair + ð1 − α Þρwater
[19] approaches belong to two of the best possible implicit free surface ð5Þ
reconstruction methods, while particularly VOF is relatively simple to
treat topological changes of the interface and it is naturally where, α is the air volume fraction in the cell. The interface between
conservative. In the present work, water and air were used as working two phases was tracked by the volume fraction. Conservation of α can
Fig. 4. The comparison between experimental and CFD-predicted horizontal flows. a) GVN = 1.4 (Vsg = 0.2 ms− 1), LVN = 0.3 (Vsl = 0.05 ms− 1). b) GVN = 14.6 (Vsg = 2.38 ms− 1),
LVN = 0.3 (Vsl = 0.05 ms− 1). c) GVN = 87.3 (Vsg = 14.3 ms− 1), LVN = 2.6 (Vsl = 0.42 ms− 1).
854 A. Alizadehdakhel et al. / International Communications in Heat and Mass Transfer 36 (2009) 850–856
be represented by the interface mass balance using the following cases. The calculations were performed by combination of the PISO
equation: algorithm for pressure–velocity coupling and a second order upwind
calculation scheme for the determination of momentum and volume
Aα fraction. The convergence criterion was based on the residual value of
+ u:jα = 0: ð6Þ the calculated variables, namely mass, velocity components and
At
volume fraction. In the present calculations, the numerical computa-
tion was considered converged when the scaled residuals of the
The cell phase is gas where α = 1, while α = 0 means that the
different variables are lowered by four orders of magnitude.
whole volume has been occupied by the liquid. It can be concluded
that the gas/liquid interface exists in the regions that α lies between
zero and one. 3.4. Mesh geometry
3.2. Initial and boundary conditions Two different geometries were checked to model the two-phase
flow in the tube. In the first layout, a simple rectangle, in the size of the
Uniform velocity inlets were employed as boundary conditions at testing section, was used to model the flow regime in the tube. In this
the gas and liquid inlets. An atmospheric pressure outlet condition layout, the input edge was divided into the liquid inlet (in the lower
was fixed for the outlet to avoid difficulties with backflow at the outlet part) and the gas inlet (in the upper part). The obtained results from
of the tube. No-slip boundary condition was imposed at the tube walls. this layout showed that although the predicted flow regimes are
The influence of the gravitational force on the flow was taken into similar to the experimental photographs, there are large differences
account. The initial volume fraction of water was set to zero for all between predicted and experimental pressure drops. Therefore, this
cases. modeling was not satisfactory. In the more accurate and consequently
more CPU time consuming modeling procedure the whole experi-
3.3. Solution strategy and convergence criterion mental tube including the calming section (first 2 m tube before the
test section) which contains the gas–liquid inlet junction was
Because of the dynamic behavior of the two-phase flow, unsteady modeled. The predicted pressure values were found at the positions
state calculations with a time step of 0.001 s were carried out for all of the experimental pressure transducers. These results were quite
Fig. 5. The comparison between experimental and CFD-predicted vertical flows. (a) GVN = 0.7 (Vsg = 0.1 ms− 1) LVN = 0.4 (Vsl = 0.07 ms− 1). (b) GVN = 5.6 (Vsg = 0.9 ms− 1)
LVN = 0.4 (Vsl = 0.07 ms− 1). (c) GVN = 23.7 ms− 1), LVN = 0.4 (Vsl = 0.07 ms− 1).
A. Alizadehdakhel et al. / International Communications in Heat and Mass Transfer 36 (2009) 850–856 855
GVN LVN ΔP/L (Pa m− 1) An artificial neural network was designed and trained to predict
Experimental Predicted Error % the average pressure drop of the two-phase flow in the studied tubes.
Vertical The input layer of the network consists of three variables (gas velocity
0.7 0.42 6961 6910 − 0.7 number, liquid velocity number and line slope) and the output layer
5.6 0.42 2605 2650 1.7 contains one variable (average pressure drop). In order to determine
145.2 0.42 1961 2061 5.1 the optimum number of neurons in the hidden layer, 3, 6, 9, 12, 15 and
Horizontal
20 numbers of neurons were used. Two-third of the total available
1.4 0.3 92 98 6.5 data (443 data sets) was used to train the network and the remainder
14.5 0.3 238 249 4.6 was used to validate the model results. Correlation coefficient (r2) and
145.2 0.3 1176 1113 − 5.4 mean square error (MSE) were used as the criterions for selecting the
best network architecture. A network with the higher values of r2 and
the lower values of MSE results more precise predictions. The MSE and
r2 values for different numbers of neurons in the hidden layer were
satisfactory when compared with the measured results. Therefore, obtained as shown in Table 2. The results from this table show that for
this strategy was used in the whole CFD modeling for different setups. the training set, by increasing the number of hidden nodes, the MSE
An investigation was done on the predicted pressure drops to find the decreases and r2 tends to increase. However, for the validation set, the
optimum grid size for the new geometry. Different sizes of the grid were network parameters got better as the number of neurons was
used to mesh the geometry and the predicted average pressure drops increased until it reached to fifteen hidden nodes. After that, there
were examined. Average pressure drops of 1419, 1632 and 1616 Pa m− 1 was no improvement in MSE. It can be explained by the fact that
were obtained for the horizontal layout with Vsl = 0.42 ms− 1 increasing of the nodes more than fifteen, the model over fits the
(LVN= 2.6) and Vsg = 14.3 ms− 1 (GVN=87.5) when the domain was training data and cannot generalize the rules to new data as well.
mesh respectively into 146,112, 280,662 and 438,898 control surfaces. Therefore, the ANN with fifteen hidden neurons was chosen as the
This examination showed that increasing of the number of control best network in this study.
volumes from 146,112 to 280,662 caused a significant difference in the In the present work, three different transfer functions (Log-
predicted pressure drops. However, employing more number of control Sigmoid, Hyperbolic-Tangent Sigmoid and Linear) were investigated
volumes up to 438,898 just had a negligible effect on the calculated to find the best one for the network with optimum number (fifteen) of
pressure drops. Therefore, to save CPU time, the second mesh layout was neurons in the hidden layer and a linear transfer function was
chosen as the optimum size of the control volumes. employed for the output layer transfer function. The obtained MSE
and r2 of validation set for this study are also given in Table 2. This
table shows that Log-Sigmoid transfer function gives better results
3.5. CFD modeling results than Hyperbolic-Tangent Sigmoid and Linear transfer function.
Therefore, a network with 15 nodes in the hidden layer, three inputs
Predicted flow regimes and pressure drops of horizontal and and one output, with Hyperbolic-Tangent Sigmoid and Linear transfer
vertical flows were compared with experimental photos as well as function for hidden and output layers was used, respectively.
measured pressure drops. The CFD modeling of horizontal flow was In Fig. 6, the measured pressure drops have been compared with
carried out at three different inlet conditions to model the flow the ANN and CFD-predicted results. Turning to Fig. 6(a), the figure
pattern of stratified-wavy, slug and annular regimes. Comparisons shows that the linear line fitted to ANN data is very close to the
between the predicted contour plots of liquid phase and experimental “predicted equals to experimental” line. In addition, the plotted data
photographs are shown in Fig. 4. As can be seen in the figure, there are very close to this line. In the other words, the results show that the
exist good agreements between predicted flow patterns and experi- presented ANN model can predict the pressure drop of the two-phase
mental photographs. In the horizontal annular flow of Fig. 4(c), due to flow at new conditions of various gas flow rate, liquid flow rate, and
the gravity effect, the liquid layer flowing near the lower wall of the line slope with a good precision. On the other hand, in order to
tube is thicker than that on the upper wall. compare the ANN predicted pressure drops with the CFD predictions,
For the vertical setup, three flow regimes of plug, slug and annular the CFD-predicted pressure drops of Table 1 and the corresponding
were modeled and presented in Fig. 5. Turning to Fig. 5(a), the values of experimental and ANN predictions are plotted in Fig. 6(b). As
parabolic head of the gas bubbles as well as the flow disturbance it can be seen in this figure, the CFD predictions are more accurate
under the bubbles is well predicted by CFD. The figure reveals that the than the ANN. However, regarding the level of accuracy and easy to
gas bubbles try to push the liquid upward and the both go toward the use advantage of ANN, the ANNs trained with the large number of
outlet. For the slug flow regime, Fig. 5(b), the same patterns happened
with larger bubbles. However, in this pattern the liquid tends to slide
down close to the tube wall and the downward sliding liquid Table 2
movement is more observable at this regime. Finally, in Fig. 5(c) due Modeling results obtained from different numbers of neurons and different transfer
functions of the hidden layer.
to high gas superficial velocity, the interaction between the gas and
liquid causes the liquid to go upward as thin layer adhered to the tube Model Transfer function Neurons in Training set Validation set
wall. In this annular flow regime, the gas phase flows in the center of no. of hidden layer the hidden layer MSE r2 MSE r2
the tube. In addition, there is a possibility of some liquid entrained in 1 Log-Sigmoid 3 0.991 0.9317 1.3205 0.9037
the gas phase, as illustrated in this figure. 2 Log-Sigmoid 6 0.173 0.9908 0.171 0.9880
In addition to visual validation of the flow regime, a comparison 3 Log-Sigmoid 9 0.164 0.9890 0.277 0.9885
4 Log-Sigmoid 12 0.0938 0.9938 0.161 0.9907
between the predicted and experimental time-averaged pressure
5 Log-Sigmoid 15 0.0253 0.9952 0.043 0.9931
drops is given in Table 1. From modeling results, no relationship can be 6 Log-Sigmoid 20 0.0219 0.9985 0.159 0.9889
found between the GVN, LVN and the prediction accuracy for the 7 Hyperbolic-Tangent 15 0.091 0.9899 0.121 0.9854
horizontal or vertical setup. However, the results show that the Sigmoid
pressure drop can be predicted with a good precision and the absolute 8 Linear 15 0.132 0.9792 0.541 0.9673
Acknowledgment
References
[1] G.F. Hewitt, Flow regimes, in: G. Hetsroni (Ed.), Handbook of Multiphase Systems,
Hemisphere, New York, 1982.
[2] J.S. Cole, G.F. Donnelly, P.L. Spedding, Friction factors in two phase horizontal pipe
flow, International Communications in Heat and Mass Transfer 31 (7) (2004)
909–917.
[3] S. Wongwises, W. Kongkiatwanitch, Interfacial friction factor in vertical upward
gas–liquid annular two-phase flow, International Communications in Heat and
Mass Transfer 28 (3) (2001) 323–336.
[4] P.L. Spedding, E. Benard, G.F. Donnelly, Prediction of pressure drop in multiphase
horizontal pipe flow, International Communications in Heat and Mass Transfer 33
(9) (2006) 1053–1062.
[5] J. Xu, Y. Wu, Z. Shi, L. Lao, D. Li, Studies on two-phase co-current air/non-
Newtonian shear-thinning fluid flows in inclined smooth pipes, International
Journal of Multiphase Flow 33 (9) (2007) 948–969.
[6] Y. Taitel, A.E. Dukler, A model for predicting flow regime transitions in horizontal
Fig. 6. The ANN and CFD modeling validations. a) ANN predicted and expected pressure
and near horizontal gas–liquid flow, AICHE Journal 22 (1) (2004) 47–55.
drops. b) CFD, ANN and experimental pressure drops. [7] W.C. Lyons, sixth ed., Standard Handbook of Petroleum and Natural Gas Engineering,
vol. 2, Gulf Professional Publishing, Huoston, Texas, 1996, pp. 426–533.
[8] T. Xie, S.M. Ghiaasiaan, S. Karrila, Artificial neural network approach for flow
regime classification in gas–liquid–fiber flows based on frequency domain analysis
available field data can supersede the common semi-empirical of pressure signals, Chemical Engineering Science 59 (11) (2004) 2241–2251.
methods for predicting the two-phase flow pressure drop in pipelines. [9] S. Cai, H. Toral, J. Qiu, J.S. Archer, Neural network based objective flow regime
identification in air–water two phase flow, Canadian Journal of Chemical
Engineering 72 (3) (1994) 440–445.
5. Conclusion [10] E.S.A. Osman, M.A. Aggour, Artificial neural network model for accurate prediction
of pressure drop in horizontal and near-horizontal multiphase flow, Journal of
Experiments were carried out to investigate the two-phase flow Petroleum Science and Technology 20 (1) (2002) 1–15.
[11] M.E. Shippen, S.S.L. Scott, A neural network model for prediction of liquid holdup
pressure drops and flow regimes in a tube. The plug, slug and annular in two-phase horizontal flow, Proceedings of the SPE Annual Technical Conference
flow patterns were observed in the tube vertical setup. In addition, the and Exhibition, San Antonio, Texas, 2002.
stratified-wavy, slug and annular regimes were recognized in the [12] T.R. Nigmatulin, F.J. Bonetto, Shape of Taylor bubbles in vertical tubes, Interna-
tional Communications in Heat and Mass Transfer 24 (8) (1997) 1177–1185.
horizontal tube. The experimental results showed that the average [13] J.S. Cole, G.F. Donnelly, P.L. Spedding, Trials of momentum balance for two-phase
pressure drop in vertical two-phase flows decreases with increasing horizontal pipe flow, International Communications in Heat and Mass Transfer 31
gas velocity at low gas flow rates, while at high gas flow rates it (7) (2004) 919–927.
[14] S. Ghorai, K.D.P. Nigam, CFD modeling of flow profiles and interfacial phenomena
increases with increasing gas velocity. The observed pressure drop in two-phase flow in pipes, Chemical Engineering and Processing Journal 45 (1)
variation trends at different gas/liquid flow rates in downward (2006) 55–65.
inclined tubes were similar to that of the horizontal setup. On the [15] L. Guang-yao, W. Jing, J. Zhi-hai, Experimental and numerical investigations on
horizontal oil–gas flow, Journal of Hydrodynamics 19 (6) (2007) 683–689.
other hand, the pressure drop behaviors across the tube in upward [16] S.C.K. De Schepper, G.J. Heynderickx, G.B. Marin, CFD modeling of all gas–liquid
inclined flows were similar to that of vertical flow. In the present and vapor–liquid flow regimes predicted by the Baker chart, Chemical Engineering
study, it has been found that CFD-predicted results can be sensitive to Journal 138 (1–3) (2008) 349–357.
[17] K. Ekambara, R.S. Sanders, K. Nandakumar, J.H. Masliyah, CFD simulation of bubbly
model simplification. The CFD results showed that proper modeling of
two-phase flow in horizontal pipes, Chemical Engineering Journal 144 (2008)
the flow inlet geometry is quite important in the reliable prediction of 277–288.
the two-phase flow behavior and simplifying the inlet geometries may [18] C.W. Hirt, B.D. Nichols, Volume of Fluid (VOF) method for the dynamics of free
cause significant errors. The CFD predictions were compared with the boundaries, Journal of Computational Physics 39 (1) (1981) 201–225.
[19] M. Lappa, A CFD level-set method for soft tissue growth: theory and fundamental
experimental measurements and corresponding photographs. The equations, Journal of Biomechanics 38 (1) (2005) 185–190.
results showed that by employing proper geometry, mesh and [20] Fluent user's guide, Version 6.2, Fluent Inc, Lebanon, NH, 2005.