Journal of Natural Gas Science and Engineering 33 (2016) 624e633

Contents lists available at ScienceDirect

Journal of Natural Gas Science and Engineering

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

An improved Ant Colony Optimization (ACO) technique for estimation

of flow functions (kr and Pc) from core-flood experiments
Muhammad Yaralidarani, Hamidreza Shahverdi*
Department of Chemical Engineering, Isfahan University of Technology, Isfahan 84156-83111, Iran

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

Article history: Oil and gas production from petroleum reservoirs are significantly affected by rock and fluid properties.
Received 30 January 2016 Relative permeability (kr) and capillary pressure (Pc) are two key flow functions in the petroleum
Received in revised form reservoir that are employed in numerical modeling to predict oil and gas production in future. The
18 May 2016
history matching (optimization) techniques are generally used to accurately estimate the flow functions
Accepted 25 May 2016
from the results of core flood experiments.
Available online 31 May 2016
In this study, Ant Colony optimization (ACO) method is modified for the application in the continuous
inverse problem to adjust some unknown variables by either minimization or maximization of objective
Keywords:
Ant Colony Optimization
function. Then this technique is implemented to estimate oil, water and gas flow functions from core-
Gas flood experiments. Some new ideas and innovations are proposed to improve the performance and
Oil convergence of ACO algorithm. This algorithm is an automated history matching technique that can
Reservoir estimate relative permeability and capillary pressure simultaneously. Moreover, this feature enables us to
Relative permeability incorporate different mechanisms (i.e. viscous force, gravitational force and capillary force) of the core
Flow function experiments in the estimation of flow functions. Having developed the algorithm, the validity of this
method is tested using two sets of coreflood experiments including gas-oil and oil-water systems. The
comparison between actual values of flow functions (kr and Pc) and those obtained from ACO method
depicted good agreement and adequate accuracy. Furthermore, the investigation of objective function
value versus iteration number demonstrated that the algorithm is converging to the optimal value.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction
Relative Permeability (kr) and Capillary Pressure (Pc) are two
important properties in simulation of oil and gas reservoirs that can
directly affect the fluid flow in porous media as well as oil and gas
production. These parameters are named as flow function in porous
media and governed by fluid saturation. Many researches have
been directed toward accurate estimation of these functions in
petroleum industry (Chen et al., 1999; Kerig and Watson, 1986;
Schembre and Kovscek, 2006). The capillary pressure and relative
permeability are not independent and one can influence the other.
The simultaneous calculation of flow functions from coreflood
experiment is of great importance for petroleum industry partic-
ularly in enhanced oil recovery (EOR) techniques.
Relative permeabilities curve is measured in the core laboratory
by two techniques, steady state and unsteady-state experiments. In
* Corresponding author.
E-mail address: hr_shahverdi@cc.iut.ac.ir (H. Shahverdi).
the steady state method, two immiscible fluids (e.g. water and oil)
are injected cocurrently at a specific ratio through the core until the
same production ratios are achieved from the outlet of the core. In
this method, the relative permeability data can directly be calcu-
lated from Darcy law. However, this method has some disadvan-
tages like being a time consuming process, unable to capture the
values of low saturation and ignoring the effect of capillary forces
whereas in the unsteady state method, one of the fluids is injected
into the core to displace the resident fluid phases. This test is much
quicker than the steady-state test but the flow functions (kr and Pc)
cannot be directly drawn from experimental results and need in-
verse calculation. There are two methods for estimation of relative
permeability from unsteady state tests, analytical (explicit) method
and optimization (implicit) method. The most common analytical
method is JBN (Johnson et al., 1959) which consists of numerical
differentiation of pressure and production values. In spite of
simplicity, the JBN method has serious limitations including:
ignoring capillary and gravitational forces, considering flow in one
dimension, and assuming the fluids to be incompressible and

http://dx.doi.org/10.1016/j.jngse.2016.05.067
1875-5100/© 2016 Elsevier B.V. All rights reserved.
 M. Yaralidarani, H. Shahverdi / Journal of Natural Gas Science and Engineering 33 (2016) 624e633
isothermal. Furthermore, the numerical differentiating from
experimental data points causes erroneous results in particular for
data which has high fluctuation. In order to overcome the JBN
limitations and obtain a more accurate flow function, the optimi-
zation or history matching technique is used. In this approach, a
relative permeability and capillary pressure model are tuned in an
optimization procedure to reach an acceptable match between the
laboratory results of the fluid production, pressure drop and the
corresponding values calculated by the model (Shahverdi et al.,
2011). The optimization methods are divided into two types of
deterministic (gradient) and stochastic approaches. Kerig et al.
(1988) calculated the two-phase relative permeability values from
unsteady state test using gradient method. Mejia et al. (1996)
applied this method to predict three-phase relative permeability
from coreflood experiment. The gradient optimization method
originally was developed in 1963 by Levenberg & Marquardt (More ,
1978). As the name implies the gradient methods takes derivative
of the explicit objective function to find the optimal value (e.g.
Newton method). The main drawback of this method is its de-
pendency to initial guess which could lead to erroneous results
when dealing with large number of unknown parameters.
The second type of optimization methods are meta-heuristic
(stochastic) methods that search for the optimal value based on
the stochastic algorithm. Most of these methods are inspired from
rules that exist in nature. The main advantage of meta-heuristic
methods is wide search domain and the ability to avoid trapping
in the local optimal value. The most common meta-heuristic
methods are Genetic Algorithm (GA) (Kalyanmoy et al., 2002),
Ant colony Optimization (ACO) (Dorigo and Stutzle, 2004), Particle
Swarm Optimization (PSO) (Riccardo et al., 2007) and so on. Meta-
heuristic optimization algorithms have been used in oil industry for
estimation of some scalar parameters in petroleum reservoirs (e.g.
porosity and permeability of rock) but these techniques have
seldom been employed for estimation of capillary pressure and
relative permeability values. Razavi and Jalali-Farahani (2010)
applied ACO algorithm for calculation of absolute permeability (K)
and porosity (Ø) by history matching of wells pressure and well
production rate. They also implemented both ACO and GA methods
in a two dimensional oil reservoir model to maximize cumulative
oil production by tuning the location of production and injection
wells, well radius, well pressure and rate of fluid injection. They
obtained a good agreement between the consequences of GA and
ACO models. Hajizadeh et al. (2011) applied ACO technique for
history matching of oil reservoir by tuning some unknown pa-
rameters which consists of absolute permeability and porosity.
They achieved a reasonable agreement between actual parameters
and those obtained by ACO. Hajizadeh et al. suggested ACO method
as an efficient approach in modeling oil reservoirs. Irani and Nasimi
(2012) proposed a proper hybrid of back-propagation (BP) algo-
rithm and ant colony optimization algorithm (ACA) which could
improve the performance of neural networks. They applied this
new method to predict the permeability of a reservoir. The results
showed that the proposed ACA-BP scheme was more efficient and
effective than the BP algorithm. Zhang and Wang (2014) performed
the inversion of reservoir parameters with improved niche ant
colony algorithms (INACA). The method was successful at inverting
porosity and permeability and it was indicated that the relative
error of single parameter inversion can be maintained at less than
0.4 percent. Finally, the numerical simulation showed that the
proposed INACA method is not only an effective convergent opti-
mization method but also well suited one for solving ill-posed
problems. Jung et al. (2015) presented application of ant colony
optimization in estimation of the reservoir properties such as
permeability, skin factor and wellbore storage coefficient from an
insufficient field data. They found that the results have successfully
625
matched to those obtained from genetic algorithm, the non-linear
regression and the modified LevenbergeMarquardt method. They
also proposed the ant colony method as a good way of optimization
in petroleum engineering area, especially well testing analysis.
In this study, we have used ACO optimization method for
simultaneous prediction of relative permeability and capillary
pressure of immiscible fluids in the porous media. However, to the
best of our knowledge, no researchers have been employed this
approach so far in the oil industry for estimation of flow functions.
Furthermore, the previous studies have mainly focused on predic-
tion of scalar properties in oil and gas fields (e.g. pressure, porosity
and absolute permeability) whereas in this study, authors estimate
the flow function curves which are dependent on fluid saturation.
Moreover, some significant innovations have been proposed in this
study to improve the performance of the traditional ACO algorithm.
The authors believe the proposed ACO in this research is more
compatible with real ant behavior in nature.
The methodology of modified ACO devised in this study is
described for the general optimization problem. Then the model is
validated using unsteady-state coreflood experiment to estimate
relative permeability and capillary pressure simultaneously.
2. Ant colony theory (ACO)
One of the most interesting types of meta-heuristic optimiza-
tion techniques is ant colony optimization (ACO), which was orig-
inally developed by Dorigo (1992). This method was firstly
proposed to solve the discrete space problems (e.g. Travelling
Salesman Problem (Dorigo and Gambardella, 1997), Quadratic
Assignment (Maniezzo and Colorni, 1999) but later was developed
to solve the continuous domain problems (Socha and Dorigo, 2008;
Seçkiner et al., 2013; Xiao and Li, 2011). The ACO algorithm was
devised based on the behavior of ants in nature in which they
search for the smallest (optimum) path between their nest and
food. The investigation on the ants’ life depicted that all ants leave a
substance called pheromones on the pathway while moving. The
amount of this substance in every path is reduced by time because
of evaporation. In the other hand, the pheromone of a path is
increased by those ants that passed through that particular path.
Hence, the amount of pheromone in a path is controlled by two
factors, the rate of evaporation and the number of ants who passed
through that particular path. The pheromones left behind on the
paths can be recognized by the other ants using the sense of smell
such that the ants try to find those paths that have more
Fig. 1. Paths selected by various ants from nest to food.
626 M. Yaralidarani, H. Shahverdi / Journal of Natural Gas Science and Engineering 33 (2016) 624e633

pheromone. Fig. 1 demonstrates a schematic map of the paths from

nest to food source discovered by the various ants. Assuming two
ants commence moving from the nest to the food source (Fig. 1),
one takes the longer path whereas the other selects the shorter
path also each ant leaves the pheromone in its pathway. Since, the
smaller paths take shorter time to be travelled, the pheromone of
these paths would be more compared to that of the longer paths.
This fact causes the shorter (optimum) path to have a higher chance
of selection by other ants. Although the shorter paths have the
higher chance of selection by the ants, the other paths still have the
possibility to be selected by the ants. This fact results in a wide
domain of exploration by various ants and thus discovering a new
path that can be potentially an optimal solution. This feature en-
ables ACO algorithm to find the global optimum value rather than
being trapped into the local optimum value.
The original version of ACO which was devised for the contin-
uous problems, creates an archive of solutions with the constant
number of members (solutions). A weight factor is associated to
Fig. 2. The created path by the various ants.
each of members reflecting the goodness (quality) of that member.
This weight factor is function of the objective function and the
pheromone value which is updated in every iteration. In each used as the length of path in the ant algorithm.
iteration “k” new solutions are created by sampling from the It should be noted that, at the beginning of the optimization
archive. The sampling is performed by a probability function there is no path between start and end point because the unknown
considering the quality of each solution (weight factor). After that parameters have not been guessed yet. The paths (nodes) are
the new members created in the previous step are added to the created gradually by the ants (iterations) as optimization proceeds.
archive and the archive is sorted based on the quality of each This is the main difference between the ACO algorithm proposed in
member. The best members are kept in the list while the worse are this research and the one developed originally to solve the TSP
removed from the archive. The number of the member in the problem (Dorigo, 1992). The paths in TSP (the distance between the
archive remains constant during the various iterations. cities) have previously been identified whereas the paths in our
algorithm are associated with unknown parameters that will be
3. Implementation of ACO algorithm for optimization created progressively by the process of problem solving.
problems The modified ACO algorithm is expressed in three main steps as
follows:
In this section, the mathematical algorithm of the improved ACO Step 1: Creating initial paths: At this stage, the initial paths (initial
algorithm devised in this study will be discussed. First, the nota- guess) are generated randomly by the primary ants. At this stage,
tions presenting the formulation of ACO are described below: different values for “Si” are generated randomly using a represen-
Path: is a solution for the problem which consisted of various tative and constrained formula which depends on the physic of the
parameters. These parameters will be described later in this paper. problem. Then the corresponding objective function to each vari-
Si: The unknown variables (vector) which are going to be esti- able (Si) is calculated. Fig. 3 demonstrates a schematic of “k” initial
mated by ACO. This vector is associated to the path i. paths which all are connected to the start node. This should be
Ji: The objective function of the vector Si which certainly de- noted that the number of initial paths dictates the extension of
pends on the unknown variables Si. The objective function is the search domain which may significantly influence the convergence
function that is going to be minimized or maximized by changing Si performance. For complex problems that deal with a larger number
in an iterative process. For many engineering problems, this func- of unknown variables, more initial paths are required. In this
tion is defined as the difference between experimental and research, five initial paths are generated.
modeling values. For these problems, the target is to minimize the Step 2: Pheromone updating: Pheromone Updating is a handy
objective function. tool for measuring the accuracy of created solutions. The paths
Fi: The amount of pheromone left behind in the path i by the
ants.
In a “n variable problem”, Si is defined as Si ¼ {R1,i, R2,i, …, Rn,i } in
which Rj,i (1  j  n) are the unknown parameters. It is worth noting
that each ant is equivalent to an iteration in the optimization loop.
The ants that are used to create the initial paths (initial guess) are
exempted from the iterations.
Fig. 2 shows the schematic of paths created by the various ants
in finding the food. The final solution will be achieved while the
minimum objective function (food) is obtained. The start point in
Fig. 2 accords to the initial solution (far from the food) which has
highest objective function and the end point means the lowest
objection function or final solution. Since, the target is to achieve
the minimum path to the food, the length of each path is equal to
the inverse of objective function. In other words, the lower objec-
tive function results in a longer path towards the food. For maxi-
mization problems (e.g. benefits function in economics) in which
the goal is to obtain the highest value, the objective function itself is Fig. 3. A schematic of paths generated by various ants.
 M. Yaralidarani, H. Shahverdi / Journal of Natural Gas Science and Engineering 33 (2016) 624e633 627

whose objective functions are smaller have more pheromones thus (Si). For instance, in Fig. 3, the paths “k þ 1” and “k þ 2” are
they have more chances to be selected by the ants. Those paths linked to path 2 demonstrating that “Skþ1” and “Skþ2” are pro-
which have larger objective function (smaller “ 1/J ”), on the other duced by some modifications on “S2”. While some other paths
hand, have less pheromones and thus less chance to be selected. are independent depicting that there is not any relationship
This fact depicts that the pheromone of any path has an inverse between their variables, “Si” (e.g. path 1, 2 and 3 in Fig. 3). The
relationship with the objective function of that path. Hence, the maximum number of paths that can be generated from a
pheromone of path “i” can be represented by the following particular path is called maximum split which is a crucial
formula: parameter in convergence and needs to be adjusted properly.
The higher split leads to a wider search domain and more
8
> chance of capturing the global optimum value but later
< ð1  EÞ F n þ 1 L:Ph:U
Finþ1 ¼
i Ji (1) convergence. The maximum number of split in this research is
>
: n selected as five.
ð1  EÞ Fi G:Ph:U
New ant (iteration) commences travelling from the start node by
Where “ Finþ1 ” stands for pheromone of path “i” at iteration “n þ 1”
choosing one path among initial paths generated in step 1. The
and “n” is the iteration (ant) number. The L.Ph.U and G.Ph.U are
selection of path at each node is performed by roulette wheel
indicators of local and global pheromone updating, respectively.
technique (Goldberg, 1989).
The L.Ph.U is utilized for the paths that have been selected by the
Having selected one path, the ant arrives to the next node that
current ant (iteration) and “1/Ji” represents the added pheromone
may have one of three following circumstance:
left behind in the path “i” by the current ants. The G.Ph.U is used for
those paths that the current iteration (ant) has not selected. One of
a) The ant arrives to a node that has no path in front. At this sit-
the specifications of the pheromone is evaporation by time leading
uation the ant generates new path (new “Si” and “Ji”) related to
to reduction of the pheromone over time. The Pheromone evapo-
the current path. In other words, the ant produces new variable
ration is a negative feedback on the path goodness and prevents
“Si” by major modification on variable of the current path (“Si-1”).
excessive accumulation of pheromones on the path. In Eq. (1), “E”
This modification is equivalent to the mutation stage in Genetic
represents the pheromone evaporation rate which should be a
Algorithm.
value between zero and one. The term “EFni ” depicts the amount of
b) The ant arrives to a node that has maximum split (five paths) in
evaporated pheromone of path “i” over time, hence the “(1-E) Fni ”
front. At this condition the ant will choose one of these paths in
represents the remaining amount of pheromone in that path. The
front using the roulette wheel approach.
term “(1-E). Fni ” illustrates the fact that if a path is not passed by any
c) The number of paths in front of a node is a value between one to
ant, its pheromone will decrease progressively over time and the
maximum split (i.e. five paths). At this state, ant either generates
possibility of selection of the path will be diminished. In this study,
a new path or chooses one of the existing paths. For this pur-
the evaporation rate is considered as 0.15.
pose, a virtual path with a random probability is considered (e.g.
Step 3: Moving ants (main loop): After creating initial paths
path “k þ 3” in Fig. 3) then the probability of all paths connected
(initial guesses), the ants start travelling from start point towards
to the existing node (paths “k þ 1”, “k þ 2”, “k þ 3” in Fig. 3) are
end point either by passing through the existing paths or creating
normalized using Eqs. (4)e(6). After that, the roulette wheel
new paths. This step is the main loop of optimization procedure. In
technique is applied to select one of these paths. If the ant select
this research, we define following principles and rules for optimi-
the virtual path (“k þ 3”), then this path is accounted as a real
zation by ACO:
path and a variable (“Skþ3”) is assigned to that path (“k þ 3”).
Subsequently its objective function (“Jkþ3”) and pheromone (by
1 Each path is equivalent to a solution for the problem which
L.Ph.U) are calculated. In the case that the ant has not selected
consists of four parameters as follows:
the virtual path the virtual path is deleted.

Path i ¼ ½Si ; Ji ; Fi ; Pi  (2) pold

kþ1
pnew
kþ1 ¼ (4)
Where “Si” is a guess for the unknown variables, “Ji” is the objective pold
kþ1
þ pold
kþ2
þ pold
kþ3
function which corresponds to “Si”, “Fi” is the pheromone of path “i”
given by Eq. (1), “Pi” is the possibility assigned to the path “i” which
pold
kþ2
is related to the pheromone by the following equation: pnew
kþ2 ¼ (5)
pold
kþ1
þ pold
kþ2
þ pold
kþ3
Fa
Pi ¼ Pk i (3)
a pold
i¼1 Fi kþ3
pnew
kþ3 ¼ (6)
pold
kþ1
þ pold
kþ2
þ pold
kþ3
Where a is a constant parameter designated to accelerate the
convergence speed. We found the value of 1.5 for a, as an appro- This should be highlighted that any ant is allowed to produce
priate value in this study, obtained by trial and error. The “k” in Eq. only one path. In other words, each ant keeps going through the
(3) is the total number of paths until the current iteration. This pre-generated paths until it reaches a node that needs a new path
should be noted that “Si” and “Ji” of each path are constant during to be generated. Having produced the new path, the ant is stopped
running various iterations whereas the possibility (Pi) and the and a new ant commences travelling from the start node (new
pheromone (Fi) of each path are varied by iteration because new iteration starts).
paths are added in each iteration. This should be noted that the idea of “virtual path” is a kind of
“Exploration” and has the merit of creating various solutions which
2 Each path is connected to other paths via a node as shown in increases the chance of reaching the optimized solution. Explora-
Fig. 3. Those paths which are connected to each other via a node tion consists of probing a much larger portion of the search space
represent mathematical relationships between their variables with the hope of finding other promising solutions that are yet to be
628 M. Yaralidarani, H. Shahverdi / Journal of Natural Gas Science and Engineering 33 (2016) 624e633

Fig. 4. Algorithm of improved ACO optimization.

Fig. 5. Prediction of relative permeability and capillary pressure using ACO model.

refined. This operation amounts then to diversifying the search in
order to avoid getting trapped in a local optimum. In other words,
exploration is related to global search and improves dynamic effi-
ciency. The complete flowchart of the improved ACO described in
the above stages is depicted in Fig. 4.
Regarding the ACO parameters (e.g. pheromone evaporation,
maximum split number, parameter “a” in probability formula, etc.),
it should be emphasized that these parameters strongly depend
upon the type and size of the optimization problem. Hence, for any
optimization problem, these parameters need to be determined
carefully using sensitivity analysis.
As described above, the new features and parameters are
considered in the ACO method to improve the optimization pro-
cess. This modification is mainly focused on the better controlling
of balance between exploration/exploitation in order to direct the
algorithm to the global optimized value and the higher conver-
gence speed.
4. Implementing ACO algorithm to estimate flow functions in
porous media
In this section, the ACO approach is used to estimate the flow
functions (kr and Pc) in the petroleum reservoirs from the results of
unsteady state coreflood experiments. Different modes of the cor-
eflood experiments are performed in the oil industry in order to
estimate various parameters of the reservoir rock such as porosity,
absolute permeability, relative permeability and capillary pressure
curves. In the unsteady state coreflood experiment, the sample of

Table 1
Rock properties and dimensions of case 1.

Connate water saturation Average absolute permeability (mD) Porosity Length of model (cm) Diameter (cm)

0.1 2620 0.24 50 5.64

 M. Yaralidarani, H. Shahverdi / Journal of Natural Gas Science and Engineering 33 (2016) 624e633 629

Table 2
Fluid properties used in case 1.

Phase Water Oil

Density (gr/cc) 1.1 0.80

Viscosity (cp) 1.14 2.52

Fig. 7. Cumulative oil production as a fraction of initial oil in place (IOIP) versus
Injected pore volume for water injection (case 1).

Fig. 6. Core gridding and permeability distribution of case#1.

reservoir rock (core) is saturated with oil then water or gas is

injected through the core to displace oil inside the core. The
effluent productions from the outlet of the core as well as the inlet
and outlet pressure are recorded during the test. These results
alongside the boundary constraints (i.e. injection and production
pressure or flow rate) and initial condition are all employed in the
fluid flow equations (i.e. continuity and Darcy equation) to simulate
the flow in the rock. The full description of simulation in porous
media for petroleum reservoir is given in other references (Ertekin
et al., 2001; Ahmed, 2006; Aziz and Settari, 1979). The estimation of
flow functions from the core experiment is an inverse problem that
requires optimization algorithm as illustrated in Fig. 5. The rock and
fluid properties (i.e. size of core, porosity, permeability, fluid den-
sity and viscosity) measured in the laboratory together with an Fig. 8. Pressure drop across the core (psi) versus injected pore volume for water in-
initial guess for relative permeability and capillary pressure are jection (case 1).
used in a coreflood simulator to obtain the fluid production and
pressure versus time. This should be highlighted that the flow
5. Model validation
function curves (kr and Pc) are generated using an existing empir-
ical correlation (e.g. Corey, LET, …) by considering the specific
In this section, two sets of coreflood experiments are used to
constraints to get a meaningful and a physical trend for kr. The
verify the accuracy of the ACO algorithm proposed in this study.
relative permeability of each phase should be monotonically
First case: In this case, a synthetic core and fluid data were used
increasing with respect to its own saturation and its value is limited
to design an unsteady-state core experiment by a black oil simu-
between zero and one. The mathematical expressions of these
lator. The rock and fluid properties used in this case are given in
constraints are as follows:
Table 1, Table 2. The relative permeability and capillary pressure
0  kri  1 i ¼ oil; water; gas (7) data versus water saturation was imported into the simulator as the
table. The core was modeled by 50 blocks in X direction, 5 blocks in
Y direction and 10 blocks in Z direction. In order to consider the
vðkri Þ
0 i ¼ oil; water; gas (8) heterogeneity of the rock, a normal distribution of porosity and
vðSi Þ permeability with the mean values of 0.24 and 2620 mD were
Having simulated the core, the difference between experimental assigned to the model. The model geometry and gridding is pre-
production, pressure and those of the simulation data are calcu- sented in Fig. 6. The initial and boundary condition of the model is
lated as objective function. Then the optimization algorithm (ACO) given in Table 3. Having run the simulator, the fluid production and
is used to minimize the objective function by adjusting the un- pressure data resulted from the simulation were reported as
known parameters (kr and Pc). experimental data.

Table 3
Initial and boundary conditions used in the coreflood simulation of case 1.

Initial pressure (atm) Initial water saturation Initial gas saturation Water injection rate (cc/hr) Total fluid production (cc/hr) (outlet)

203 0.2 0 40 40
630 M. Yaralidarani, H. Shahverdi / Journal of Natural Gas Science and Engineering 33 (2016) 624e633

Fig. 11. Objective function versus iteration number for case 1.

Fig. 9. Relative permeability of water and oil versus water saturation obtained from
the history matching and the experiment for case#1.

Sao o  Saorw
o

Krow ¼ bo ao (10)
1  Sorw

cw co
Pcow ¼  aw   ao (11)
Sw Swi 1Sw Sorw
1Swi 1Sorw

Where Sw is water saturation, Swr is the minimum water saturation

at which point water phase starts to flow, So is oil saturation, krwo
and krow are water and oil relative permeability in oil-water system,
respectively. Pcow is oil-water capillary pressure.
According to the above equations, the unknown vector is
expressed as follows:

S ¼ f bw ; Swr ; aw ; ; bo ; Sorw ; ao ; cw ; aw ; co ; ao g (12)

This should be highlighted that the specific constraints (Eqs. (7)
and (8)) were imposed on the unknown vector (Eq. (12)) in order to
Fig. 10. Oil-water capillary pressure versus water saturation obtained from the
produce physical shape for kr and Pc during history matching.
experiment and the model for case 1. The objective function used in the ACO algorithm is defined as
follows:
After that, the flow function table (kr and Pc) were accounted as
N  2
unknown parameters and the modified ACO algorithm (Fig. 5) was 1 X Exp Sim
J ¼ Wοw QOw  QOw
used to adjust these functions. In other words, the ACO algorithm s2οw i¼1 i i

changed the relative permeability and capillary pressure iteratively

N 
X 2
to obtain an acceptable agreement between experimental data (i.e. 1
þ WDР;w DРExp  DРSim (13)
fluid production and differential pressure) and those of simulation. s2 DР;w i¼1
w;i w;i

The relative permeability and capillary pressure in the ACO al-

gorithm can be represented by any existing empirical correlation.
The modified Corey model (Eqs. (9) and (10)) (Corey et al., 1956)
and Skjaeveland’s capillary pressure model (Eq. (11)) (Skjaeveland
et al., 2000) were considered in history matching of case#1.
Saww  Sawrw
Krwo ¼ bw
1  Saww
In the above equation, sοw and sDP,w are standard deviations of
experimental values for the oil production and pressure drop,
respectively. These quantities represent the errors associated with
the coreflood experiments, which can be obtained by repetition of
the experiment. The terms Wοw and WDР,w are weighted co-
efficients compensating the magnitude difference between the two
(9) quantities (oil production and pressure drop across the core). The
proper adjustment of these factors allows us to match fluid pro-
duction and pressure to the same extend.

Table 4
The estimated parameters of modified Corey model and Skjaeveland model by the ACO method for case 1.

Relative permeability (modified Corey) Capillary pressure (Skjaeveland)

bw Swr aw bο Sοrw aο cw aw co ao

2.508 0.0997 5.8876 1.629 0.3000 3.4334 0.000426 0.84465 0.001786 2.1887
 M. Yaralidarani, H. Shahverdi / Journal of Natural Gas Science and Engineering 33 (2016) 624e633 631

Table 5
Rock properties and core dimensions for case 2.

Connate water saturation Absolute permeability (md) Porosity Core length (cm) Core diameter (cm)

0.31 200 0.22 6.0 4.502

Table 6
Fluid properties at 25  C for case 2.

Phase Water Oil Gas

Density (gr/cc) 1.00 0.83 0.22

Viscosity (cp) 1.06 1.77 0.0187

Fig. 7 shows cumulative oil production (as a fraction of initial

oil in place) versus injected pore volume (the total volume of
injected fluid as a fraction of total core pore volume) obtained
from the ACO model and the experiment. Fig. 8 demonstrates the
differential pressure across the core against injected pore volume
obtained from the ACO model and the experiment. As can be seen,
the break through time and trend of experimental data are
reasonably matched by ACO model for both fluid production and
pressure.
Fig. 9 demonstrates relative permeability of oil and water versus Fig. 12. Cumulative oil production as a fraction of initial oil in-place (IOIP) in core
water saturation resulted from the history matching and the versus total injected fluid as a fraction of core pore volume for case 2.

experiment. It should be noted that the relative permeability data

resulted from the core experiments are only reliable for the satu-
ration range in which experiment is occurred. The oil-water capil-
lary pressure versus water saturation obtained from the history
matching shows good agreement with the experimental one as
compared in Fig. 10. The best estimated parameters of capillary
pressure model (the Skjaeveland model, Eq. (11)) and relative
permeability model (the modified Corey model, Eqs.( (9) and (10))
are provided in Table 4.
In order to assess the convergence and performance of the ACO
algorithm, the objective function of case 1 was monitored during
various iterations in Fig. 11. This figure illustrates that at the initial
iterations the objective function has the highest value (or ant is far
from the food) while the objective function value decreases grad-
ually as the optimization proceeds.
The reduction of the objective function during the execution of
the ACO algorithm reflects the reliable performance and conver-
gence towards the desirable solution.
Second case: In this section, the ACO model is verified using the
Oak experimental data (Oak, 1989). The rock and fluid data as well Fig. 13. Differential pressure across the core versus total injected fluid as a fraction of
core pore volume fore case 2.
as the initial and boundary condition of this case are provided in
Table 5, Table 6 and Table 7. In this test, the gas is injected at fixed
rate into the horizontal core initially saturated with oil and connate
water. The values of the pressure drop across the core and cumu- N  2
lative oil production were recorded as experimental data during the 1 X Exp Sim
J ¼ Wοg QOgi  QOg
flooding process. Similar to the previous case, the simulation of this s2οg i¼1 i

test was performed using a one dimensional black oil coreflood N 

X 2
1
simulator. þ WDР;g DРExp  DРSim (14)
Similar to the water injection test, the objective function used in s2DР;g i¼1
g;i g;i

the ACO process is defined by the following equation:

Where sοg and sDP,g are standard deviations of the experimental oil

Table 7
Initial and boundary conditions of case 2.

Initial pressure (atm) Initial fluid saturations Gas injection rate (cc/hr) Number of grids

Water Oil Gas

10.00 0.31 0.69 0.00 10 20

632 M. Yaralidarani, H. Shahverdi / Journal of Natural Gas Science and Engineering 33 (2016) 624e633

Fig. 14. Oil and gas relative permeability versus oil saturation for gas injection test Fig. 16. Objective function variation versus iteration number obtained from simulation
(case 2) obtained from the experiment and the ACO model. of case 2 by the ACO model.

 Lg
1  S*o
o
Krog ¼ Krgo  L g  Tg (16)
1  S*o þ Eg S*o

cg
Pcgo ¼  ag (17)
S*g

So is the normalized oil saturation defined as follows:

So  Sorg
S*o ¼ (18)
1  Sorg  Swi
S*g is the normalized gas saturation defined as follows:

Fig. 15. Gas-oil capillary pressure versus oil saturation for gas injection test (case 2)
obtained from the experiment and the ACO model.
production and pressure drop associated with the experimental
data. The terms Wοg and WDР,g are the weighted coefficients
compensating the difference in the order of magnitude between
the oil production and pressure drop across the core. The relative
permeability curves in this case are represented by the LET model
(Eqs. (15) and (16)) (Lomeland et al., 2005) which consists of more
unknown parameters compared to the Corey model (used in the
first case). The gas-oil capillary pressure curve (Pcgo) is represented
by a shortened form of the Skjaeveland model (Eq. (17))
(Skjaeveland et al., 2000). This should be highlighted that because
gas phase behaves as a non-wetting phase against the oil and rock
surface, the gas-oil capillary pressure should be considered as a
positive value. Hence, the positive part of Skjaeveland model is
used in gas injection test to represent the gas-oil capillary pressure.
 L o
S*o
o
Krog ¼ Krog  Lo  To (15)
S*o þ Eo 1  S*o
1  Sg  Sorg  Swi
S*g ¼ (19)
1  Sorg
Where So and Sg are saturations associated with oil and gas phase,
respectively. Sorg stands for residual saturation of gas phase (the
minimum gas phase saturation) and Swi represents irreducible or
connate water saturation.
According to the Eqs. (15)e(19), the unknown vector to be used
in the ACO model is expressed as follows (Eq. (20)):
n o
o o
S ¼ Krog ; Lo ; Eo ; To ; Sorg ; Krog ; Lg ; Eg ; Tg ; cg ; ag (20)
The ACO algorithm was implemented for this test to obtain the
above unknown vector in an iterative process by getting a match
between experimental and simulation data.
Fig. 12 shows the cumulative oil production (as a fraction of
initial oil in-place in core) versus total injected fluid (as a fraction of
core pore volume) obtained from the experiment and the ACO
model for the second case. Fig. 13 depicts differential pressure
across the core during gas injection versus injected fluid (pore
volume) resulted from the experiment and the ACO model. Both
production and pressure consequences represent sufficient agree-
ment between the experimental and the model data.
The comparison between relative permeability of the experi-
ment and that of ACO model obtained from case 2 are provided in

Table 8
Estimated values of Skjaeveland and LET model by the ACO method (case 2).

Relative permeability (LET) Capillary pressure

o
Krog Lo Eo To Sorg o
Krgo Lg Eg Tg cg ag

0.8429 2.8099 2.0754 0.6979 0.18 0.5170 1.4778 2.7579 0.9037 6.0990 1.0844
 M. Yaralidarani, H. Shahverdi / Journal of Natural Gas Science and Engineering 33 (2016) 624e633 633

Fig. 14. The gas-oil capillary pressure for the ACO model and the sow Standard deviations of experimental values for the oil
experiment of case 2 are demonstrated in Fig. 15. As can be seen, production in oil-water system
both relative permeability and capillary pressure curves drawn sDP,w Standard deviations of experimental values for the
from the ACO have acceptable agreement with the corresponding pressure drop in oil-water system
experimental data. The parameters associated with the LET model sog Standard deviations of experimental values for the oil
and Skjaeveland model calculated by the ACO for case 2 are pro- production in oil-gas system
vided in Table 8. sDP,g Standard deviations of experimental values for the
To ensure the accuracy of the ACO model, the objective function pressure drop in oil-gas system
variation versus the iteration number during ACO implementation
is recorded in Fig. 16. The decreasing trend of the objective function References
in this Figure shows the ACO algorithm converging to the minimum
value. Ahmed, T., 2006. Reservoir Engineering Handbook. Society of Petroleum Engineers,
Richardson, TX.
Aziz, K., Settari, A., 1979. Petroleum Reservoir Simulation. Applied Science, Pub-
6. Conclusions lishers, London.
Chen, J., Hopmans, J.W., Grismer, M.E., 1999. Parameter estimation of two-fluid
capillary pressureesaturation and permeability functions. Adv. Water Resour.
The main conclusions drawn from this study are: 22 (5), 479e493.
Corey, A., et al., 1956. Three-phase relative permeability. J. Petroleum Technol. 8
(11), 63e65.
1 The new idea and innovations are incorporated in the Ant Col- Dorigo, M., 1992. Optimization, learning and natural algorithms. In: Systems and
ony optimization (ACO) algorithm for the continuous inverse Information, Electronic Engineering Department. Politecnico di Milano, Italy.
problem in order to adjust some unknown variables by either Dorigo, M., Gambardella, L.M., 1997. Ant colony system: a cooperative learning
approach to the traveling salesman problem. IEEE Trans. Evol. Comput. 53e66.
minimization or maximization of objective function. Dorigo, M., Stutzle, T., 2004. Ant Colony Optimization. MIT Press, United States of
2 The proposed ACO technique is implemented to estimate oil, America.
water and gas flow functions from the coreflood experiments. Ertekin, T., Abou-Kaseem, J.H., King, G.R., 2001. Basic Applied Reservoir Simulation.
Society of petroleum engineers, Richardson, TX.
Linking the ACO algorithm to the numerical simulator, enables
Goldberg, David E., 1989. Genetic Algorithms in Search, Optimization, and Machine
us to incorporate different mechanisms (i.e. viscous force, Learning. Addion wesley 1989.
gravitational force and capillary force) of the core experiments Hajizadeh, Y., Christie, M., Demyanov, V., 2011. Ant colony optimization for history
matching and uncertainty quantification of reservoir models. J. Petroleum Sci.
in the estimation of flow functions. This feature is an advantage
Eng. 77 (1), 78e92.
on top of the JBN method that ignores some physics of the flow Irani, R., Nasimi, R., 2012. An evolving neural network using an Ant Colony Algo-
in calculation of relative permeability. rithm for a permeability estimation of the reservoir. Petroleum Sci. Technol. 30
3 Having developed the algorithm, the validity of this method is (4), 375e384.
Johnson, E.F., Bossler, D.P., Naumann, V.O., 1959. Calculation of relative permeability
tested using two sets of coreflood experiments including gas-oil from displacement experiments. Trans. AIME 216, 370e372.
and oil-water system. The comparison between actual values of Jung, J., Seo, H., Yoo, I., et al., 2015. Analysis of well testing data using ant colony
flow functions (kr and Pc) and those obtained from the ACO optimization. Geosystem Eng. 18 (5), 1e6.
Kalyanmoy, Deb, Pratap, A., Agarwal, S., Meyarivan, T.A.M.T., 2002. A fast and elitist
method depicted good agreement and adequate accuracy. multiobjective genetic algorithm: NSGA-II. Evol. Comput. IEEE Trans. 6 (2),
182e197.
Kerig, P.D., Watson, A.T., 1986. Relative-permeability estimation from displacement
Nomenclature experiments: an error analysis. SPE Reserv. Eng. 1 (02), 175e182.
Lomeland, Frode, Ebeltoft, Einar, Thomas, Wibeke Hammervold, 2005. A new ver-
satile relative permeability correlation. In: International Symposium of the
Si Unknown Vector Society of Core Analysts, Toronto, Canada, pp. 1e12.
Ji Objective function Maniezzo, V., Colorni, A., 1999. The Ant System applied to the quadratic assignment
Fi value of Pheromone Problem. IEEE Trans. Knowl. Data Eng. 11 (5), 769e778.
Mejia, G., Watson, A., Nordtvedt, J., 1996. Estimation of three-phase flow functions
R Unknown Parameter
in porous media. AIChE J. 42 (7), 1957e1967.
E Pheromone Evaporation Rate More, J.J., 1978. The Levenberg-Marquardt algorithm: implementation and theory.
N Number of Iteration In: Numerical Analysis. Springer, pp. 105e116.
Pi Probability Oak, M.J., 1989. Two-and Three-phase Relative Permeability of Water-wet Berea
Sandstoned Appendix to Report 89312ART0196. Rep. F89-P-74, 210-10-13,
k Total Number of Paths 89312ART0200.
r Random Number Riccardo, Poli, Kennedy, James, Blackwell, Tim, 2007. Particle swarm optimization.
Sw Water Saturation Swarm Intell. 1 (1), 33e57.
Razavi, S., Jalali-Farahani, F., 2010. Optimization and parameters estimation in pe-
So Oil Saturation troleum engineering problems using ant colony algorithm. J. Petroleum Sci. Eng.
Sg gas Saturation 74 (3), 147e153.
Sorw residual oil saturation in oilewater system Schembre, J.M., Kovscek, A.R., 2006. Estimation of dynamic relative permeability
and capillary pressure from countercurrent imbibition experiments. Transp.
Sorg residual oil saturation in oilegas system porous media 65 (1), 31e51.
Swi Connate water saturation Seçkiner, Serap Ulusam, Erog lu, Yunus, Emrullah, Merve, Dereli, Türkay, 2013. Ant
Krwo two-phase water relative permeability in oilewater colony optimization for continuous functions by using novel pheromone
updating. Appl. Math. Comput. 219 (9), 4163e4175.
system
Shahverdi, Hamidreza, Sohrabi, Mehran, Jamiolahmady, Mahmoud, 2011. A new
Krow two-phase oil relative permeability in oilewater system algorithm for estimating three-phase relative permeability from unsteady-state
Krog two-phase oil relative permeability in oilegas system core experiments. Transp. porous media 90 (3), 911e926.
Skjaeveland, S., et al., 2000. Capillary pressure correlation for mixed-wet reservoirs.
Krgo two-phase gas relative permeability in oilegas system
SPE Reserv. Eval. Eng. 3 (01), 60e67.
Pcow two-phase gas-oil capillary pressure Socha, K., Dorigo, M., 2008. Ant colony optimization for continuous domains. Eur. J.
Pcgo two-phase oil-water capillary pressure operational Res. 185 (3), 1155e1173.
Qo cumulative oil production Xiao, Jing, Li, LiangPing, 2011. A hybrid ant colony optimization for continuous
domains. Expert Syst. Appl. 38 (9), 11072e11077.
DPw pressure drop across the core in oil-water system Zhang, X.M., Wang, L.R., 2014. Two-parameter inversion of fluid-saturated porous
DPg pressure drop across the core in oil-gas system medium with Niche Ant Colony Algorithm. Math. Problems Eng. (4)