Development and Testing of Two-Phase

Relative Permeability Predictors Using

Artificial Neural Networks
N. Silpngarmlers, SPE, B. Guler,* SPE, T. Ertekin, SPE, and A.S. Grader, SPE, Penn State U.

Summary input parameters that have stronger influence on the output, are
In this paper, we report liquid/liquid and liquid/gas two-phase incorporated as additional inputs to the models. The models are
relative permeability predictors that are developed using artificial then trained and tested using the experimental data obtained from
neural networks (ANNs). In the development stage, some of the the literature. The results from the training and testing stages are
relative permeability data from literature are used during the train- presented for both models. The differences between both models
ing stage while some other sets are preserved to test the prediction are also discussed.
capabilities of the models. Various rock and fluid properties, in- In structuring the liquid/liquid (oil/water) ANN model, the con-
cluding endpoint saturations, porosity, permeability, viscosity and nection weights are updated in an incremental fashion so that the
interfacial tension, and some functional links (mathematical weights are adjusted according to a steepest descent protocol ob-
groups coupling various rock and fluid properties) constitute the tained after each data pattern is presented to the model. With the
input parameters of the models. The models are found to success- completion of the training session, performance of the model is
fully predict the field and experimental relative permeability data. tested using the data sets not exposed to the model. The relative
permeability characteristics predicted by the model for these data
Introduction
sets are found to be in good agreement with the data reported in the
Relative permeabilities are essential rock-fluid properties required literature. The input layer of the liquid/gas (oil/gas) ANN model
for almost all calculations of multiphase flow dynamics in porous with the exception of interfacial tension incorporates the same set
media. A good characterization of relative permeabilities enables of rock and fluid properties used in the liquid/liquid model. The
petroleum engineers to evaluate reservoir performance, forecast liquid/gas model differs from the liquid/liquid model in terms of its
ultimate recovery, and investigate the efficiency of improved oil re- network topology, types of functional links used, and the mechan-
covery techniques. Acquisition of accurate relative permeability data ics of how the connection weights are updated. The individual
is crucial and has always been of interest in the petroleum industry. topologies of the liquid/liquid and liquid/gas models are also dif-
Relative permeability characteristics can be obtained from ferent in terms of the number of neurons placed in the input and
laboratory measurement of a core sample or can be estimated using hidden layers. The more influential parameters are identified by
empirical correlations. Laboratory determinations of relative per- large connection weights that imply strong signals originating
meabilities are labor-intensive and can be complicated. The em- from those input neurons to the output neurons. The parameters
pirical models to estimate relative permeabilities based on rock identified in this fashion are used in structuring the functional
and fluid properties have experienced relatively mediocre success links. In the liquid/gas model, the connection weights are updated
owing to our limited understanding of the parameters and mecha- in a batch mode so that the steepest descent gradients observed
nisms that control the relative permeability characteristics. after presenting each data pattern to the model are averaged and
The two-phase relative permeabilities are direct nonlinear func- later used in adjusting the weights. During the testing stage, the
tions of phase saturations. They are known to be affected by sev- liquid/gas relative permeability predictor has also successfully pre-
eral other parameters other than phase saturations, such as satura- dicted the field and experimental data sets. The differences in
tion history, pore-size distribution and pore structure, wettability, network topologies and functional links of the two models imply
overburden pressure, porosity, permeability, interfacial tension, that liquid/liquid and liquid/gas relative permeability relationships
fluid density, fluid viscosity, initial wetting-phase saturation, im- are controlled by a variety of different parameters.
mobile third-phase saturation, and flow rate.1 The relative influ-
ences of these parameters on relative permeability characteristics
Overview of Artificial Neural Networks (ANNs)
are not yet clearly understood and quantified.
ANNs provide a powerful toolbox to perform nonlinear, mul- ANNs are developed by creating artificial neurons, which are
tidimensional interpolations. This feature of ANNs makes it pos- simple processing elements (PE) massively interconnected in order
sible to capture the existing nonlinear relationships that are most of to mimic a small portion of the serial- and parallel-information
the time not well understood between the input and output param- processing ability of the biological neural network. There are many
eters. Thus, ANNs can be effectively used to implicitly incorporate different types of ANNs, each of which has different strengths
the controlling mechanisms and parameters into the models which particular to its applications. The abilities of different networks can
are developed for relative permeability prediction. In this study, be related to their structure, dynamics, and learning methods.
two-phase, liquid/liquid, and liquid/gas relative permeability pre- ANNs can be used for pattern recognition, signal filtering, data
dictors are developed using a backpropagation algorithm. The segmentation, and so on. They offer the advantages of learning
methodology used in the development of two-phase liquid/liquid from examples, self-organization, fast data processing, and ease of
and two-phase liquid/gas relative permeability ANN predictors is insertion into existing and newly developed systems.2
described in this paper. Some of the available rock and fluid prop- In this study, two-phase liquid/liquid and liquid/gas relative
erties are used as input parameters of the models. Some functional permeability predictors were developed using a backpropagation
links, which represent various mathematical relationships among network (BPN), which is one type of feed-forward neural network
using generalized delta rule, which is a powerful learning rule. The
important characteristic of the BPN is found in its ability to map a
* Now with the U. of Texas at Austin.
set of input stimuli to a set of output by extracting the feature or
information from the input patterns. Its key distinguishing feature
Copyright 2002 Society of Petroleum Engineers
is that it can be designed and trained to accomplish a wide variety
This paper (SPE 79547) was revised for publication from paper SPE 69392, first presented of mappings, some of which are very complex.23 This can be
at the 2001 SPE Latin American and Caribbean Petroleum Engineering Conference, Bue-
nos Aires, 2528 March. Original manuscript received for review 19 April 2001. Revised
achieved because the hidden neurons in hidden layer(s) of the
manuscript received 8 May 2002. Manuscript peer approved 19 May 2002. network learn to respond to correlations or relationships among

September 2002 SPE Journal 299

different input parameters found in the input patterns. Because of range of [0,1] is used as a transfer characteristic for each neuron in
this characteristic, the ability of the network to generalize can be the hidden and output layers. The unipolar sigmoid function is
developed by training it with different exemplars. In the training used because it is bounded (between 0 and 1) and continuous and,
stage of a BPN, the information of input parameters is passed from therefore, differentiable everywhere with a positive slope. The
input layer to output layer. The calculated errors between the ac- function is given as
tual output and the output from the network are then propagated
back to adjust the connection weights in a sequential manner to 1
improve the predictive capabilities of the network. Once the error fy = . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (1)
becomes smaller than the specified criterion, the network is con- 1 + ey
sidered to be trained, the learning process is stopped, and connec-
tion weights are fixed. The network is now ready for the testing or The developed BPNs incorporate the momentum and learning rate
prediction stage in which the information will not be passed back- into the model in order to improve their performances.
ward. It should be noted that since the learning of BPN is super- The input-output patterns stored in the training database are
vised through the output neurons corresponding to its given input used in the training stage. The sets of input parameters are intro-
neurons, its performance is superior when a given pattern has duced one by one to the network, and calculated relative perme-
similar characteristics to those it has learned. In order for the ability values are compared with the experimental values corre-
network to respond correctly to the new pattern or correlation sponding to the presented set of input parameters. If a disparity is
existing between input and output, the examples of the new pattern observed, the error will be propagated back from the output layer
need to be given along with the corresponding outputs to retrain to adjust the connection weights between layers. For the oil/water
the network.2 relative permeability model, the connection weights are updated in
BPN is probably the most well known and widely used network an incremental or continuous mode. This implies that the weights
among the current types of neural network systems available. Even are adjusted according to a steepest-descent protocol established
though backpropagation algorithm is powerful and simple to after each training pattern was presented to the model. This train-
implement, BPNs have some drawbacks, such as slow conver- ing process continues until the average error of each pattern de-
gence and the possibility that the network converges to a local creases below a desired criterion.
minimum rather than the global minimum. Some improvements When the training is completed, the network is tested for its
can be made to overcome these drawbacks. The rate of conver- learning and generalization capabilities. The test for its learning
gence can be affected by a learning rate that determines how fast ability is conducted by testing its capability to reproduce or recall
a network will learn relationship between input and output pat- the outputs for the set of input that was used in the training. The
terns. The learning rate has a value between 0 and 1. The smaller test for its generalization ability is carried out by investigating its
the value of the learning rate, the slower the learning process will capability to predict the outputs for the input sets that were not
be. Adjusting the learning rate during the course of training can included in the training patterns. If the relative permeability values
accelerate the convergence. One approach in varying the learning predicted by the model are in good agreement with the experimen-
rate is to adjust it according to the performance of the algorithm, tally deduced relative permeability values, the network is then
so that the learning rate is decreased after each epoch (one pass of considered as having satisfactory generalization or predictive ca-
all input-output patterns) if the global error increases, and is in- pabilities. Otherwise, the structure of the network needs to be
creased when the global error decreases. The approach to prevent- modified in order to improve its prediction performance. Modifi-
ing the network not getting stuck in a local minimum is to incor- cations can be made by changing the number of neurons in the
porate a momentum term to the generalized delta rule. The mo- hidden layer, the momentum value, and the number of hidden
mentum term behaves like a low-pass filter that smoothes out the layers to search for a better structure using a somewhat heuristic
noise or oscillations in the weight surface while still keeping the approach. Another important modification that can be made to the
weight vector moving in the same direction. Another feature of ANN structure is the incorporation of the functional links that are
momentum is that it tends to accelerate convergence when the mathematical relationships combining the rock and fluid properties
weight vector is moving in a consistent direction (momentum has related to input neurons. The functional links are explicitly incor-
a constant value between 0 and 1). porated as additional input parameters to the model; therefore, they
bring in additional input neurons to the network. The network
Oil/Water Relative Permeability Predictor structure that gives the fastest convergence and the least average
In the development of a liquid/liquid (oil/water) relative perme- error is reserved as the final model. The four successive stages of
ability predictor,4 a large number of oil/water relative permeability the oil/water relative permeability model development are summa-
data are collected from literature. Some of these data sets are used rized in Fig. 1.
during the training stage, while the other sets are preserved to test Stage 1. In the initial stage of the development, the input pa-
the prediction ability of the model. rameters are the 10 rock and fluid properties listed earlier along
with additional functional links, which are inverses and squares of
Input Parameters. Because of the difficulty in finding two-phase each rock and fluid property. These input parameters constitute 30
relative permeability values along with all rock and fluid proper- neurons in the input layer. The output neurons of the network are
ties at the conditions of the experiments conducted, only the most oil and water relative permeabilities. Thus, each stored pattern
commonly reported and easily measurable rock and fluid proper- consists of 30 input parameters and 2 output values at the corre-
ties were selected as the main input parameters to the model. These sponding saturation value of each input pattern. Each stored pat-
selected parameters must also have significant influence to the tern corresponds to one point on relative permeability curve. It
relative permeability. The fluid properties chosen as the key input should be noted that for each relative permeability experiment,
parameters are viscosities and densities of phases, interfacial ten- except the water saturation, the other input parameters were con-
sion between the phases, and wetting-phase saturation. In addition, stant. Thirteen relative permeability experiments (84 patterns)
porosity, absolute permeability, residual oil saturation, and irre- were used in the training while two experiments (13 patterns) were
ducible water saturation are identified as the key input parameters preserved for the prediction purpose. The learning at this stage was
describing the rock properties. These two endpoint saturations found to be successful; however, the results from the prediction
are utilized in all relative permeability models because they rep- stage were not considered to be totally satisfactory. In particular,
resent the starting and the endpoints of two-phase relative perme- the oil relative permeability predictions were not accurate. In the
ability curves. next stage of development, we seek alternative ways to improve
the predictive capability of the network.
Model Development. Three-layer (input, hidden, and output lay- Stage 2. The network structure was modified by replacing the
ers) BPNs were developed as oil/water relative permeability pre- functional links incorporated in Stage 1 with the new nine func-
dictors. A unipolar sigmoid function producing outputs in the tional links that represent some combinations of the rock and fluid

300 September 2002 SPE Journal

 thus, each network gives an output of its own phase relative per-
meability. The number of training data sets was increased to 65
sets, which contain 588 patterns, while the same two data sets (13
patterns) were preserved to test the prediction capability of the
network. The new functional links known as normalized satura-
tions of water and oil were calculated using the definitions given
in Eqs. 2 and 3, respectively:

Sw Swirr
S*w = , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (2)
1 Swirr

1 Sw Sor
S*o = . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (3)
1 Sor

These normalized saturations scaled all of the other input pa-

rameters including other functional links. These networks are pre-
served as final two-phase, oil/water, relative permeability predic-
tors. A schematic representation of these networks (ANN-4) is
shown in Fig. 2. All of the modifications made were found to
improve the overall quality of the predictions. It should be noted
that the training time was again decreased by more than half,
which implies that the implemented modifications help the net-
work to learn much more effectively.
In the course of the study, four different submodels are gener-
ated as oil/water relative permeability predictors. Each submodel
has two individual relative permeability models representing the
Fig. 1Developmental stages of oil/water relative permeability
oil relative permeability and water relative permeability character-
ANN models (krwn1 and kron1 are the values of krw and kro
predicted at the previous saturation, respectively). istics. The differences among these submodels are the input pa-
rameters and their respective functional links incorporated to the
network. The construction of these sub-models is presented in
properties as shown in Table 1. At this stage, it was observed that Fig. 3. The ANN-4.1 is obtained by excluding phase viscosities
phase densities had no significant influence on the relative perme- and their functional links from the input parameter set of ANN-4.
ability values, but they made the training of the network more Excluding the interfacial tension and its functional links from
cumbersome. Thus, they were removed from the input parameter the input structure of ANN-4 generates ANN-4.2. ANN-4.3 is
set of the network. During this stage, the training database was generated by the exclusion of interfacial tension, phase viscosities,
expanded by including data sets from another 25 relative perme- and their functional links from the input structure of ANN-4.
ability experiments. The testing of the ANN yielded similar results Finally, when porosity and absolute permeability are excluded
to those in the previous stage, in which the learning was found to from the input structure of ANN-4.3, ANN-4.4 is obtained. ANN-4
be successful but the generalization was not uniformly satisfactory. is the most sophisticated model in terms of its required input
However, the training time was decreased by half. parameters. The simplest model is ANN-4.4 because it requires
Stage 3. In this stage of the development, the only modification only water saturation and endpoint saturations of both phases as
made to the Stage 2 network was to incorporate the relative per- input parameters.
meability values predicted at the previous saturation as additional
input parameters to the network. The motive behind this strategy
was the fact that relative permeabilities exhibit smoothly varying S* 1

characteristics. The knowledge of the relative permeability values

at the previous saturation can help determine the position of the S* Swirr 2

next data point. The inclusion of this history dependency of the S* Sor 3
two successive relative permeability values helped the network in 1
learning, so that the training time was found to decrease further. S* k 4
2
The overall quality of the prediction results is also somewhat im-
S* 5
proved, especially as the endpoint relative permeability values 3
were captured more accurately. S* ln ( w /o / wo) 6
Stage 4. During the final stage of the development, several 4
modifications are incorporated to the network. These modifica- S* ln( wo) 7
5
tions include the expansion of the training database and introduc- S* woSor kr
8 1 1
tion of new functional links that use normalized saturation. Two 6
*
separate networks were created for oil and water phases. Each S woSwirr 9
network is specialized in its own phase-relative permeability; *
7
S wo 10
8
S* ln ( w /o) 11
9
S* ( w /o)0.08 12
10
S* ( w /o) 13
0.08
S* o 14

*
S w 15

Fig. 2Schematic diagram of ANN-4.

September 2002 SPE Journal 301

 ANN-4 ANN-4 without ANN-4.1
phase viscosities
and related
functional links

ANN-4 without
interfacial tension ANN-4.2
and related
functional links

ANN-4 without phase

viscosities, interfacial ANN-4.3
tension, and related
functional links
from Fig. 4, ANN-4, which incorporates the largest number of
input parameters, gives the best prediction results.
ANN-4.3 without The average deviations between the predicted values and the
porosity and ANN-4.4 experimental data are calculated for each phase relative perme-
permeability ability model as plotted in Fig. 5. These average deviations are
computed over the total data sets used both in training and pre-
diction. From Fig. 5, it can be seen that ANN-4 gives the best
Fig. 3Final set of five oil/water relative permeability models. results for both oil and water models. The performance of the
model becomes attenuated when the number of input parameters
and functional links are decreased. However, the simplest model
The choice of which model to use depends on the availability (ANN-4.4) still gives reasonably satisfactory results.
of the rock and fluid properties. The input parameters and their Inclusion of New Training DataField Data. Four sets of
corresponding functional links required for each submodel is sum- water/oil relative permeability data provided by the Saudi Aramco
marized in Table 2. Laboratory Research and Development Center were used to test
Note that lists of the input parameters for oil relative perme- the predictive capability of the model using the ANN-4. The pre-
ability and water relative permeability models are the same, but dicted results without including any of the new data set are shown
entry values of these two models are different because of the as dashed lines in Figs. 6 through 9, and the diamonds denote the
normalized saturation for each phase, which multiplies other input experimental data. The solid lines represent the predicted values
parameters and functional links. when one of the new data sets was included in the training process.
The results indicate that when none of the new data sets were
incorporated in the training, predictions were not good because the
Performance of the Oil/Water Relative Permeability Models. network had not learned the characteristics of these new data sets.
The five final oil/water relative permeability predictors were For this reason, the predictions were significantly improved after
trained with available data. An experimental data set is preserved one of the new data sets was included in the training. In addition,
to test the prediction capability of the models. The prediction the inclusion of this new data set does not degrade the performance
results from the models are plotted against the experimental data as of the network; therefore, the new data sets can be added to the
shown in Fig. 4. The plots show that as the complexity of the existing database in order to expand the range of input parameters
model increases, the predicted relative permeability values are in and the relative permeability characteristics stored in the database.
better agreement with the experimental values. As can be seen The network is then retrained with the new training database.

Oil/Gas Relative Permeability Predictor. Our initial efforts on

0.7
the development of liquid/gas (oil/gas) relative permeability pre-
ANN-4.4 dictors included the use of the liquid/liquid (oil/water) relative
ANN-4.3
0.6
ANN-4.2
ANN-4.1
Relative Permeability

0.5 ANN-4
kro -Exp
krw -Exp
0.4

0.3

0.2

0.1

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7
Sw

Fig. 4Experimental and predicted relative permeabilities for

various network models. Fig. 5Average deviation of oil/water relative permeability models.

302 September 2002 SPE Journal

 1 1
Experimental kro
Experimental kro
0.9 0.9
Experimental krw
Relative Permeability 0.8 Experimental krw 0.8

Relative Permeability
ANN-4
0.7 ANN-4 0.7

0.6 0.6 ANN-4 with Data Set 1

ANN-4 with Data Set 1 in training
0.5 in training 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Sw Sw

Fig. 6Experimental and predicted relative permeabilities for Fig. 7Experimental and predicted relative permeabilities for
field Data Set 1. field Data Set 2.

permeability model to predict oil/gas relative permeabilities. The from the literature,58 while some of the data are provided by the
predictions that resulted from this initial approach were found to Saudi Aramco Laboratory Research and Development Center.
be totally unsatisfactory, indicating that the oil/water relative per-
meability model could not correctly capture the oil/gas relative Training Database. It is important to remember that prediction
permeability characteristics. This observation implies that the sys- capabilities of ANN models become attenuated when they are
tem properties and mechanisms that control the flow of fluids in forced to predict outside of the range of the training database
liquid/gas systems differ significantly from those found in liquid/ (extrapolation). Therefore, it is crucial to ensure that the training
liquid systems. Therefore, a different network structure, with dif- database covers the widest range of all of the input parameters. The
ferent functional links, is developed for oil/gas relative permeabil- general characteristics of the collected oil/gas relative permeability
ity models. data sets are summarized in Table 3. Nineteen experimental data
sets, which contain 242 patterns, are available. Out of the nineteen
Input Parameters. Absolute permeability and porosity are the data sets, sixteen data sets (208 patterns) were included in the
rock properties input for the model. Fluid properties that are input training, while three data sets (34 patterns) were set aside for
include phase viscosities and oil saturation. The information of prediction purposes. The data sets selected for the training data-
saturation at which each phase is immobile, residual oil saturation base ensure the coverage of the entire range of the available data.
and critical gas saturation, are also input. Some of the core samples
used in the relative permeability experiments were natural cores Model Development. In developing oil/gas relative permeability
retrieved from the reservoir with a small amount of immobile models, an approach similar to that used in oil/water relative per-
brine. Therefore, irreducible water saturation is also included in the meability models has been exercised. Moreover, the knowledge
set of input parameters. Some of the experimental data are obtained and experience acquired from development of oil/water models are

1 1
Experimental kro Experimental k ro
0.9 0.9
Experimental k rw
0.8 Experimental krw 0.8
Relative Permeability

A N N -4
Relative Permeability

0.7 ANN-4 0.7

A N N -4 w ith D ata Set 1
0.6 0.6 in training
ANN-4 with Data Set 1
0.5 in training 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Sw Sw

Fig. 8Experimental and predicted relative permeabilities for Fig. 9Experimental and predicted relative permeabilities for
field Data Set 3. field Data Set 4.

September 2002 SPE Journal 303

 approaches, batch (or periodic) and incremental (or continuous),
imported for implementation in the oil/gas model. Again, two are essentially the same. It has been argued that in the incremental
separate models are constructed for oil and gas relative perme- weight-update approach, one particular pattern may lead the
abilities. Both of these models use backpropagation networks con- weight vectors into the local minima when the learning rate is
sisting of three layers (input, hidden, and output). A unipolar sigmoid large. However, there are possibilities that in using this approach,
function is applied as a transfer function in both models. The learning one pattern may move the weights closer to the global minima or
of the network is achieved by continuous or incremental updating of to the desired solution.9
connection weights based on the steepest descent protocol. After the weight update process is modified, the search for
Stage 1. Initially, eight input parameters were used to provide additional helpful functional links is continued. An additional
input stimuli to both networks. The corresponding output from functional link that was observed to be helpful for both models is
each network is its phase-relative permeability at a given oil satu- normalized porosity, which is defined by
ration. Similar to the oil/water models, the input patterns are pre-
sented to the network one by one. Each input pattern corresponds * = 1 Sor Swirr. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (6)
to a point on the relative permeability curve. The best model Stage 3. In this stage, the goal was to search for more helpful
topologies were searched by varying number of neurons in the functional links for both models. The approach was to first identify
hidden layer, the momentum constant, and the learning rate. It was the input parameters that are more influential to the outputs. Com-
observed that structures for both models were different in terms of paring the final connection weights attained at the end of a suc-
the number of neurons in the hidden layer and the momentum cessful training session identifies the more influential parameters.
constant used. It was further observed that while the learning ses- The large connection weights imply strong signals between the
sions were successful, the generalization was not satisfactory. Sev- two connected neurons. The three largest weights, from the output
eral functional links, including the relative permeability values neuron to the hidden neurons, and then to the input neurons, are
obtained at the previous oil saturation value, were considered for identified. The results indicate that the input parameters more in-
their potential contributions to improve the learning and general- fluential to the oil relative permeability are critical gas saturation
ization capabilities of the networks. Our analysis along these lines (Sgc) and the square of the normalized oil saturation (So*2). The
indicates that functional links, which helped improve prediction more influential input parameters to gas relative permeability are
capability of the two models, were different. Functional links that identified as viscosity of gas and the square of the normalized gas
were found to improve the prediction performance of the oil rela- saturation (Sog*2). Parameters identified in this fashion are used in
tive permeability model are normalized saturation (So*) and its constructing the additional functional links, which magnify the effect
square, as well as the ratio of viscosity of gas to viscosity of oil of such input parameters on the output. The functional links incor-
(g/o). The functional links that help the learning of the gas porated in each relative permeability model are shown in Table 4.
relative permeability model are normalized saturation (Sog*), its Stage 4. The artificial neural networks developed so far per-
square, and the ratio of permeability to porosity (k/). The defi- formed reasonably well in initial experimentation. However, net-
nitions of the two normalized saturations used in these models are works still experienced some degree of difficulty in capturing the
slightly different, as shown in Eqs. 4 and 5: slopes of the relative permeability curves in some cases. Therefore,
So Sor at this stage of development, we focused on providing information
S*o = , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (4) on the slope of each relative permeability curve to each network.
1 Sor This information is computed between two consecutive points on
So Sor the curve. Because this piece of information is not available be-
S*og = . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (5) forehand in predicting relative permeability values, the slope cannot
1 Sgc Sor
be included as input parameters of the network. However, it can be
The observations made in Stage 1 again support our strategy of used as an additional output so that the network can extract the rela-
constructing a separate model for each phase as each model uses tionship between the relative permeability value and the slope asso-
different functional links; because these functional links are uti- ciated with it at a given saturation. This implementation was found to
lized as additional input parameters to the models, they bring in yield better prediction performance of both networks. The four stages
additional neurons into the input layer. of development used in oil/gas relative permeability models can be
Stage 2. In the first stage of the development, even though the summarized as shown in Fig. 10.
prediction capability of the models was improved, there were some The structures of the final oil relative permeability and gas
fluctuations in the predicted values of relative permeabilities. In relative permeability models are shown in Figs. 11 and 12, re-
other words, the predicted relative permeability curves exhibited a spectively. Eight rock and fluid properties in conjunction with 13
somewhat nonmonotonically increasing or decreasing behavior. At functional links, as indicated in Figs. 11 and 12, constitute input
this stage, a different approach in updating the connection weights parameters of the models.
is investigated by implementing a batch or periodic weight-
updating scheme. This is accomplished by averaging the steepest Performance of the Oil/Gas Relative Permeability Models. The
descent gradients, calculated after presenting each training pattern predictive capabilities of the oil and gas relative permeability mod-
together, to produce a more accurate estimate of the gradient. The els were also tested against experimental and field data. The pre-
connection weights are then adjusted according to the average diction results are shown in Figs. 13 through 16. The solid lines
gradient after all of the patterns are introduced to the network. This denote the predicted relative permeability values, and the dia-
approach was found to be effective in eliminating the fluctuations monds represent the actual data. Fig. 13 shows the results from the
in the predicted relative permeability values. Theoretically, the two learning test of the network by reproducing the data set that was

304 September 2002 SPE Journal

 ( )

Functional

Fig. 11Schematic diagram of the kro model of the oil/gas rela-

tive permeability predictor.

core sample used in obtaining experimental relative permeability

data shown in Fig. 15 is a limestone core. A composite core re-
trieved from a reservoir is used to obtain experimental relative
permeability data presented in Fig. 16. It can be seen from the
prediction plots (Figs. 14 through 16) that both models yield ac-
ceptable results in generating the experimental and field data
Fig. 10Developmental stages of oil/gas relative permeability within a certain margin of error. From the four relative permeabil-
ANN models. ity data sets shown in Figs. 13 through 16, it can be noted that
relative permeability curves studied exhibit significantly different
characteristics in terms of their endpoint saturations, magnitudes of
used in the training stage. The results show an almost perfect recall the relative permeability values, and crossover points of the curves.
performance of the models. This exercise shows that the developed models perform well for a
Figs. 14 through 16 show the results from the generalization relatively wide spectrum of relative permeability characteristics.
test in which the network was used to predict the data sets that it
was not exposed to during the training session. Note that the ex- Model Performance Comparison with the Existing Correla-
perimental data shown in Fig. 14 are obtained from a synthetic tions. The prediction performances of the models are compared
core sample composed of polytetrafluoroethylene (PTFE). The with the existing correlations, which are Coreys and Honarpours
correlations.1 Coreys correlations are chosen for the comparison
because they represent one of the more commonly used correla-

1
0.9 Experimental kro
irr

0.8 Experimental krg

Relative Permeability

ANN predicted
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1
So

Fig. 12Schematic diagram of the krg model of the oil/gas rela- Fig. 13Experimental and predicted oil/gas relative permeabil-
tive permeability predictor. ity data for a data set included in the training.

September 2002 SPE Journal 305

 1 1
0.9 Experimental kro 0.9 Experimental kro

Relative Permeability
Relative Permeability
0.8 Experimental krg 0.8 Experimental krg
ANN predicted 0.7
0.7 ANN predicted
0.6
0.6
0.5
0.5
0.4
0.4 0.3
0.3 0.2
0.2 0.1
0.1 0
0 0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1 So

So Fig. 15Experimental and predicted oil/gas relative perme-

ability data for a data set preserved for prediction (limestone
Fig. 14Experimental and predicted oil/gas relative perme- core data).
ability data for a data set preserved for prediction (synthetic
core data).
should be sufficient for the network to learn the existing relation-
ships among the input and output parameters. If the training data
tions in estimating two-phase relative permeabilities. Honarpours covers a wide range of characteristics, the networks will develop
correlations were developed from a large number of experimental more effective generalization abilities because they have been ex-
data specifically for gas/oil relative permeabilities. Moreover, both posed to a broader spectrum of characteristics that can be encoun-
Coreys and Honarpours correlations require only the rock and tered. Besides the number of training data sets, the quality of the
fluid properties available in this study. The comparative perfor- data is also very important, because the networks are trying to
mances of the three models can be seen in Figs. 17 through 19. extract the existing information from the data sets used in training.
From these comparisons, it can be seen that the artificial neural Erroneous information can lead the network in the wrong direction
networks developed in this study give better overall predictions or adversely affect the networks performance in learning. For this
than the correlations for the cases presented here. This observa- reason, the incorporation of data sets into the training database
tion reinforces the attractive feature of the networks as general- should be done with caution. The quality and consistency of a new
ized approximators. data set can be roughly investigated by including the new data set
in the training database. After the training is terminated (either by
Concept Analysis the error criteria or the number of epochs), the network is asked to
This study shows that ANNs as soft computing algorithms can be reproduce the same data set. If the network experiences difficulties
applied to obtain alternative solutions for problems in the petro- in reproducing a particular data set while most of the other data
leum industry. In this study we focus on obtaining the accurate sets are recalled correctly, then the ANN practitioner should be
two-phase relative permeability data using the available rock and concerned about the quality of the new data set. A similar heuristic
fluid properties. BPNs can be used for this purpose because of their approach can involve the inclusion of the new data set in training
ability in mapping input stimuli with corresponding output(s). This and comparing the networks performance before and after the
can be achieved through the ability of the network in learning the introduction of the new data set. If the networks performance
existing relationships among the input parameters and the out- remains the same or improves, the new data set can be confidently
put(s). Because ANNs are data-based algorithms, the data used in used in expanding the networks capabilities.
training the networks are crucial both quantitatively and qualita- The learning of the networks is also affected by the input pa-
tively. The number of data sets included in the training database rameters presented to the network and by the mode of presentation.
Identification of the important parameters can be achieved through
the knowledge in the subject matter coupled with a somewhat
1
0.9 Experimental kro 1
Relative Permeability

0.8 Experimental krg 0.9

0.7
Relative Permeability

ANN predicted 0.8

0.6 0.7
0.5 0.6
0.4 0.5 Experimental kro

0.3 Experimental krg

0.4
ANN predicted
0.2 0.3
Corey correlation
0.1 0.2
Honarpour correlation
0 0.1
0 0.2 0.4 0.6 0.8 1 0
So 0 0.2 0.4 0.6 0.8 1
So
Fig. 16Experimental and predicted oil/gas relative perme-
ability data for a data set preserved for prediction (composite Fig. 17Model performance comparison for the synthetic
core data). core sample.

306 September 2002 SPE Journal

 1 After the error decreases below a prespecified value, the training is
terminated, and the network performance is tested using the inde-
0.9 Experimental kro pendent testing data. Because the testing data set is independent of

Relative Permeability
0.8 Experimental krg the training data, the result from the test is a good measurement of
0.7 generalization of the network. Using this approach, if the network
ANN learns the underlying relationship of the data, the testing perfor-
0.6
Corey correlation mance of the network will improve with training. However, if the
0.5 network stops learning the common characteristics of any data set,
0.4 Honarpour correlation but instead learns characteristics that are only true for the training
0.3 data sets, performance on the testing set will degrade. In such
situations, the convergence criterion for the error should be relaxed
0.2 so that the training will be stopped earlier.
0.1
0 Summary and Conclusions
0 0.2 0.4 0.6 0.8 1
The ANN-based relative permeability models can be used to mini-
So mize the number of relative permeability experiments. Relatively
small experimental data points (along with rock and fluid proper-
Fig. 18Model performance comparison for the limestone ties) are needed to train the network to generate a complete set of
core sample. relative permeability curves. Moreover, the models can be used to
estimate the two-phase relative permeabilities at any saturation
value in the reservoir using the information learned from a few
heuristic approach. The importance of a parameter can be evalu- core samples extracted from different regions of the reservoir. This
ated by investigating the networks performance when such a pa- becomes extremely useful in reservoir simulation studies, as it
rameter is input into the network. Another approach is to compare gives flexibility in generating separate relative permeability data
the magnitudes of the connection weights between the neurons for different regions of the reservoir. The models have a
from the output layer back to the input layer. The large connection dynamic learning feature, which allows them to learn additional char-
weights indicate strong signals originating from the input neuron acteristics of the relative permeabilities through the expansion of the
to the output neuron, which imply the influence of such input training database as more experimental data become available.
parameters to the output. Focusing on the way the influential input In this paper, we presented five artificial neural networks that
parameters are presented to the network can then enhance the have been developed as oil/water relative permeability predictors.
network performance. This approach is also exercised in this study. Each model uses a different set of input parameters and functional
From this study, it was also found that the networks that exhibit links. The oil/gas relative permeability predictors using the most
perfect recalls do not always give good predictions. This is espe- commonly reported rock and fluid properties are also developed.
cially true if the training data contain some noises and a small All data used in the studies are obtained from relative permeability
convergence criterion is used. In the presence of a small conver- experiments. Following are the observations and conclusions that
gence criterion, the networks are forced to learn and match the data have been derived from these exercises:
that contain noises. Thus, the network can become overtrained, 1. Property-based functional links help in improving the effi-
which leads to the memorization of the data instead of the learning ciency of the ANNs.
of the existing relationships. In such cases, there is a tradeoff 2. Incorporating normalized saturations captures relative perme-
between the test and prediction results. This implies that when the ability values more accurately.
networks give a perfect match for the data used in the training, the 3. Constructing separate models for each phase provides an op-
quality of prediction is decreased. On the contrary, when the con- portunity to focus on the controlling parameters for each phase.
vergence criterion is relaxed, the reproduction of the data used in 4. Submodels increase the applicability and practicality of the
the training is degraded, but prediction results are improved. This proposed approach.
observation reinforces the importance of the quality of the training 5. The factors that control the flow dynamics in liquid/liquid sys-
data. There are many ways to avoid or prevent the problem of tems are observed to be different from those in liquid/gas systems.
overfitting. One of the simplest and most widely used methods 6. In liquid/gas systems, the parameters thatinfluence the flow of
of avoiding overfitting, and which is used in this study, is the early gas are different from those influential to oil flow.
termination of training. The implementation of this method is to 7. Batch weight updating approach was found to yield better re-
divide the data into two groups, which are training sets and testing sults for oil/gas relative permeability models. This observation
sets. The networks are trained using only data in training sets. may be specific to the quality of the collected experimental
data.
8. Reliable relative permeability data are crucial to the learning
1
and prediction performance of the networks.
0.9 Experimental kro 9. ANNs possess dynamic training database so that their range of
Relative Permeability

0.8 Experimental krg applicability can be extended by incorporating more experi-

0.7 ANN mental data as they become available.
0.6 Corey correlation 10. ANN has the potential of providing a toolbox for identifying
the parameters controlling relative permeability characteristics.
0.5 Honarpour correlation
0.4
0.3 Nomenclature
0.2 k absolute permeability, L2, md
0.1 kri relative permeability to phase i, fraction
0 Si saturation of phase i, fraction
0 0.2 0.4 0.6 0.8 1 y net incoming input to a neuron
So i viscosity of phase i, m/Lt, cp
ij interfacial tension between phases i and j, m/t2,
Fig. 19Model performance comparison for the composite dyne/cm
core sample. porosity, percent

September 2002 SPE Journal 307

Subscripts 9. Bose, N.K. and Liang, P.: Neural Network Fundamentals with Graphs,
c critical Algorithms, and Applications, McGraw-Hill, New York City (1996).
g gas
irr irreducible
o oil
r residual SI Metric Conversion Factors
w water cp 1.0* E03 Pas
Superscripts dyne 1.0* E02 mN
* normalized md 9.869233 E04 m2
*Conversion factor is exact.
Acknowledgments
We acknowledge the partial financial support provided by the
Petroleum and Natural Gas Engineering program of the Dept. of
Energy and Geo-Environmental Engineering at Pennsylvania State Nuntawan Silpngarmlers is a PhD candidate in the petroleum
U. and the Consortium for Virtual Operations Research (CVOR) and natural gas engineering (PNGE) program at Pennsylvania
housed within Penn State and West Virginia universities. We ex- State U. e-mail: [email protected]. Her main research inter-
tend our gratitude to Saudi Aramco Research and Development ests are in the area of numerical reservoir simulation, well-test
Center for providing experimental relative permeability data sets analysis, EOR, and applications of artificial neural networks in
reservoir engineering. Silpngarmlers holds a BSc degree in
used in training and testing of the relative permeability models
chemistry from Kasetsart U. in Bangkok, Thailand, and an MS
developed in this study. degree in petroleum and natural gas engineering from Penn-
sylvania State U. Baris Guler is a research assistant at the Cen-
References
ter for Petroleum and Geosystems Engineering at the U. of
1. Honarpour, M., Koederitz, L., and Harvey, A.H.: Relative Permeability Texas at Austin, where he is working toward his PhD degree. His
of Petroleum Reservoirs, CRC Press Inc., Boca Raton, Florida (1986). research interests include reservoir engineering, modeling,
2. Maren, A.J., Harston, C.T., and Pap, R.M.: Handbook of Neural Com- simulation, EOR, and high-performance computing. Guler
puting Applications, Academic Press Inc., San Diego, California holds a BS degree in PNGE from the Middle East Technical U.,
(1990). Ankara, Turkey, and an MS degree in PNGE from Pennsylvania
3. Hagan, M.T., Demuth, H.B., and Beale.M.: Neural Network Design, State U. Turgay Ertekin is a professor of PNGE and George E.
ITP/PWS Press, Boston, Massachusetts (1996). Trimble Chair at Pennsylvania State U., where he has been a
faculty member since 1978. His main research interests are in
4. Guler, B., Ertekin, T., and Grader, A.S.: An Artificial Neural Network
the area of mathematical modeling of fluid-flow dynamics in
Based Relative Permeability Predictor, paper 99-91 presented at the porous media, coalbed methane reservoirs, well-test analysis,
1999 Petroleum Conference of the South Saskatchewan Section, The and artificial neural network applications to reservoir engineer-
Petroleum Society of CIM, Regina, Canada, 1821 October. ing problems. Ertekin holds BS and MS degrees in petroleum
5. Stegemeier, G.L.: The Relationship of Relative Permeability to Con- engineering from Middle East Technical U., Ankara, Turkey,
tact Angles, MS thesis, U. of Texas at Austin, Austin, Texas (1954). and a PhD degree in PNGE from Pennsylvania State U. Ertekin
6. Hvolboll, V.T.: The Effect of Overburden Pressure on Relative Per- is the 1998 recipient of the SPE Distinguished Achievement
meability, MS thesis, U. of Oklahoma, Norman, Oklahoma (1955). Award for Petroleum Engineering Faculty and the 2001 recipi-
7. McCaffery, F.G.: The Effect of Wettability on Relative Permeability ent of the SPE Lester C. Uren Award. A.S. Grader is a professor
of PNGE at Pennsylvania State U. His research interests include
and Imbibition in Porous Media, PhD dissertation, U. of Calgary,
multiphase fluid flow in fractured and nonfractured porous me-
Calgary (1973). dia, high-resolution CT imaging of fluid-flow processes, well
8. Asar, H.K.: Influence of Interfacial Tension on Gas-Oil Relative Per- testing, and reservoir engineering. He holds a BSc degree from
meability in a Gas-Condensate System, PhD dissertation, U. of South- the Technion, Israel Inst. of Technology, and MS and PhD de-
ern California, Los Angeles (1980). grees from Stanford U.

308 September 2002 SPE Journal