SPE 98012

Reservoir Characterization Using Intelligent Seismic Inversion

Emre Artun, SPE, West Virginia University; Shahab D. Mohaghegh, SPE, West Virginia University; Jaime Toro, West
Virginia University; Tom Wilson, West Virginia University; Alejandro Sanchez, Anadarko Petroleum Corporation

Copyright 2005, Society of Petroleum Engineers

drilling new wells during field development.
This paper was prepared for presentation at the 2005 SPE Eastern Regional Meeting held in
Morgantown, W.V., 14–16 September 2005.
Introduction
This paper was selected for presentation by an SPE Program Committee following review of
information contained in a proposal submitted by the author(s). Contents of the paper, as
Reservoir characterization requires building a spatial model of
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to the reservoir by using appropriate data gathered from previous
correction by the author(s). The material, as presented, does not necessarily reflect any
position of the Society of Petroleum Engineers, its officers, or members. Papers presented at studies. This spatial model is then used in flow simulators,
SPE meetings are subject to publication review by Editorial Committees of the Society of
Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper
which can predict reservoir performance. An accurate and
for commercial purposes without the written consent of the Society of Petroleum Engineers is reliable reservoir characterization study is indispensable in
prohibited. Permission to reproduce in print is restricted to a proposal of not more than 300
words; illustrations may not be copied. The proposal must contain conspicuous reservoir management. The major challenge in today’s
acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O.
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.
reservoir characterization is to integrate all different kinds of
data to obtain an accurate and high-resolution reservoir model.
Abstract The concept of data analysis forms the basis of reservoir
Today, the major challenge in reservoir characterization is characterization. Uncertainty, unreliability, and large variety
integrating data coming from different sources in varying of scales due to the different origins of the data must be taken
scales, in order to obtain an accurate and high-resolution into consideration. Together with the immense size of the data
reservoir model. The role of seismic data in this integration is sets that must be dealt with, these issues bring complex
often limited to providing a structural model for the reservoir. problems, which are hard to address with conventional tools.
Its relatively low resolution usually limits its further use. That’s why unconventional computation tools have gained
However, its areal coverage and availability suggest that it has much interest in data analysis in recent years. Among those
the potential of providing valuable data for more detailed modern tools; intelligent systems, which mimic the
reservoir characterization studies through the process of mechanism of the human mind, are a way of dealing with
seismic inversion. In this paper, a novel intelligent seismic imprecision and partial truth1. It should not surprise us that
inversion methodology is presented to achieve a desirable using intelligent systems in reservoir characterization studies
correlation between relatively low-frequency seismic signals, has become a widely-used method in the petroleum
and the much higher frequency wireline-log data. Vertical engineering literature. Some previous intelligent reservoir
seismic profile (VSP) is used as an intermediate step between characterization applications include, but are not limited to,
the well logs and the surface seismic. A synthetic seismic synthetic log generation2,3,4, permeability estimation from
model is developed by using real data and seismic logs5,6, and predicting bulk volume of oil7.
interpretation. In the example presented here, the model Let us consider different types of data used in reservoir
represents the Atoka and Morrow formations, and the characterization: core samples provide very high resolution
overlying Pennsylvanian sequence of the Buffalo Valley Field information about the reservoir (fraction of inches), while
in New Mexico. Generalized regression neural network seismic data have a resolution in tens of feet, and well logs
(GRNN) is used to build two independent correlation models have in one of inches. Because of its low resolution, seismic
between; 1) Surface seismic and VSP, 2) VSP and well logs. data is routinely used only to attain a structural view of the
After generating virtual VSP’s from the surface seismic, well reservoir. On the other hand, unlike core samples or well logs,
logs are predicted by using the correlation between VSP and which are only available at isolated localities of a reservoir,
well logs. The values of the density log, which is a surrogate seismic data frequently provides 3D coverage over a large
for reservoir porosity, are predicted for each seismic trace area. Because of this areal coverage, researchers have always
through the seismic line with a classification approach having aimed to use seismic data in reservoir description. Inverse
a correlation coefficient of 0.81. The same methodology is modeling of reservoir properties from the seismic data is
then applied to real data taken from the Buffalo Valley Field, known as seismic inversion in the literature. The process
to predict interwell gamma ray and neutron porosity logs presented in this paper includes modeling of the well logs
through the seismic line of interest. The same procedure can from seismic data, which is also an inverse modeling process
be applied to a complete 3D seismic block to obtain 3D (Figure 1). This approach attracts a lot of interest and is very
distributions of reservoir properties with less uncertainty than important because of the necessary shift from exploration to
the geostatistical estimation methods. The intelligent seismic development of existing fields8.
inversion method should help to increase the success of Seismic inversion has been applied by several authors with
different approaches. Hampson et al.8 and Leiphart and Hart9
2 Reservoir Characterization Using Intelligent Seismic Inversion SPE 98012

have compared different techniques such as multiple linear Theoretical Background

regression, backpropagation and probabilistic neural networks. Seismic Surveys. The seismic method is the most widely used
They have predicted porosity logs from seismic attributes and tool in the exploration of hydrocarbon reservoirs. It is useful
both have suggested using probabilistic neural networks in this for obtaining a structural view of the subsurface geology. The
type of problems considering its mathematical simplicity and basic theory of the seismic method is based on the movements
success. of signals through the subsurface. Reflection method is the
Balch et al.10 have used fuzzy ranking to see which type of most widely used seismic method, which is useful in
seismic attribute is related to the target reservoir property. identifying formation tops. A seismic trace is the response of a
They have modeled correlations between those selected single seismic detector to the seismic energy propagation
attributes and porosity, water saturation and net pay thickness through the earth. If these traces are displayed side-by-side,
by using a backpropagation neural network. Chawathe et al.11 then it is called a seismic record. A processing stage comes
used neural networks to predict gamma ray log from seismic after obtaining a seismic record, to enhance the signal, to
attributes; amplitude, phase, frequency, reflection strength, minimize the noise and increase the resolution. These
and quadrature. However, they have used higher-resolution processed images are than compiled together to produce the
crosswell seismic data instead of surface seismic as a new final output of the seismic survey: a seismic section14.
approach. Soto and Holditch12 have used the same types of Vertical seismic profiling (VSP) differs from conventional
attributes from surface seismic to predict the gamma ray log seismic surveys in the location of signal receivers. In VSP
with neural networks. surveys, the receivers are located in the borehole instead of at
Reeves et al.13 introduced a new methodology, which the earth’s surface. Because the earth acts as a low-pass filter,
divides the whole seismic inversion problem into two parts. placing of the receivers at depth reduces the distance that the
They have considered cross-well tomography as an signal has to travel through the earth, thus yielding higher
intermediate step in their procedure, after finding a correlation frequency (higher resolution) data. VSP surveys are very
between surface seismic and cross-well seismic. They have similar to velocity surveys in terms of where the sources and
suggested producing virtual cross-well seismic data, before receivers are located. However, they differ from each other
dealing with logs. Giving Chawathe et al.’s11 work as an with two issues14:
example, they stated that well logs can easily be predicted 1) The distance between geophone recording depths (smaller
from virtual cross-well seismic data. According to the authors, in VSP, every 15-40 meters)
using crosswell seismic as an intermediate scale data can 2) Collection of information (Only first break times are
provide improved vertical resolution, increase constraints and collected in velocity surveys. In VSP, upgoing and downgoing
reduce the uncertainty of reservoir description. events are also collected for several seconds.)
In this study, a similar methodology is followed. Instead of Seismic attributes are all the information obtained from
cross-well seismic which is rather hard to obtain, vertical seismic data, either by direct measurements or by logical or
seismic profile (VSP) is incorporated into the study as the experience-based reasoning15. The attributes used in this study
intermediate scale data. This is due to the fact that VSP is can be briefly defined as;
available more frequently, and is less expensive to obtain than
cross-well tomography. It is a common type of data that can Amplitude: Measure of the strength of the reflected signal.
be found in many fields. Indicates changes in physical properties of various lithological
Together with the integration of a third type of data, entities. It can sometimes be used to detect gas presence11.
another unique feature of this study was developing and Hilbert Transform: This amounts to a 90-degree phase
integrating a synthetic model to the research, before dealing rotation. Amplitude and Hilbert transform are combined as
with real data. Having a synthetic model that we had full Cartesian components of a trace signal16.
control of gave us the opportunity to develop and test the Instantaneous Phase: Phase angles range from -180 degrees to
proposed methodology better before applying it to real data. +180 degrees. Envelope and phase are combined as polar
Our synthetic model represents the gas-producing Atoka and components of a trace signal16.
Morrow formations and the overlying Pennsylvanian sequence Average Energy: This attribute integrates the envelope
in the Buffalo Valley Field in New Mexico. Surface seismic between paraphase events. It highlights stratigraphic detail
and VSP responses of this model are computed. Artificial through energy fluctuations across traces. Values are in
neural networks are used to develop two independent degrees16.
correlation models between; 1) Surface seismic and VSP, 2) Envelope: Represents the reflection strength. The envelope is
VSP and well logs. Density log has been selected as the target independent of the phase and it relates directly to the acoustic
log, and is predicted from the seismic line. In the second case, impedance contrasts16.
seismic field data have been used to predict gamma ray and Frequency: This attribute describes how long it takes the
neutron porosity distributions through a seismic section. phase to complete 360 degrees of rotation16.
In the following sections a theoretical background, which Paraphase: This attribute is the instantaneous phase with the
includes brief explanations of seismic surveys and artificial predictable trend removed. As such, it assists visualizing the
neural networks -under the light of the methodology of this structural picture because phase tracks geologic boundaries16.
study- are included. Then, the methodologies followed in the
synthetic and the field cases are presented. After that, the Artificial Neural Networks. Artificial neural networks
results and their discussions are followed by conclusions. (ANN) can be broadly defined as information processing
systems that mimic the human mind as a mathematical model
SPE 98012 Artun, Mohaghegh, Toro, Wilson & Sanchez 3

representation of the biological neural networks. ANN have

gained an increasing popularity in different fields of p
engineering in the past few decades, because of their C i = ∑ X j − X ij (3)
capability of extracting complex and non-linear relationships. j =1

Their mechanism is based on the following assumptions17: n

⎛ C ⎞
1) Information processing occurs in many simple elements ∑Y i
exp⎜⎜ − i ⎟⎟
that are called neurons (processing elements). Yˆ ( X ) =
i =1 ⎝ σ ⎠ (4)
2) Signals are passed between neurons over connection links. n
⎛ C ⎞
3) Each connection link has an associated weight, which, in a ∑ exp⎜⎜ − i ⎟⎟
typical neural network, multiplies the signal being transmitted. i =1 ⎝ σ ⎠
4) Each neuron applies an activation function (usually non-
linear) to its net input to determine its output signal. The estimate Y(X) is defined as a weighted average of the
Figure 2 shows a typical neuron (processing element). observed values, Yi, where each observed value is weighted
Outputs (In) coming from another neuron are multiplied by exponentially according to its Euclidian or city block
their corresponding weights (Wn), and summed up. An distance21. σ is the smoothing factor, and optimum smoothing
activation function is then applied to the summation, and the factor is determined after several runs according to the mean-
output of that neuron is now calculated and ready to be squared error of the estimate, which must be kept at minimum.
transferred to another neuron18. There are many different types This process is referred to as the training of the network. If a
of neural network architectures and algorithms available. In number of iterations pass with no improvement in the mean-
this study, a generalized regression neural network (GRNN) is squared error, that smoothing factor is determined as the
used. GRNN is a modification to probabilistic neural network optimum one for that data set. In the production phase, that
that has been suggested by authors, who have studied seismic smoothing factor is applied to data sets that the network has
inversion8,9. GRNN has also been successfully used in not seen before. While applying the network to a new set of
geological pattern recognition applications such as synthetic data, increasing the smoothing factor would result in
log generation10 and total organic carbon content prediction decreasing the range of output values.
from logs19. Besides, GRNN has been used in finding a GRNN is known to be particularly useful in approximating
correlation between cross-well seismic and surface continuous functions, such as logs, or other types of geological
seismic13,20. Huang et al.19 described GRNN as an easy-to- patterns. It may have multidimensional input, and it will fit
implement tool, which has efficient training capabilities, and multidimensional surfaces through data22. It is a three-layer
the ability to handle incomplete patterns. network. In the hidden layer, there must be one hidden neuron
for each training pattern.
Generalized Regression Neural Network. Introduced by
Specht21 in 1991, GRNN is a one-pass learning algorithm with Case 1: Synthetic Model Study
a highly parallel structure. It is a memory-based network, Description of the model. The model was developed using
which provides estimates of continuous variables, and Struct, a modeling package in the Geographix DiscoveryTM
converges to the underlying regression surface. This approach Suite of Landmark Graphics®. It is a comparative example of
is freed from the necessity of assuming a specific functional the stratigraphic section of the Buffalo Valley Field, which
form. Instead, the appropriate form is expressed as a includes Atoka and Morrow formations together with the
probability density function (pdf), which can be determined overlying Pennsylvanian sequence, where stratigraphic
from the observed data. General regression uses y (a scalar complexity increases with depth. It is developed by using a
random variable), the X (a particular measured value of a forward modeling process, which has simulated straight rays
vector random variable x), and the non-parametric estimator of traveling from the surface and avoiding diffraction at
the joint probability density function f(x, y). After defining the interfering events. The model has been defined with properties
scalar Euclidian distance function, Di2; like thickness, geometry, lateral distribution, density and
interval velocity of the rocks, which were gathered from the
Di2 = ( X − X i ) T ( X − X i ) (1) actual data set23:
Horizontal dimension - 4.9 mi (7.9 km), equivalent to the
dimension of the 3D seismic survey.
performing the integrations results with the following:
Vertical dimension - Real depth - approximately 7000 ft; and
n
⎛ D i2 ⎞ thickness of the sequence – approximately 2000 ft. This data
∑ Y i
exp ⎜⎜ − ⎟
2σ ⎟⎠ was measured from well logs.
ˆ
Y (X ) =
i =1 ⎝ (2) Geometry of the sand channels and sand/limestone layers -
n
⎛ D i2 ⎞
∑ exp ⎜−
⎜ 2σ ⎟
⎟ After considering the well log interpretation and seismic
visualization, channels and layers were defined as having
i =1 ⎝ ⎠
thicknesses between 10 and 80 ft.
Density and interval velocity of shales and sand bodies - The
In order to define it in a simpler mathematical form, (4) is
average density and velocity values were derived from well
proposed instead of (2), which has given similar results.
logs are shown in Table 1.
Instead of the Euclidian distance, it uses the city block
distance, Ci;
4 Reservoir Characterization Using Intelligent Seismic Inversion SPE 98012

Surface Seismic and VSP-Derived Models. Two separate cross-sectional distributions of density, acoustic velocity and
models were obtained by extracting wavelets from actual seismic attributes for both surface seismic and VSP were
surface seismic and VSP for the time window of the Atoka plotted.
and Morrow formations. The synthetic surface seismic The main aim was to determine the most appropriate
response was computed by using a wavelet derived from the portion of the synthetic data set that should be used in training
zone of interest in line 1034 of the 3D surface seismic data, of the network. The success of a neural network model
and the synthetic VSP response (Figure 3) was computed by depends strongly on the way it is trained. Thus, the training
using a Butterworth wavelet derived from line 2064 of the 3D data set must be representative of the geological complexity of
VSP with a larger bandwidth. The properties of the wavelets the area being modeled. Only then, one can obtain successful
used for each case are shown in Table 2. The tuning thickness results in the prediction phase. For this purpose, an effort was
provides a measure of the vertical resolution limit of the made to identify special geological features that can be useful
seismic data, which can be calculated with the relationship; in training the neural network. These include: edges of sand
channels, extreme value points, and unique geological
VR = v/4f (5) structures. In this particular case, those kinds of features could
be found mostly in the central part of the synthetic seismic
where; VR is the vertical resolution in ft/(sec × Hz), v is the section. After several tests, we decided to use data of traces 32
interval velocity in ft/sec and f is the dominant frequency in and 57 (Figure 4). Because of the limitation in optimum
Hz. 1% noise were introduced to the model during the number of data rows, only two traces were used for training.
computations23. The network structure used in training is shown in Figure
In the model, the geological complexity increases with 5-a. As shown, data of traces 32 and 57 were used to predict a
depth. The positive amplitudes (blue in Figure 3) are produced VSP attribute from time and seven available surface seismic
at the top of the carbonates, reflecting the interface between an attributes. At the end, seven separate prediction models have
overlying weak acoustic impadance rock (shales, layers in been developed for seven attributes. After having confidence
white) and an underlying high acoustic impadance of the prediction abilities of these models, they were applied to
(carbonates). Negative amplitude is produced due to the the whole seismic line to obtain network-predicted
interface between the strong acoustic impedance carbonate, distributions of all the available seismic attributes. These
and the underlying weak impedance shale23. attributes were then plotted in order to compare them with the
actual ones.
Model Output. The model is basically a seismic line of 100
traces, which includes three wells at traces 20, 50 and 80 with Step 2: Correlation of VSP with Well Logs: The second step
the well at trace 50 having a VSP survey. These wells had well of the correlation was deriving log properties from the neural
logs of density and acoustic velocity. The available data after network-derived VSP data. Since trace 50 is the trace which
developing the model were; includes all of the data that we are dealing with (i.e. surface
1) Surface seismic and vertical seismic profile responses in seismic, VSP and well logs), this trace was used to develop the
the form of the following seismic attributes: model for this part of the correlation. The density log was
selected as the target log.
− Trace amplitude When one looks at the density log of trace 50, one can
− Average energy clearly see that there are a few different averaged values of
− Trace envelope density (Figure 6), which have been later defined as classes.
− Instantaneous frequency Each of these classes actually represents a different type of
− Hilbert transform rock layer defined in the model. Finally, instead of using
− Paraphase actual values of density, it was decided to use these classes as
the target values. By doing this, we have changed the nature of
− Instantaneous phase
the neural modeling to one of classification. Classification
networks are sometimes simpler than networks that are built to
2) Density and acoustic velocity distributions.
predict continuous values. In the initial trial; classes were
Methodology. The methodology in this study includes two defined as (Figure 6);
major steps of correlation as proposed (Figure 1);
- Class 1: ρ ≈ 1.9 g/cc
1) Correlation of surface seismic with VSP;
2) Correlation of VSP with well logs. - Class 2: ρ ≈ 2.3 g/cc
- Class 3: ρ ≈ 2.65 g/cc
The data used is within the interval 0.8-1.124 seconds
(6600-9000 ft.), as it represents a major exploration target in
the area. Above this interval, there is seismic noise that was Thus, the network structure was like the one in Figure 5-b.
All of the VSP attributes and time were used as inputs to
ignored as input data. In the following section, these two steps
predict one of the density classes.
of correlation will be explained in details.

Step 1: Correlation of Surface Seismic with VSP: An effort Case 2: Field Study (B. Valley Field, New Mexico)
Available data. The following data were used in this study23:
was made to find a correlation between surface seismic
attributes and VSP attributes. First, the model was visualized: 1) A 3D seismic survey, loaned by WesternGeco for this
study, covering an area of 24 mi2 with a vibroseis source,
SPE 98012 Artun, Mohaghegh, Toro, Wilson & Sanchez 5

sweep frequencies ranging from 8 to 98 Hz, and a sweep reliable training model by using all the available data. In this
length of 17 seconds. case, 90% of the total of 871 rows of data from five wells was
2) A vertical seismic profile volume, with a two-LRS-315 used for training, while 10% was used for testing during
vibrators source, covering an area of approximately 3.5 mi2 training. Two types of logs: gamma ray and neutron porosity
in the southeast corner of the area. were selected as target logs. Because of the large number of
3) Logs from around 40 wells were available as either paper available attributes a Key Performance Indicator (KPI) study
or electronic copies. Available types of logs included was carried out, to determine the most influential attributes24.
gamma ray, neutron, density, sonic, spontaneous potential, As a result of this study each attribute had a coefficient
and resistivity. Paper copies have been scanned, and then between 0 and 1, which basically shows the influence of that
digitized together with other available electronic copies. attribute on the selected output (i.e. logs). The attributes
Average total depth of those wells were ranging from 8000 (including the time) were ranked based on their KPI
to 9000 ft, intersecting the Atoka and Morrow formations. coefficients and the ones having a coefficient larger than 0.5
were used as inputs. Those rankings for gamma ray and
Unfortunately, not all of the available well logs were of neutron porosity are shown in Table 3. The best thirteen
high visual quality. That made them hard to accurately attributes for the gamma ray, and the best eight attributes for
digitize. It is a proven fact that quality of the data has an the neutron porosity had coefficients higher than 0.5, and they
important role in building reliable neural network models. were used as the input attributes. Training results were
Reasonable amount of noise in the data is acceptable, and even analyzed for each well separately, and upon applying those
useful for a more realistic model. However, it is impossible to models to the other traces on the seismic line, distributions of
rely on a model, which has been built with unreliable data. gamma ray logs and neutron porosity logs were produced.
Rolon’s study4 clearly states this fact, by comparing two
similar studies with well log data having different levels of Results and Discussions
quality. Given these issues, the digitized log files were Case 1. Table 4 shows the correlation statistics of the training
evaluated and selected for neural network design based on models developed for this step of the study. Correlation
overall quality, to avoid building poor prediction models. Only coefficient, r, and r-squared values are included as accuracy
good quality logs were used. indicators of the match. Results for pattern (overall), training,
Within the context of log quality and distribution, and the calibration, and verification sets are included. Figure 8 shows
location of the VSP well, a seismic line (Figure 7) was the results of the training as actual vs. network-predicted plots
extracted from the 3D survey area for use in the neural for training, calibration, and verification sets of seven
network design and evaluation. The line trends SE-NW attributes. Each set of data is shown with a different symbol
through the survey area, and passes through five wells and color. Statistical and visual results indicate that reliable
including the VSP well. The line included 173 seismic traces; models have been developed. These models have been applied
the VSP well, well-1, located on trace 16 (Figure 7). Other to the whole line (i.e. other traces of the synthetic model). The
wells are located on traces 55, 90, 123, and 153, respectively. seismic attribute distributions have been re-produced. They
The data shown in Figure 7 are seismic amplitudes. have been re-plotted to be able to compare with the actual
VSP data were available only for well-1. Surface seismic distributions. Figure 9-a shows the comparison plots for three
data used in the analysis extended from 0.92-1.1 seconds (two- attributes: amplitude, average energy, and envelope; and
way time). This interval includes reflection events associated Figure 9-b shows the plots for three other attributes:
with the Atoka-Morrow target interval A total of 27 seismic frequency, Hilbert transform, and paraphase, which are
attributes were available. Eleven of these attributes were obtained after applying the models for these attributes to the
provided by the Kingdom Suite, and other 16 attributes were entire line. These plots also include the actual log lines of each
calculated using theoretical relationships. Additional attributes attribute at traces 20, 35, 50, 65, and, 80, which makes it
included a variety of derivatives, instantaneous energy, power easier to assess the quality of the produced distributions.
and acceleration, quality factor, acoustic impedance For the second step, where the density has been predicted;
inversions, and various residuals and smoothed outputs of actual vs. network plot of the training model is shown in
other attributes. Figure 10-a. Although the network seemed to be successful in
most parts, it has missed some points in the middle-valued
Methodology. The methodology as the one employed in the region which was assumed to belong to Class 2: 2.3 g/cc. In
synthetic case study was followed in this case. Since there was order to overcome this problem, it has been decided to
only one available VSP survey (Well-1), the required introduce another class to our model. Class 4 was assigned to
correlation models for VSP prediction were restricted to data density values around 2.09 g/cc (Figure 10-b). That has helped
from that well. The seismic and attribute data were resampled the network to be successful in that region (Figure 10-b), and
to 0.0005 seconds (half-millisecond). Well log responses in this network model was chosen to be used in further studies.
depth were converted to time using modified well and surface The correlation statistics of this model are shown in Table 5
seismic velocity functions. The log responses were also together with the results for the velocity log. Actual VSP
resampled to half-millisecond intervals. attributes were applied to this model, and density has been
In the log prediction stage, the approach that has been predicted along the seismic line with a correlation coefficient
carried out was to use all the available well data for training, of 0.82. This has been repeated with the predicted VSP
and to apply the model to other parts of the seismic line to attributes coming from the first correlation step. The density
obtain the distributions. The main idea is to have a more has again been predicted through the seismic line with a
6 Reservoir Characterization Using Intelligent Seismic Inversion SPE 98012

correlation coefficient value of 0.81. Figure 11 shows the final 2) Benefits of using a synthetic seismic model were clearly
actual vs. network-predicted density distributions. It is seen seen in developing the appropriate methodology by identifying
that, the lithological units as well as the sand channels have the most appropriate neural network algorithm and testing to
been successfully predicted by the neural network. With this use other tools. The synthetic model not only made it easy to
final result, the goal of predicting density log through the two- be familiar with the type of data that is dealt with, but also
step intelligent seismic inversion methodology has been provided the opportunity of comparing results for the whole
accomplished. The same procedure has also been applied to seismic section. Although it was not applicable for the field
velocity and acoustic impedance. study, the synthetic model was used for a classification
(lithology identification) approach.
Case 2. As in the synthetic case, surface seismic – VSP
correlation models have been developed, by using time and 3) Although there are several examples of using artificial
eleven surface seismic attributes to predict the single VSP neural networks to correlate seismic attributes to reservoir
attribute. The attributes used in this study were the ones which properties, or well logs, this study is first of its kind simply
have been generated from the Kingdom Suite as the other because of integrating three types and scales of data (surface
attributes can be derived by using those ones. Table 6 shows seismic, VSP, and well logs).
the correlation statistics for these models for each attribute.
Results for pattern, training, calibration, and verification sets 4) To determine the data that should be used for training,
are included. Verification results for phase, paraphase, model visualization can be helpful. It was seen in the first
frequency, and signed frequency seemed to be low. However, correlation step of the synthetic model study that; including
when plotted it seemed that they were not as bad as their diverse geological characteristics in the training set can
correlation coefficients would suggest. Their actual vs. increase the prediction abilities of the neural network models.
network graphs are shown in Figure 12. Actual vs. network
plots for other attributes are shown in Figure 13. These models Acknowledgements
were then applied to other traces on the seismic line, with the This research was supported by a grant from the U.S.
available surface seismic data. VSP distributions were Department of Energy (DE-FC26-03NT41629). Authors
produced through the entire seismic line. would like to express their appreciation and gratitude to the
In the log prediction stage, gamma ray and neutron porosity project manager Mr. Thomas Mroz, for his help throughout
were selected as the target logs. The results for these two types the life of this project. Seismic and VSP data of the Buffalo
of logs are presented below separately. Valley Field, used in this study were provided courtesy of
WesternGeco. Authors also would like to acknowledge
Gamma Ray: Correlation statistics are shown in Table 7 Janaina Pereira, for her help in digitizing well logs for the
for pattern, training, and testing sets, respectively. Results are field study.
showing both overall statistics, and separate well statistics.
Graphs that show actual and network-predicted log lines are Nomenclature
shown in Figure 14 for all wells. Figure 15 shows the C = City block distance V = Interval velocity
predicted gamma ray distribution of the seismic line using D = Euclidian distance VR = Vertical resolution
these models. Actual log lines of five wells are also shown in f = Dominant frequency W = Weight of a link
the plot which may give a better perspective of comparisons. I = Input of a neuron Y = Dependent variable
i, j = Random numbers X = Independent variable
Neutron Porosity: Neutron porosity logs were available in n = No. of sample observations σ = Smoothing factor
only two wells; well-1 and well-2. Thus, these two wells were
used in training. Correlation statistics of these models are References
shown in Table 8. Figure 16 shows the actual and network- 1. Nikravesh. M. and Aminzadeh, F.: ‘Past, present and future
predicted logs for two wells. Predicted distribution of neutron intelligent reservoir characterization trends’, Journal of Petroleum
porosity is shown in Figure 17. Science and Engineering, Vol. 31, pp.67-79, 2001.
2. Mohaghegh, S.D., Richardson, M., Ameri, S.: ‘Virtual Magnetic
Imaging Logs: Generation of Synthetic MRI Logs from
Conclusions Conventional Well Logs’, paper SPE 51075, 1998 SPE Eastern
In this study, a novel approach to the seismic inversion Regional Meeting Proceedings, Nov. 9-11, Pittsburgh,
problem is introduced. A two-scale-step, intelligent seismic Pennsylvania.
inversion methodology is successfully implemented on a 3. Mohaghegh, S.D., Goddard, C., Popa, A., Ameri, S., and Bhuiyan,
synthetic model, and real data of the Buffalo Valley Field, M.: ‘Reservoir Characterization through Synthetic Logs’, paper
New Mexico. The main conclusions of this study can be SPE 65675, 2000 SPE Eastern Regional Meeting Proceedings, Oct.
summarized as follows: 17-19, Morgantown, West Virginia.
1) Even being in a complex and non-linear manner, it was 4. Rolon, L.: ‘Developing Intelligent Synthetic Logs: Application to
shown that a relationship does exist between various seismic Upper Devonian Units in PA’, M.Sc. thesis, West Virginia
University, Morgantown, West Virginia, 2004.
attributes and reservoir properties. Such a relationship can be 5. Mohaghegh, S.D., Arefi, R., Ameri, S., and Rose, D.: ‘Design and
extracted by using artificial neural networks, one of the major Development of an Artificial Neural Network for Estimation of
soft computing tools. Generalized regression neural network Formation Permeability’, SPE 28237, 1994 SPE Petroleum
(GRNN), which is a useful algorithm for approximating Computer Conference Proceedings, Jul. 31 - Aug. 3, Dallas, Texas.
continuous functions, was applied successfully in this study.
 SPE 98012 Artun, Mohaghegh, Toro, Wilson & Sanchez 7

6. Arpat, G.B., Gumrah, F., Yeten, B.: ‘The neighborhood approach Table 1: Density and interval velocity values for different types
to prediction of permeability from wireline logs and limited core layers.
plug analysis data using backpropagation artificial neural
networks’, Journal of Petroleum Science and Engineering, Vol. Lithology Density, g/cc Velocity, ft/s
20, pp.1-8, 1998. Sandstone 2.3-2.6 12,500-16,000
Shale 1.9-2.1 9,000-11,000
7. Weiss, W.W., Balch, R.S., Stubbs, B.S.: ‘How Artificial
Limestone 2.5-2.71 14,000-17,000
Intelligence Methods Can Forecast Oil Production’, paper SPE
75143, 2002 SPE/DOE Improved Oil Recovery Symposium Table 2: Properties of wavelets used to compute seismic
Proceedings, April 13-17, Tulsa, Oklahoma. responses.
8. Hampson, D.P., Schuelke, J.S., and J.A. Quirein: ‘Use of
multiattribute transforms to predict log properties from seismic Property 3D Seismic VSP
data’, Geophysics, Vol. 66, No. 1, pp. 220-236, 2001. Bandwidth, Hz 11-95 20-110
9. Leiphart, D.J., Hart, B.S.: ‘Comparison of linear regression and a Dominant Frequency, Hz 45 65
probabilistic neural network to predict porosity from 3D seismic Dominant Period, sec 0.022 0.0154
attributes in Lower Brushy Canyon channeled sandstones, Interval Velocity, ft/sec 15,000-16,000 15,000-16,000
southeast New Mexico’, Geophysics, Vol. 66, No. 5, pp. 1349- Tuning Thickness, ft 83-88 57-61
1358, 2001
Table 3: Results of the Key Permormance Indicators for gamma
10. Balch, R.S., Stubbs, B.S., Weiss, W.W., Wo, S.: ‘Using Artificial
ray and neutron porosity. Shaded ones were used as inputs.
Intelligence to Correlate Multiple Seismic Attributes to Reservoir
Properties’, paper SPE 56733, 1999 SPE Annual Technical Rank Gamma Ray Neutron Porosity
Conference and Exhibition Proceedings, Oct. 3-6, Houston, Texas. 1 Peak-to-Trough Amp Weighted Phase
11. Chawathe, A., Ouenes, A., Weiss, W.W.: ‘Interwell property 2 Smoothed Envelope Time
mapping using crosswell seismic attributes’, paper SPE 38747, 3 Finite Difference Peak-to-Trough
1997 SPE Annual Technical Conference and Exhibition 4 Average Frequency Instant Energy
Proceedings, Oct. 5-8, San Antonio, Texas. 5 Instant Energy Smoothed Envelope
12. Soto B. R., and Holditch, S.A.: ‘Development of reservoir 6 Envelope Envelope
characterization models using core, well log, and 3D seismic data 7 Time Amplitude
and intelligent software’, paper SPE 57457, 1999 SPE Eastern 8 Amp Weighted Phase HF Inversion
Regional Conference and Exhibition Proceedings, Oct. 21-22, 9 Amplitude Finite Difference
Charleston, West Virginia. 10 Phase Average Energy
13. Reeves, S., Mohaghegh, S.D., Fairborn, J., and Luca, G.: 11 Hilbert Average Frequency
‘Feasibility Assessment of a New Approach for Integrating Multi- 12 Smoothed Inversion Paraphase
13 HF Inversion Hilbert
Scale Data for High-Resolution Reservoir Characterization’, paper st
14 1 D Amp Smoothed Inversion
SPE 77759, 2002 SPE Annual Technical Conference and 15 Average Energy Decay Rate
Exhibition Proceedings, Sep. 29 - Oct. 2, San Antonio, Texas. 16 Inst. Accel Phase
14. Gadallah, M.R.: Reservoir Seismology: Geophysics in Non- 17 Inst. Q Fac Signed Freq
technical Language, PennWell Books, Tulsa, Oklahoma, 1994. 18 Decay Rate Frequency
15. Taner, M.T.: ‘Seismic Attributes’, CSEG Recorder, pp. 48-56, 19
nd
2 D Amp
st
1 D Amp
September, 2001. 20 Res Env Inst Res Env
16. Kingdom Suite software tutorial, Seismic Micro-Technology, Inc., 21 Frequency Inst Abs Amp
Houston, Texas. 22 Inst Abs Amp Inst. Q Fac
st
17. Fausett, L.: Fundamentals of Neural Networks: Architectures, 23 Inst Res Env 1 D Env
st
Algorithms, and Applications, Prentice-Hall, Englewood Cliffs, 24 1 D Env Res Env
New Jersey, 1994. 25 Inst Power Inst. Accel
nd
18. Mohaghegh, S.D.: ‘Virtual Intelligence Applications in Petroleum 26 Signed Freq 2 D Amp
nd
Engineering - Part 1: Artificial Neural Networks.’ Journal of 27 Paraphase 2 D Env
nd
Petroleum Technology, Distinguished Author Series September 28 2 D Env Inst Power
2000, pp 64-73.
19. Huang, Z., Williamson, M.A.: ‘Geological pattern recognition and Table 4: Correlation statistics for seven VSP attributes, for the
modeling with a General Regression Neural Network’, Canadian first correlation step of the synthetic model study.
Journal of Exploration Geophysics, Vol. 30, No. 1, pp. 60-66,
1994. Amplitude A. Energy Envelope Frequency
20. Luca, G.: ‘Towards High Resolution Reservoir Characterization’, r
2
r r2 r r2 r r2 r
M.Sc. Thesis, West Virginia University, Morgantown, West Pattern 1.00 1.00 0.83 0.91 0.97 0.98 0.76 0.87
Virginia, 2001. Training 1.00 1.00 0.84 0.92 0.98 0.99 0.76 0.87
21. Specht, D.: ‘A General Regression Neural Network’, Vol. 2, No. 6, Calibration 1.00 1.00 0.79 0.89 0.90 0.95 0.81 0.90
IEEE Transactions on Neural Networks, November, 1991. Verification 1.00 1.00 0.75 0.86 0.92 0.95 0.73 0.86
22. NeuroShell 2, release 4.0 tutorial, Ward Systems Group, Inc.,
Frederick, Maryland, 1993-1998.
23. Sanchez, A.A.: ‘3D seismic interpretation and synthetic modeling Hilbert Trans. Paraphase Phase
2 2 2
of the Atoka and Morrow formations, in the Buffalo Valley Field r r r r r r
Pattern 0.92 0.96 0.88 0.94 0.93 0.96
(Delaware Basin, New Mexico, Chaves County) for reservoir
characterization using neural networks’, M.Sc. thesis, West Training 0.93 0.96 0.89 0.94 0.97 0.99
Virginia University, Morgantown, West Virginia, 2004. Calibration 0.90 0.95 0.85 0.92 1.00 1.00
24. KPI module of Intelligent Reservoir Characterization & Analysis, Verification 0.96 0.98 0.85 0.92 0.52 0.72
IRCA. http://www.intelligentsolutionsinc.com/irca.htm
 8 Reservoir Characterization Using Intelligent Seismic Inversion SPE 98012

Table 5: Correlation statistics for density and velocity, for the

second correlation step of the synthetic model study.

Density Aco. Velocity

r2 r r2 r
Pattern 0.82 0.91 0.90 0.95
Training 0.94 0.97 0.93 0.96
Calibration 0.54 0.73 0.97 0.99
Verification 0.54 0.73 0.49 0.70

Table 6: Correlation statistics for eleven VSP attributes, for the

first correlation step of the real case study.

Amplitude A. Energy Envelope Frequency

r2 r r2 r r2 r r2 r
Figure 1: Modeling high-frequency logs from low-frequency
Pattern 0.97 0.98 0.96 0.98 0.98 0.99 0.34 0.58
seismic signals. Schematic view of seismic inversion process
Training 1.00 1.00 0.98 0.99 0.98 0.99 0.92 0.96 with the proposed correlation map. Two major steps of
Calibration 0.91 0.95 0.98 0.99 0.98 0.99 0.83 0.91 correlation: 1)Surface Seismic – VSP, 2) VSP – Well logs. (Top
2
Verification 0.90 0.95 0.85 0.92 0.85 0.92 0.21 0.46 portion: taken from Hampson et al. )

Hilbert Trans. Paraphase Phase

r2 r r2 r r2 r
Pattern 0.98 0.99 0.79 0.89 0.64 0.80
Training 0.98 0.99 0.98 0.89 0.96 0.98
Calibration 0.98 0.99 0.90 0.95 0.74 0.86
Verification 0.92 0.96 0.29 0.54 0.41 0.64

Finite Peak-to- Signed Inversion

Difference Trough Frequency
r2 r r2 r r2 r r2 r
Pattern 0.92 0.96 0.93 0.96 0.46 0.96 0.98 0.99
Training 0.98 0.99 0.98 0.99 0.98 0.99 0.98 0.99
Calibration 0.98 0.99 0.90 0.95 0.86 0.93 0.98 0.99 Figure 2: Schematic diagram of an artificial neuron or a
18
Verification 0.83 0.91 0.79 0.89 0.09 0.30 0.98 0.99 processing element. (Taken from Mohaghegh )

Table 7: Correlation statistics for the pattern, training, and

calibration sets of the training model for gamma ray log
prediction.

Pattern Training Calibration

r2 r r2 r r2 r
All 0.77 0.88 0.79 0.89 0.60 0.77
Well-1 0.58 0.76 0.57 0.75 0.73 0.85
Well-2 0.75 0.86 0.76 0.87 0.67 0.82
Well-3 0.66 0.81 0.72 0.85 0.27 0.52
Well-4 0.76 0.90 0.77 0.88 0.27 0.52
Well-5 0.81 0.90 0.82 0.91 0.69 0.83

Table 8: Correlation statistics for the pattern, training, and

calibration sets of the training model for neutron porosity log
prediction.

Pattern Training Calibration

r2 r r2 r r2 r
All 0.95 0.97 0.96 0.98 0.82 0.91
Well-1 0.95 0.98 0.97 0.98 0.77 0.88
Figure 3: VSP amplitude response of the synthetic model.
Well-2 0.94 0.97 0.95 0.97 0.89 0.94
SPE 98012 Artun, Mohaghegh, Toro, Wilson & Sanchez 9

Figure 4: Cross-sectional envelope distribution, which was used

to determine special features that can be useful in the network
training. Finally, traces 32 and 57 were decided upon to be used
in training.

a)
Figure 6: Density log of Trace - 50. First classification proposed:
Three classes at density values 1.9, 2.3, and 2.65

b)

Figure 5: Network structures used for training. a) Data of traces

32 and 57 have been used, to predict a VSP attribute from time
and surface seismic attributes. b) Data of Trace - 50 have been
used, to predict one of the density classes from time and VSP Figure 7: Amplitude distribution of the seismic line. Red dashed
attributes. lines are the five wells that the line is passing through. Well-1 is
the well with the VSP survey.
10 Reservoir Characterization Using Intelligent Seismic Inversion SPE 98012

Figure 8: Actual vs. network results for each attribute after training the network for surface seismic - VSP correlation. Results for training,
calibration, and verification sets are included with different symbols.
SPE 98012 Artun, Mohaghegh, Toro, Wilson & Sanchez 11

Figure 9-a: Actual (left) and network-predicted (right) distributions for VSP attributes: amplitude, average energy, and envelope.
12 Reservoir Characterization Using Intelligent Seismic Inversion SPE 98012

Figure 9-b: Actual (left) and network-predicted (right) distributions for VSP attributes: frequency, Hilbert transform, and paraphase.
SPE 98012 Artun, Mohaghegh, Toro, Wilson & Sanchez 13

a)

b)

Figure 10: Defining density classes, and the corresponding training results. a) Three classes (1.9, 2.3, 2.65). r = 0.82 for training set,
b) Four classes (1.9, 2.09, 2.3, 2.65). r = 0.94 for training set.
14 Reservoir Characterization Using Intelligent Seismic Inversion SPE 98012

Figure 11: Actual (left), and network-predicted (right) distributions of density values (r = 0.81).

Figure 12: Actual vs. network plots for paraphase, phase, frequency, and signed frequency shown as logs. Although having relatively low
correlation coefficient values, visual plots show promise.
SPE 98012 Artun, Mohaghegh, Toro, Wilson & Sanchez 15

Figure 13: Actual vs. network plots for seven of the attributes, that had satisfactory correlation statistics. Training, calibration, and
verification sets are included with different symbols and colors.
16 Reservoir Characterization Using Intelligent Seismic Inversion SPE 98012

Figure 14: Actual and network-predicted gamma ray logs for wells 1, 2, 3, 4, and 5. Values for training, and calibration sets are included. VSP
attributes were used as inputs.
SPE 98012 Artun, Mohaghegh, Toro, Wilson & Sanchez 17

Figure 15: Network-predicted gamma ray distribution through the seismic line of interest. Actual log lines are also shown for five wells on the
line. VSP attributes were used as inputs.

Figure 16: Actual and network-predicted neutron porosity logs for wells 1, and 2. Values for training, and calibration sets are included. VSP
attributes were used as inputs.
18 Reservoir Characterization Using Intelligent Seismic Inversion SPE 98012

Figure 17: Network-predicted gamma ray distribution through the seismic line of interest. Actual log lines are also shown for two wells on the
line. VSP attributes were used as inputs.