Intelligent Reservoir Characterization (IRESC)

odecrackΤεχνίτη Νοημοσύνη και Ρομποτική

29 Οκτ 2013 (πριν από 3 χρόνια και 10 μήνες)

72 εμφανίσεις


In this study, a new integrated methodology
will be

developed to identify the nonlinear relationship
and mapping between 3
D seismic data, production log
and will be applied to producing field. The method uses
conventional techniques such as
geostatistical and
classical pattern recognition [1]

in conjuncti
on with
modern techniques such as soft computing
computing, fuzzy logic, genetic computing, and
probabilistic reasoning) [2
. An important task of our
research is to use clustering techniques recognize the
optimal location of a new well to be dril
led based on 3
seismic data and available production log/data or other
viable logs. The classification task will be accomplished
in three ways; 1) k
means clustering, 2) fuzzy clustering,
and 3) neural network clustering to recognize the
similarity cube
s. Then the relationship between each
cluster and production log will be recognized around the
wellbore and the results will be used to reconstruct and
extrapolate the production log away from the wellbore.
This advanced 3
D seismic and log analysis and
terpretation can be used to predict; 1)
mapping between
production data and seismic data, 2) reservoir
connectivity based on multi
attributes analysis, 3) pay
zone estimation, and 4) optimum well placement.



In reservoir engineering, it is i
mportant to characterize
how 3
D seismic information (attributes) is related to
production, lithology and geology as well as well logs
(e.g. porosity, density, gamma ray, etc.) [5
Knowing the 3
D seismic data will help to reconstruct
the 3
D volume o
f relevant reservoir information away
from the wellbore. However, data from well logs and
D seismic attributes are often difficult to analyze
because of their complexity and our limited ability to
understand and use the information content of these

Unfortunately, only linear and simple nonlinear
information can be extracted from these data by
standard statistical methods.

As an alternative, neural networks [11] and fuzzy logic
[12] have the potential to establish a model from
nonlinear, complex, a
nd multi
dimensional data and
find wide application in analyzing experimental,
industrial, and field data. In recent years, this has
stimulated a growing interest among reservoir
engineers, geologists, and geophysicists [2, 4
Boadu [5] and Nikravesh e
t al. [2] applied artificial
neural networks and neuro
fuzzy successfully to find
the relationship between seismic and rock properties
for sandstones rocks. Monson and Pita [10], Chawathe,
et al. [7] and Nikravesh [4] applied artificial neural
networks and

fuzzy techniques successfully to
find the relationship between 3
D seismic attributes
and well logs and to extrapolate mapping away from
the wellbore to reconstruct the log responses.

In this paper, we analyze 3
D seismic data to recognize
the mos
t important patterns, structures, relationships,
and characteristics based on classical pattern
recognition techniques, neural networks and fuzzy
logic models. The nonlinear mapping between
production data and 3
D seismic will be identified.
Finally, based

on integrated clustering techniques, the
optimal locations to drill new wells will be predicted.



Cluster analysis encompasses a number of different
classification algorithms, which can be used to
organize observed data into meaningful

structures. For
example, k
means is an algorithm to assign k centers to
represent the clustering of N points (k<N). These
points are iteratively adjusted so that each point is
assigned to one of the clusters, and each of the clusters
is the mean of its a
ssigned points. In general, the k
means technique will produce exactly k different
clusters of the greatest possible distinction.
Alternatively, fuzzy techniques can be used as a mean
for clustering. Fuzzy clustering partitions a data set into
fuzzy cluste
rs such that each data point can belong to
multiple clusters. Fuzzy c
means (FCM) is a well
known fuzzy clustering technique that generalizes the
classical (hard) c
means algorithm. In addition, the
organizing map technique known as the
Kohonen’s sel
organizing feature map can be used as
an alternative for clustering purposes [13]. This
technique converts patterns of arbitrary dimensionality
(the pattern space) into the response of one

or two
dimensional arrays of neurons (the feature space). This
nsupervised learning model can discover any
Intelligent Reservoir Characterization (IRESC)

Masoud Nikravesh, Mahnaz Hassibi, BISC Program, Computer Sciences Division, EECS Department
University of California, Berkeley, CA 94720, USA

relationship of interest such as patterns, features,
correlations, or regularities in the input data, and
translate the discovered relationship into outputs.

Figure 1 shows schematically the flow of information

and techniques to be used for intelligent reservoir
characterization (IRESC). The main goal will be to
integrate soft data such as geological data with hard
data such as 3
D seismic, production data, etc. to build
a reservoir and stratigraphic model. In
this study, we
will only concentrate on integrating 3
D seismic data
and production data to build similarity cubes based on
clustering techniques. In this case study, using three
different techniques [13
15]; 1) k
means, 2) neural
network (self
map), and 3) fuzzy c
we will use 3
D seismic attributes to find similarity
cubes and clusters (clusters can be interpreted as
lithofacies, homogeneous classes or similar patterns
that exist in the data). Then the relationship between
each cluster an
d production log will be recognized
around the wellbore and the results will be used to
reconstruct and extrapolate the production log away
from the wellbore. The results from clustering will be
superimposed on the reconstructed production log and
locations to drill new wells will be determined.



Our example is from a field that produces from the
Ellenburger Dolomite. The Ellenburger is one of the
most prolific gas producers in the conterminous United
States, with greater than 13 TCF of

production from
fields in west Texas. The Ellenburger Dolomite was
deposited on an Early Ordovician passive margin in
shallow subtidal to intertidal environments. Reservoir
description indicates the study area is affected by a
related, collapsed p
aleocave system that acts as
the primary reservoir in the field studied [16, 17].

The 3
D seismic volume used for this study has
3,178,500 data points. Two hundred, fifty
two data
points intersect the seismic traces. Finally, 89
production log data points

are available for analysis (19
production and 70 non
production). In the training
phase for clustering and mapping purposes, a sub set of
seismic data, which is representative of the 3
D seismic
cube, production log data and area of interest, was
. The subset was designed as a section passing
through all the wells as shown in Figure 2. Section has
105,150 data points. However, 35,751 data points are
selected for clustering purposes, representing the main
focus area. Figure 3 shows the schematic di
agram of
how the well path intersects the seismic traces. For
clustering and mapping, there are two windows that
must be optimized, 1) the seismic window and 2) the
well log window. There are over one hundred seismic
attributes that exist. Therefore, an op
timal number of
seismic attributes needs to be recognized. An optimal
number of clusters to be recognized must to be chosen
depending on the nature of the problem. Figure 4
shows the iterative technique that has been used to
select an optimal number of clu
sters, seismic attributes,
and optimal processing windows for the seismic
section shown in Figure 3. Knowledge of experts such
as geological layering has also been used to constrain
the maximum number of clusters to be selected. In this
study, six attribu
tes have been selected (Raw Seismic,
Cosine Instantaneous Phase, Instantaneous Amplitude,
Instantaneous Phase, Instantaneous Frequency, and
Integrate Absolute Amplitude) out of 19 attributes
calculated. Ten clusters were recognized and a window
of 1 has be
en used as the optimal window size for the
seismic and a window of 3 for the production log data.
Finally, based on qualitative analysis, specific clusters
that have the potential to be in the producing zones are
selected (Figure 5). Figure 5 shows the cl
uster (group
of clusters) selected using three different techniques, k
means (statistical), neural network, fuzzy c
(fuzzy logic) clustering. By comparing Figures 5.a,
Figure 5.b, and Figure 5.c one can conclude that (in
this study) that all the tech
niques predicted the same
cluster (or group of clusters) based on our objectives
(producing zones). However, this may not always be
the case. In addition, the information that can be
extracted based on different techniques will be
different. For example,
clusters using classical
techniques will have sharp boundaries whereas those
generated using the fuzzy technique will have fuzzy

Based on the clusters recognized in Figure 5 and the
production log, a subset of the cluster has been selected

shown in Figure 6. In this sub
cluster, the
relationship between production log and clusters has
been recognized and the production log has been
reconstructed and extrapolated away from the wellbore.
Finally, the production log and the cluster data are
perimposed at each point in the 3
D seismic cube.
Figure 7 shows a typical time
slice of a 3
D seismic
cube that has been reconstructed with the extrapolated
production log and cluster data.. Three criteria have
been used to select potential locations for
infill drilling
or recompletion, 1) continuity of the selected cluster
and production log, 2) size and shape of the cluster and
production log, 3) existence of the production log

inside the selected cluster. Based on these criteria
locations of the new we
lls are selected and two such
locations are shown in Figure 7, one with high
continuity and potential for high production and one
with low continuity and potential for low production.
The neighboring wells that are already in production
confirm such a pred
iction as shown in Figure 7.



In this study, a new integrated methodology was
developed to identify a nonlinear relationship and
mapping between 3
D seismic data, production log.
The technique was applied toa producing field.
advanced 3
D seismic and log analysis and
interpretation methodology uses conventional
statistical techniques in addition to modern soft
computing techniques. It can be used to predict; 1)
mapping between production log and seismic data, 2)
reservoir connectivity bas
ed on multi
analysis, 3) pay zone estimation, and 4) optimum well
It is important to note that these optimal
locations are based on static data (3
D data and original
well logs) and can be considered only as potential
locations for ne
w wells to be drilled. Therefore, final
decisions for selecting location of new drilling sites
should be confirmed by dynamic reservoir modeling
and simulation and by future performance prediction of
the reservoir through reservoir simulation and history
atching techniques. Finally, this new methodology
combined with techniques presented by Nikravesh [4,
6] and Nikravesh et al. [2, 3] can be used to reconstruct
the well logs such as porosity, density, resistivity, etc.
away from the wellbore. Therefore, n
et pay zone
thickness, reservoir models, and geological
representations will be accurately identified as the main
goal of the IRESC project. This
accurate reservoir
characterization through data integration is a key step
in reservoir modeling, management,
and production



Aminzadeh, F. and S. Chatterjee, Applications of clustering in
exploration seismology, Geoexploration, 23 (1984/85) 147


Nikravesh, M., B. Novak, and F. Aminzadeh, Data mining and
fusion with integrated neuro
fuzzy agents: Rock properties and
seismic attenuation, JCIS'98, The Fourth Joint Conference on
Information Sciences, North Carolina, USA, October 23


Nikravesh, M, and Aminzadeh, F., Knowledge discovery from
data bases: Intelligent data mining

ACT Inc. and LBNL
Proposal, submitted to SBIR
NASA, 1997.


Nikravesh, M. “Mining and fusion of petroleum data with fuzzy
logic and neural network agents”,
CopyRight© Report
DOE, and DeepLook Industry Consortium, 1998.


Boadu, F.K., Rock prope
rties and seismic attenuation; Neural
network analysis, Pure Applied Geophysics, 149 (1997), pp.


Nikravesh, M., Neural network knowledge
based modeling of
rock properties based on well log databases, SPE 46206, 1998
SPE Western Regional Meeting,
Bakersfield, California, 10
May, 1998.


Chawathe, A., A. Quenes, and W.W. Weiss, Interwell property
mapping using crosswell seismic attributes, SPE 38747, SPE
Annual Technical Conference and Exhibition, San Antonio,
TX, 5
8 Oct. 1997


Yoshioka, K, N. Shi
mada, and Y. Ishii, Application of neural
networks and cokriging for predicting reservoir porosity
thickness, GeoArabia, Vol 1, No 3, 1996.


Schuelke, J.S., J.A. Quirein, J.F. Sarg,, D.A. Altany, and P.E.
Hunt, Reservoir architecture and porosity distrib
Pegasus field, West Texas
An integrated sequence
seismic attribute study using neural networks,
SEG’97 meeting , Dallas, TX 1997.


Monson G.D., and J.A. Pita, Neural network prediction of
logs for net pay and reservoir property

Greater Zafiro field area, Equatorial Guinea, SEG’97 meeting ,
Dallas, TX 1997.


Nielsen, R., Theory of backpropagation neural networks,
presented at IEEE Proc., Int. Conf. Neural Network,
Washington DC, 1989.


Zadeh, L.A., Fuzzy sets
, Information and Control, 8, 1965, 338


Kohonen, T. Self
Organizing Maps, Second Edition, Springer,


Lin, C.T., C.S.G. Lee, Neural fuzzy systems, Prentice Hall,


The Math Works, Natick, 1995.


Adams, R.D., J.W. Collister, D.D. Ekart, R.A. Le
vey, and M.
Nikravesh, Evaluation of gas reservoirs in collapsed paleocave
systems, Ellenburger group, Perminan basin, Texas, AAPPG
Annual Meeting, San Antonio, Tx, 11
14 April, 1999.


Levey, R., M. Nikravesh, R. Adams, D. Ekart, Y. Livant, S.
Snelgrove, J.

Cllister, Evaluation of fractured and paleocave
carbonate reservoirs, AAPPG Annual Meeting, San Antonio,
Tx, 11
14 April, 1999.


Figure 2.

Figure 4.

Figure 1.

Figure 4.

Figure 2.

Figure 3.

Figure 5.

Figure 5.a

Figure 5.b

Figure 5.c

Figure 6.

Figure 7.