Well Log Estimates and Condence Intervals B Yusing Articial Neural Networks

Well log estimates and condence intervals b yusing articial neural networks
R afael E. Banchs* & Reinaldo J. Michelena, PDVSA-Intevep
Summary
Estimation of petrophysical properties of rocks from seismic attributes plays a very important role in dening
reservoir models for reservoir characterization and simulation. How ever,some degree of uncertainty is alw ays
related to such propert yestimates. This w orkpresents
a nonlinear analysis technique, based on articial neural
netw orks,for dealing with the problem of seismic-guide
w ell log estimation and its related uncertainties. Field
data examples demonstrate the potential of the method.
Introduction
Indirect estimation of petrophysical properties from other

available sources of information, such as well logs and
seismic data, pla ys a very important role in reservoir
characterization and sim ulation. Recently, neural netw ork based methods for estimating rock properties from
seismic attributes ha ve been gaining some popularity
(Shuki et all, 1994; Todorov et all, 1997 and 1998). The
main adv an tageof neural netw orksover other estimation methods is their abilit y for extracting nonlinear
relationships among data sets.
Ho wever, these neural-net estimation methods do not give
any information about the condence intervals associated
with the resulting estimates. Since it is always desirable
to estimate also condence intervals for a giv en property
estimate, stochastic methods are generally used in order
to quantify the uncertainties associated to the obtained
estimates (Deutsch and Journel, 1992).
In the presen t w ork, w e show a neural netw ork-based
method for estimating w ell logs, as well as the associated condence intervals, using 3D seismic attributes as
a guide. First, the methodology and some practical issues
related to its implementation are brie y discussed. Then,
a eld data example is shown, in which a volume of spontaneous potential and the corresponding uncertainties in
the estimation process are estimated from the seismic attributes of a 3D volume.
Data adjustment and re-sampling

The rst problem encountered when implementing the
estimation technique is that the input data set
(seismic
attributes) and the output data set (well logs) are
measured in tw o dierent domains, time and depth,
respectively. The availabilit yof some velocity information, whic hallows the con version of the w ell log data
to time, is required. When having the appropriate T-Z
curv es, the depth to time conversion of the data is easily
achiev ed b y using spline interpolators.
SEG 2000 Expanded Abstracts
A second problem is the dierence in resolution betw een

the seismic data and the well log data. Due to this problem, a resolution adjustment is required. In order to avoid
aecting convergence during the training stage of the neural net w ork, it is recommended to low-pass lter the well
log data according to the spectral content of the seismic
data. As a result of this low-pass ltering the proposed
technique is limited to estimate the desired information
at the resolution of the seismic waveeld.
Another problem results from the dierence betw een the
sampling rates of seismic data and well log data. Then,
well log data has to be do wn-sampled (or seismic data
up-sampled, or both) in order to obtain a set of inputoutput sample pairs for training the neural netw ork. The
up-sampling and down-sampling of the data sets can be
performed directly in the discrete domain by using conventional re-sampling schemes (Oppenheim and Schafer,
1989).
Averaging and attribute computation

The previous section discussed the problems arising when
trying to match the seismic and petrophysical data in the
time (vertical) dimension. Matching both data sets in
the horizontal spatial coordinates, in-line and cross-line,
also has its implications. Since the in uence of medium
properties in a given seismic trace depends on CDP, bin
size and Fresnel zone; a giv en well must be related to a
group of traces instead of to a single one. In this way, an
average trace, which has been computed by averaging the
associated group of traces, is used for the computation of
the seismic attributes.
A nal important issue to be considered is the selection
of the seismic attributes to be used. The optimal set
of seismic attributes to be used for estimating an specic property must be determined empirically since the
relationships between attributes and properties usually
change from one site to another. Among the most commonly used are amplitude, derivative of the amplitude
and second derivativ e, integrate of the amplitude, instantaneous frequency, instan taneous phase and average frequency (Taner, 1976).
Estimation of condence intervals
In order to compute, or estimate, condence intervals for

the estimated well log v olume an statistical analysis has
to be performed on the problem's model space. Then,
the estimation algorithm must be used several times to
obtain a representative number of samples of the model
space that allows to infer a distribution and compute its
parameters (Breiman, 1973).
Well log estimates and condence intervals
Eac htime the neural netw orkis trained and then used
to compute an estimate, an independent sample of the
model space is obtained. After a representativ e amount
of sim ulations,histograms for each point in the v olume
can be computed.
Although, strictly speaking, a very large amount of simulations are required (whic hmakes the procedure very
expensive from a computational poin tof view), an approximate study can be performed by using a moderate
amount of simulations.
The obtained distributions can be considered to be normal. In this manner, a mean value estimate and its standard deviation are computed, from which a relativ e error,
or condence interval, can be obtained.
It is important to notice that this analysis only provides
information about the consistency of the estimation but
not about its bias. In fact, the bias of the estimation can
be only veried at the training locations, where the actual
w ell log information is known. Nevertheless, estimates
for the bias across the whole volume can be computed by
using the biases measured at wells and the same articial
neural netw ork tec hnique described before.
Field data example

In this section, the methodology proposed above is used
to estimate a volume of well log data from the attributes
of a 3D seismic volume from eastern V enezuela. In
the particular example presented here, it is intended
to estimate spon taneouspoten tial(SP ) logs, which are
typically used as lithological indicators.
Input data and simulation parameters

Figure 1 illustrates the region of the eld under consideration, as well as all w ell locations used for training the
algorithm. The total extension of the area under consideration is 5:73 Km2 . The time interval considered ranges
from 1:6 s to 1:7 s.
A total amount of 25 spontaneous potential logs were used

for training the neural network, which was a four layer
perceptron with 10, 12, 12 and 1 neurons en each layer,
respectiv ely. The bac k propagation algorithm w asused
for training the neural netw ork (Haykin, 1994).
T en attributes were used in this experiment; such attributes w ere the number of sample, in-line coordinate,
cross-line coordinate, integral of the amplitude, integral
of the absolute amplitude, instan taneousphase, deriv ative, second derivative, average frequency and average
amplitude. The number of neurons for each layer and
the seismic attributes were empirically selected after some
experimentation.
Simulations and results

A total of 70 simulations w ere performed. So, a total
of 70 estimated SP volumes where obtained, for which
distributions were observ ed and statistics were computed.
Figures 2, 3 and 4 present the a verage SP, its standard
deviation and relative error, respectiv ely, fora line that
crosses the locations of wells A, B and C.
Well A
Well B
Well C
1620
20
40
1640
60
1660
80
100
1680
120
1700
260
270
280
290
300
310
320
330
340
Fig. 2: Average SP section extracted from the 3D volume.

60
1620
50
1640
40
30
1660
20
1680
10
1700
260
270
280
290
300
310
320
330
340
Fig. 3: Standard deviation section.

70
1620
60
1640
50
40
1660
30
1680
20
10
1700
260
270
280
290
300
310
320
330
340
Fig. 4: Relative error section.

Fig. 1: Region under consideration and well locations.
Figure 5 shows the low-pass ltered versions of the actual

SP curv es and the average of the 70 estimated SP curves
at the three wells shown in gure 2.
Well log estimates and condence intervals
It can be seen from the gure how the estimated curves Although very expensive from a computational point
adjust to the actual curves. The correlation coecients
of view, the computation of condence intervals proobtained for the actual and estimated SP curves at wells
vides useful information for interpreting the resulting
A, B and C, were 0:92, 0:83 and 0:72, respectiv ely.
w ell log estimates.
Notice also from gure 4 that the largest relativ e error
values occur at the borders of the section. This suggests
that the estimation technique performs better when inter- Acknowledgments
polating than when extrapolating. This proves the fact
that the a vailabilit y of enough ewll data is of critical im- We thank PDVSA-Intevep for permission to publish this
w ork and Alejandro Valenciano for providing the seismic
portance for the success of the method.
and well log data.
Well A
Well B
1600
1600
1610
1610
1610
1620
1620
1620
1630
1630
1630
1640
1640
1640
1650
1650
1650
1660
1660
1660
1670
1670
1670
1680
1680
1680
1690
1690
1690
1700
150
100
50
1700
150
100
50
References
Well C
1600
1700
150
100
50
Fig. 5: Actual (blue) and estimated (green) SP values at well

locations A, B and C.
Conclusions
The proposed methodology provides a good approach

for estimating petrophysical properties and/or w ell log
data from seismic attributes, as w ell as the consistency
of such estimates, whic h allows the computation of
relativ e errors. These error values giv e a good idea of
the condence intervals associated to propert y maps.
Also, dierent from the con ventionallinear regression
techniques, this method is able to infer the nonlinear
relationships existent betw een well logs and seismic data.
Nevertheless some important considerations ha ve to be
tak en in to account when dealing with this kind of procedure:

The depth to time con version of well log data is

a very critical step. F or this reason, accurate T-Z
curves are required.
The resolution of the resulting estimates are limited

by the seismic data; i.e. the resolution of the estimated well logs will depend on the frequency content
of the processed seismic w aveeld. Ho wever, more
research needs to be done to incorporate into the
estimation high resolution information contained in
the well logs.
Breiman, L., 1973, Statistics: With a View Toward Applications, Houghton Miin Company, Boston.
Deutsch, C., Journel, A., 1992, GSLIB: Ge ostatistical Software Libr ary, Oxford University Press, New
York.
Haykin, S., 1994, Neural Net w orks:A Comprehensive
F oundation, Maxwell Macmillan, Toronto.
Oppenheim, A., Schafer, R., 1989, Discrete-Time Signal
Processing, Prentice Hall, Toronto.
Shuki, R., Schultz, P., Hattori, M., Corbett, C., 1994,
Seismic-Guided Estimation of Log Properties. Part
II: Using articial neural netw orks for nonlinear attribute calibration, The L eading Edge, V olume13,
pp 674{678.
T aner, M., 1976, Extr action and Interpr etation of the
Complex Seismic Trace, 46th Annual In ternat. Mtg.,
Soc. Expl. Geophys., Houston.
T odoro v, T., Hampson, D., Russell, B., 1997, Sonic Log
Predictions Using Seismic Attributes, CREWES Research Report, Volume 9, ch. 39.
T odoro v,T., Stew art,R., Hampson, D., 1998, P orosity Predictions Using A ttributesfrom 3C-3D Seismic Data, CREWES Research Report, V olume10,
ch. 52.

Well Log Estimates and Condence Intervals B Yusing Articial Neural Networks

Uploaded by

Copyright:

Available Formats

Well Log Estimates and Condence Intervals B Yusing Articial Neural Networks

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Well Log Estimates and Condence Intervals B Yusing Articial Neural Networks

Uploaded by

Copyright:

Available Formats

Well log estimates and con dence intervals b yusing arti cial neural networks

R afael E. Banchs* & Reinaldo J. Michelena, PDVSA-Intevep

Indirect estimation of petrophysical properties from other

Data adjustment and re-sampling

SEG 2000 Expanded Abstracts

A second problem is the di erence in resolution betw een

Averaging and attribute computation

Estimation of con dence intervals

In order to compute, or estimate, con dence intervals for

Well log estimates and con dence intervals

Field data example

Input data and simulation parameters

A total amount of 25 spontaneous potential logs were used

Simulations and results

Fig. 2: Average SP section extracted from the 3D volume.

Fig. 3: Standard deviation section.

Fig. 4: Relative error section.

SEG 2000 Expanded Abstracts

Figure 5 shows the low-pass ltered versions of the actual

Well log estimates and con dence intervals

Fig. 5: Actual (blue) and estimated (green) SP values at well

The proposed methodology provides a good approach

The depth to time con version of well log data is

The resolution of the resulting estimates are limited

SEG 2000 Expanded Abstracts

You might also like

Well log estimates and condence intervals b yusing articial neural networks

A second problem is the dierence in resolution betw een

Estimation of condence intervals

In order to compute, or estimate, condence intervals for

Well log estimates and condence intervals

Well log estimates and condence intervals