Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Multi-Attribute Analysis
Workshop
Introduction to EMERGE
The Objective of the EMERGE Program:
EMERGE is a program that analyzes well log and seismic data.
It finds a relationship between the log and seismic data at the well locations.
It uses this relationship to predict or estimate a volume of the log property
at all locations of the seismic volume.
The Data that EMERGE uses:
A seismic volume (usually 3D).
A series of wells which tie the volume.
Each well contains target log, such as porosity, which is to be predicted.
Each well also contains the information for converting from depth to time,
usually in the form of a check-shot corrected sonic log.
(Optional) One or more external attributes in the form of seismic 3D
volumes.
Theoretically, any type of log property may be used as a target for EMERGE.
Practically, the following types have been predicted successfully:
P-wave velocity
Porosity
Density
Gamma-ray
Water saturation
Lithology logs
The only requirement is that an example of the target log must exist within
each of the wells.
Since EMERGE assumes that the target log is noise-free, it is usually
important to edit the target logs before applying EMERGE.
Since EMERGE will be correlating the target logs with seismic data, the
proper depth-to-time conversion is critical. For this reason, check-shot
corrections and manual correlation are usually necessary.
Last updated: January, 2009
Inversion
Uses seismic and well log
data.
Predicts a volume of
impedance (acoustic,
elastic, shear).
Uses the convolutional
model to relate logs with
seismic.
Emerge
Starting EMERGE
10
11
To start the training process, we need to read the sonic logs from GEOVIEW
into the EMERGE main window. To do this, click on Wells / Read From
Database. The following menu appears:
The first page allows us
to select the type of
EMERGE analysis and
which wells to include.
The default is to predict
a volume using all
wells.
The left column lists all
the wells in the guide
database. We will use
them all for training, so
click on Next >>.
12
The second page is used to tell EMERGE which of the logs in the database
is the one that we are trying to predict, i.e., which one is the Target. For
this exercise, we want to predict the P-wave or sonic log, as shown. Also,
we are specifying that, although the log is measured in depth, the analysis
(Processing Domain) will be done in time. This is because the seismic data
is measured in time. We need to specify the sample rate correctly
(Processing Sample Rate), so that EMERGE can do the depth-to-time
conversion properly.
Note that the check-shot
corrected sonic log will be
used for this conversion.
Click on the Next >> button.
The Analysis Window page of
the menu will appear.
13
The Analysis Window page allows you to specify the analysis window for
training in terms of the tops that have already been entered into the
GEOVIEW database. In this project, we have entered four tops: viking,
mann, ch_top and miss. As shown below, select viking for the start of the
analysis window and miss for the end of the analysis window.
Note that the
analysis window
can be changed
later if desired.
14
At this point, the OK button becomes active, indicating that EMERGE has
enough data from the GEOVIEW database to proceed. After you click on
OK, a confirmation menu appears.
This menu appears because there
are actually two P-wave logs, the
original log and the check shot
corrected log, in each well. By
default, the most recently created
log will be used. This is the check
shot corrected log, which is shown
in the table. If you wanted to use
another log, click on the words Pwave_CheckShotCorrected for any
of the wells to produce a pullDown
menu allowing you to select the
desired log. For our case, accept
the defaults by clicking on Ok.
Last updated: January, 2009
15
The EMERGE window now shows the target log for each of the wells. The
brown horizontal lines indicate the analysis window selected on the
previous menu. We will see how to change that later. By sliding the scroll
bar, you can look at all the remaining wells.
16
Now we wish to read in the seismic data near the well locations. As
mentioned at the beginning of this exercise, the seismic data actually
consists of two SEGY files, seismic.sgy and inversion.sgy.
To load the seismic, click
on Seismic / Add Seismic
Input / From File. The
following menu appears:
17
18
In the box that defines the Type of Data, we must select the attribute type.
When the menu first appears, it shows <Unknown>. Click in this box and
choose Raw Seismic, as shown. The Attribute Name will also be set to Raw
Seismic automatically.
19
20
On this menu, we specify the geometry of the 3D volume. Note that the only
change we need to make is to set the Number of Cross-lines.
Fill in the menu as shown
(change Cross-lines to 64)
and click on Ok.
Two new
windows will
now appear.
One window
shows a table,
indicating how
the wells are
associated with
the seismic
data:
Last updated: January, 2009
21
For this data set, the well locations are correct because the X-Y coordinates
were entered appropriately in the GEOVIEW database. For other data sets,
you may need to modify the Inline and Xline columns of the table. Now click
on Ok to get the next menu.
Last updated: January, 2009
22
23
The composite seismic trace at each well location will be extracted from
the SEGY volume and the EMERGE main window will be modified to
show the additional data:
24
The final piece of data to add is an external attribute from the SEGY file
inversion.sgy. To do this, click on Seismic / Add Seismic Input / From
File. Fill out the menu as follows:
The next series of menus is identical to those used in adding the Raw
Seismic. Start by selecting the option to load a 3D volume:
25
Tell EMERGE that this is an External Attribute with the name Inversion
Result.
Click on Next >> on this menu and the one that follows. When you see the
dialog that says that the file must be scanned, click on Yes.
Last updated: January, 2009
26
Note that we are telling the program that this is a rectangular grid of lines,
with 64 Cross-lines, and 110 Inlines.
When you click Ok on this menu, two things will happen.
Last updated: January, 2009
27
28
29
30
31
The EMERGE main window shows the analysis data for each well: the
target log in red, the single seismic trace in black and the external attribute
in blue. The analysis window is also shown by brown horizontal bars on
each log. Note that this window may be different for each well.
32
(End of Exercise 1)
33
Seismic Attributes
Seismic Attributes are transforms, generally non-linear, of a seismic trace.
Two types of attributes:
Sample-based: calculated from the trace on a sample-by sample basis.
Example: amplitude envelope.
Horizon-based: calculated as averages within a window.
Example: average porosity between two horizons.
34
35
Amplitude Envelope
Amplitude Weighted Cosine Phase
Amplitude Weighted Frequency
Amplitude Weighted Phase
Average Frequency
Apparent Polarity
Cosine Instantaneous Phase
Derivative
Derivative Instantaneous Amplitude
Dominant Frequency
Filter Slices
Instantaneous Frequency
Instantaneous Phase
Integrate
Integrated Absolute Amplitude
Second Derivative
Second Derivative Instantaneous Amplitude
Time
36
EMERGE can also import external attributes. These are seismic attributes
that cannot be calculated internally because:
They are proprietary e.g. coherency
They are too complicated e.g. seismic inversion, AVO attributes, etc.
The internal attributes can be grouped into the following categories:
Instantaneous attributes
Windowed frequency attributes
Filter slices
Derivative attributes
Integrated attributes
Time (a linear ramp)
We will now look at the theory of each group of attributes with seismic
examples.
Last updated: January, 2009
37
Instantaneous Attributes
Instantaneous Attributes were first described in the classic paper by Taner et
al (Geophysics, June, 1979). They are computed from the complex trace,
C(t), which is composed of the seismic trace, s(t) and its Hilbert transform,
h(t), which is like a 90 phase shifted trace. Writing the complex trace in
polar form, as shown below, gives us the two basic attributes: the amplitude
envelope, A(t) and instantaneous phase, f(t). (Note that the term
instantaneous amplitude is used synonymously with amplitude envelope.)
C (t ) = s (t ) + ih(t )
= A(t ) exp(i (t ))
= A(t ) cos (t ) + iA(t ) sin (t )
where : i = 1
A(t ) = s (t ) + h(t )
2
h(t )
(t ) = tan
and :
s
t
(
)
Last updated: January, 2009
h(t)
A(t)
(t)
s(t)
Time
38
(t )
=
d (t )
= the instantaneous frequency
dt
39
40
41
Cosine of instantaneous
phase of inline 95.
42
43
Instantaneous frequency
of inline 95.
Amplitude weighted
frequency of inline 95.
44
45
46
47
48
Derivative Attributes
s i s i 1
,
d 1i =
t
d 1i d 1i 1 si 2si 1 si 2
.
=
d 2i =
2
t
t
The derivative examples on the next two slides are from inline 95.
Last updated: January, 2009
49
Derivative of amplitude
envelope of inline 95.
50
51
Integrated Attributes
A fifth set of attributes in EMERGE is based on the integrated seismic trace
or its amplitude envelope. The integrated values are calculated in the
following way, where si = the ith seismic or amplitude envelope sample, Ii =
the integrated value. Note that this is a running sum.
I i = si + I i 1
At the end of the running sum the integrated seismic trace is filtered by
running a default 50 point smoother along it and removing the resulting low
frequency trend. The integrated amplitude envelope is normalized by
dividing by the difference between the minimum and maximum samples
over the total number of samples. Note that the defaults can be changed in
the Attribute / Attribute Parameters menu, shown earlier.
The integrated examples on the next slide are from inline 95.
Last updated: January, 2009
52
Integrated amplitude
envelope of inline 95.
53
Time Attribute
The last attribute is the time attribute. This is simply the time value of the
seismic trace and thus forms a ramp function that can add a trend to the
computed reservoir parameter.
Here is a plot of the time attribute:
Time attribute of inline 95 (Note: it would look the same for any line in the
volume!).
Last updated: January, 2009
54
Seismic Attributes
EMERGE actually tries to find a relationship between the target log and a
combination of attributes of the seismic trace.
All attributes for a single well:
55
Cross Plotting
One way of measuring the correlation between the target data and an
attribute is to cross plot the two.
This display shows the
target log, a seismic trace
and an external attribute:
56
This is a cross plot, showing the target, P-wave, on the vertical axis
against a particular attribute.
57
y = a +b x
This line minimizes the total prediction error:
1 N
E = (y i a b * x i ) 2
N i =1
2
1 N
xy = (x i m x )(yi m y )
N i =1
where the mean is:
1 N
mx = xi
N i =1
1 N
m y = yi
N i =1
58
xy
The prediction error is the RMS difference between the actual target
log and the predicted target log.
Last updated: January, 2009
59
60
61
Click on Ok to
get this plot,
which shows the
amplitude
envelope of the
composite
seismic trace
extracted near
well 01-08.
62
63
Note that this menu will create a cross plot between the target log and any
other internal or external attribute. We may use any combination of the
wells. In addition, we may apply one of a series of non-linear transforms to
the target and/or to the attribute. Fill in the menu as shown and click Ok.
The cross plot appears:
The cross plot has used all
points within the analysis
window of every well. The
vertical axis is the target sonic
log value and the horizontal
axis is the selected attribute,
Inversion Result. A regression
curve has been fit through the
points and the normalized
correlation value of 0.47 has
been printed at the top of the
display. The normalized
correlation is a measure of
how useful this attribute is in
predicting the target log.
Last updated: January, 2009
64
When the menu has been filled in as shown above, click on Ok and the
resulting table will be displayed:
We note that the
minimum error of 298.76
results from taking the
inverse of the external
attribute Inversion
Result.
Sometimes this error can
be improved by applying
residual time-shifts
between the target logs
and the seismic data, in
addition to the check
shot corrections.
66
67
The Optimize Shifts menu allows you to select any one transform in this
case, the single attribute transform: 1/Inversion Result.
The program then tries a series of time shifts for each well to find the set
of shifts that will maximize the correlation, subject to a Maximum Shift of
10 milliseconds. With the menu set as shown, click on Ok. The Shift
Logs menu now shows the suggested shifts:
To accept these shifts, click on Ok. The
EMERGE main window will be updated
to show the shifted logs. Now
recalculate the single attribute
transforms by clicking on Attribute /
Create Single Attribute List.
68
Accept the defaults, and recompute the single attribute list with the
shifted target logs:
Note that the minimum error
has now decreased to 289.75,
corresponding to predicting the
square root of the target log
with the attribute 1/(Inversion
Result).
The Single Attribute List
shows the result of crossplotting each attribute and
ranking the result by
increasing error.
If you select any row in this table by clicking in one of the fields, and then
click the Cross Plot button at the bottom of the table, the corresponding
cross plot will be displayed.
Last updated: January, 2009
69
Click in one of the cells of the first row (Sqrt(P-wave) vs. 1/Inversion Result)
and press the Apply button. The Application Plot window will appear:
This display shows the
target log for each well
along with the
predicted log using
the selected attribute
and the derived
regression curve. To
get a closer look at the
result, click on View /
Zoom at the top of this
window and use the
mouse to select a
rectangle around the
analysis window for the
first three wells.
70
(End of Exercise 2)
71
Multiple Attributes
An extension of the conventional cross plot is to use multiple attributes.
Cross plotting against 1 attribute
(best fit is a line):
72
73
(t ) = w0 + w1 I (t ) + w2 E (t ) + w3 F (t )
where:
(t) = porosity
I(t) = acoustic impedance
E(t) = amplitude envelope
F(t) = instantaneous frequency
In matrix form:
1 = w0 + w1 I1 + w2 E1 + w3 F1
2 = w0 + w1 I 2 + w2 E2 + w3 F2
N = w0 + w1 I N + w2 E N + w3 FN
1 1 I1
1 I
2
2 =
N 1 I N
E1
E2
EN
F1 w 0 or:
F2 w 1
w 2
FN w 3
P = AW
74
W = A A
T
A P
or:
75
76
77
(3) Step 3: Find the best triplet of attributes, assuming that the first
two members are attribute1 and attribute2. For every other
attribute in the list, form all triplets:
(attribute1, attribute2, Amplitude Weighted Phase),
(attribute1, attribute2, Average Frequency), etc.
The best triplet is the one with the lowest prediction error. Call
this third attribute attribute3.
Carry on this process as long as desired.
78
Validation of Attributes
How can we know when to stop adding attributes?
Adding attributes is similar to fitting a curve through a set of points, using
a polynomial of increasing order:
Fourth Order
79
80
81
1 = w0 + w1 I1 + w2 E1 + w3 F1
2 = w0 + w1 I 2 + w2 E2 + w3 F2
N = w0 + w1 I N + w2 E N + w3 FN
82
(2) With the derived coefficients, calculate the prediction error for
Well1. This means calculate the following:
1 N
E = ( i w 0 w 1 * I i w 2 * E i w 3 * Fi ) 2
N i =1
2
where now only data points for Well1 are used. This gives us the
Validation Error for Well1, E1.
(3) Repeat this process for Well2, Well3, etc., each time leaving the
selected well out in the calculation of regression coefficients,
but using only that well for the error calculation.
(4) Calculate the Average Validation Error for all wells:
EA
Last updated: January, 2009
(
E1 + E2 + E3 + E4 + E5 )
=
5
83
The lower (black) curve shows the error calculated using the Training Data.
The upper (red) curve shows the error calculated using the Validation Data.
The figure above shows that when more than 6 attributes are used, the
Validation Error increases, meaning that these additional attributes are
over-fitting the data.
Last updated: January, 2009
84
In this exercise, we apply Multi-Attribute Analysis to the data from the first
two exercises.
To initiate the multi-attribute
transform process, click on
Attribute / Create Multi
Attribute List to see the
menu on the left.
This menu contains three
pages of parameters. The
first page is used to select
which wells will be used in
the training. To accept the
default, which is all the
wells, click on Next >>.
Last updated: January, 2009
85
The second page of the Create Multi-Attribute List menu looks like this:
The first item on the menu
determines whether we will
be creating a list of
transforms or adding a
single transform. Usually,
we want to create a list by
examining all the available
attributes using the process
of step-wise regression.
86
87
This table shows the results of the step-wise regression. Each row
corresponds to a particular multi-attribute transform and includes all the
attributes above it. For example, the first row, labeled 1/Inversion Result,
tells us that the best attribute to use alone is the inverse of Inversion
Result. The second row, Time, actually refers to a transform that uses both
1/Inversion Result and Time together as the best pair.
Last updated: January, 2009
88
89
This cross plot differs from the previous in that it shows the predicted
target value against the actual target value. The red line is not a regression
line but a line with zero intercept and slope 1, indicating perfect correlation.
The actual correlation and error are printed at the top and we can see that
the result of using 7 attributes is to achieve a correlation of just under 62%.
Last updated: January, 2009
90
91
You may want to compare this result with the prediction using the best
single attribute. To do that, click on Attribute / Display Single Attribute List,
select the first row, (Sqrt(P-wave) vs. 1/Inversion Result), and click on
Apply. You can see that the correlation increased from 51% to 62% when
we used seven attributes.
Return to the Multi-attribute Table and highlight the words Dominant
Frequency and click on the List button. The following table appears:
(End of Exercise 3)
Last updated: January, 2009
92
93
94
P = w0 + w1 A1 + w2 A2 + + wN AN
P = w0 + w1 A1 + w2 A2 + + wN AN
1
I1
E1
= w + w I + w E
0
1 2
2 2
2
3
I 3
E3
Note that the same weights are applied to all samples. Thus, for the
second sample, we get:
2 = w0 + w1 I 2 + w2 E2
Last updated: January, 2009
95
Now let the weights become 3-point convolutional operators. Using the
convolutional equation, we get:
0 I1 w2 (0) w2 ( 1)
0 E1
1
w1 (0) w1 ( 1)
= w + w ( +1) w (0) w ( 1) I + w ( +1) w (0) w ( 1) E
2
2
0
1
1
2
1
2 2
2
w1 ( +1) w1 (0) I 3 0
w2 ( +1) w2 (0) E3
3
0
This is a new system of linear equations in which each weight, wi, has been
replaced by three weights, wi(-1), wi(0), wi(1). This can be solved by leastsquares regression just as before. The only difference is that for two
attributes, we now have 3+3+1 = 7 parameters. For example, notice that for
the second sample we now get the following equation with 7, rather than 3,
terms:
96
97
98
99
100
The multi-attribute
table that is returned
has 5 different
versions of List 2,
each for a different
length convolutional
operator.
List 1 (from the
previous exercise) is
also available.
101
102
103
104
Select the sixth attribute again, and click on Apply / Training Result. A plot
appears, showing the results of applying the multi-attribute transform along
with the target logs.
After using the View
/ Zoom option, the
plot will look like
this:
105
Another useful display can be seen if you select the sixth row on the multiattribute transform list and click on Apply / Validation Result.
This display is like the
previous one, but as
the annotation points
out, each predicted log
has used an operator
calculated from the
other wells.
Effectively, this display
shows how well the
process will work on a
new well, yet to be
drilled.
(End of Exercise 4)
Last updated: January, 2009
106
107
108
This menu allows you to specify the output file name and location, as well as
specify the range of data to process. Accept the default data range by
clicking on Next >> .
Last updated: January, 2009
109
110
111
After the
application of
the transform is
complete, you
will see Inline 95
of the predicted
P-wave velocity
volume:
112
113
114
115
When you have filled in the menu as shown, click on Next >> and OK to
produce the 6 data slices. Note that the Arithmetic Mean is shown here:
This data slice shows a lowvelocity channel feature
extending horizontally across
the survey through many of
the wells in the project. You
may want to look at the other
maps to see if the channel is
defined better.
Now, close the EMERGE
program, click on File / Exit on
the EMERGE main window.
When you see the question:
Do you want to save the
project?
Last updated: January, 2009
Click Yes.
116
We want to account
for non-linear
relationships between
logs and attributes.
Log
Attribute
Non-linear prediction:
Log
Attribute
117
MLFN
PNN
RBF
Discriminant
118
Each training example consists of the input attributes plus the known
target value for a particular time sample.
Last updated: January, 2009
119
120
Within each of the Total Iterations, simulated annealing may be used to look
for improvements by searching in other areas of the parameter space. The
decision about whether to perform simulated annealing in any iteration is
controlled by the program and depends on the degree of improvement in the
previous iteration. Theoretically, more iterations is always better than fewer
because it allows more scope for finding the global minimum.
While the training is going on, the prediction error may be monitored:
Pressing Stop on this menu allows the training to be terminated at any time.
Last updated: January, 2009
121
The parameter which controls how well the network predicts the training
data is the Number of Nodes in the Hidden Layer:
The default value follows the rule-of-thumb that it should be equal to 2/3
the number of input attributes. (Note that the number of input attributes
equals the number of actual attributes multiplied by the operator length).
Increasing the Number of Nodes in the Hidden Layer will always predict the
training data more accurately, but the danger of over-training is increased.
122
5 nodes in
hidden layer:
123
These displays show the effect of changing the number of hidden layer
nodes for the simple 1-attribute case:
5 nodes in
hidden layer:
10 nodes in
hidden layer:
124
125
126
Seismic Attribute
y1
y2
y3
y?
x1
x2
x3
Let us start with the simple case in which we try to predict an unknown log
value y from a known seismic attribute value x, using the known training
values (x1 , y1), (x2 , y2 ), and (x3 , y3 ).
Last updated: January, 2009
127
y2
y?
y3
y1
x1
x2
x3
128
Linear Regression
In linear regression, we fit the
line
y = w0 + w1x
to the points.
8
6
y = 4.5
4
2
0
x=5
129
PNN
In PNN, we fit weights to the
points themselves, or:
y = w1y1 + w2y2 + w3y3 ,
where the weights are
functions of x.
y=5
x=5
130
PNN Weights
The PNN weights are given as:
d12
w1 = S exp 2 ,
d 22
w2 = S exp 2 ,
y2
y3
d
w3 = S exp , where :
y1
1
S=
d 32
d12
d 22
exp 2 + exp 2 + exp 2
2
3
2
d2 d3
d1
x1
x2 x
x3
Note that di is the distance from the i th training point to the output point, the
factor S forces the weights to sum to 1, and determines the fit.
Last updated: January, 2009
131
( x x2 ) 2
g 2 ( x) = exp
= 2.0
= 1.0
= 0.5
x
Last updated: January, 2009
x2
132
= 0.5
= 2.0
In the previous PNN result, = 1.0. The above displays show values of 0.5
and 2.0. As increases, the fit becomes smoother, but does not fit the
training points perfectly.
Last updated: January, 2009
133
PNN Validation
To determine which value
of sigma is correct, we use
cross-validation, in which
known values are left out of
the training process.
The simple example on the
right shows that the
validation points (open
circles) are fit best using a
sigma value of 2.0, even
though this value produces
a curve which does not
correctly fit the training
data.
134
Log
X
p1
p2
x1
x2
x3
p3
p?
y1
y2
y3
Now let us consider the same problem using 2 attributes, but still 3
training points and one unknown point.
Last updated: January, 2009
135
y2
p1
y1
p2
d2
d1
p
d3
p3
y3
x
x1
x3
x2
Note that the only change is that we now can think of the points in attribute
space as being 2-dimensional, and that distance is now computed by:
d i2 = ( xi x ) + ( yi y )
2
2
136
Practical PNN
137
138
The figure on the left shows the validation of multilinear regression on four
well logs, using six attributes, and the figure on the right shows the
validation of the PNN.
Last updated: January, 2009
139
140
Training the PNN means finding the best set of sigma values for each
attribute.
By definition, the best set of sigmas is the one which produces the
minimum cross-validation error.
Cross Validation means hiding data on a well-by-well basis or on a point-bypoint basis. The well-by-well default is always recommended:
141
Sigma reduced
to 1/10 the
optimized value:
142
These displays show the effect of changing the single sigma value for the
simple 1-attribute case:
Sigma optimized
automatically:
Sigma reduced to
1/2 the optimized
value:
143
These displays show the effect of changing the single sigma value for the
simple 1-attribute case:
Sigma optimized
automatically:
Sigma increased
to 2 times the
optimized value:
144
145
The RBF network is similar to the PNN in that there is a weight for each
training point and the weights are multiplied by gaussian functions of
attribute distance that are controlled by a sigma parameter.
146
RBFN
In the RBF network the fitting
function is given as:
y = w1 g1 + w2 g 2 + w3 g 3 ,
d i2
where : g i = exp 2 .
Note that gi is equal to the PNN
weight without the scaling. In
the example shown, the
individual curves (light lines)
and the final result (heavy line)
are shown. The training points
are correctly predicted.
= 1.0
147
= 0.5
= 2.0
Two different values are shown above. As sigma decreases, the weights
converge to the training values (i.e. wi = yi). As increases, the fit becomes
smoother. Also note that the training points are always correctly predicted.
Last updated: January, 2009
148
RBFN Validation
Again, we will use the crossvalidation technique to
determine which value of
sigma is correct, in which
known values are left out.
The simple example on the
right shows that the
validation points (open
circles) are fit best using a
sigma value of 1.0, even
though this value produces a
curve which is not as smooth
as for a sigma of 2.0.
149
y1 1 g12 g13 y1
y
y = g
1
g
23 2
2 12
y3 g13 g 23 1 y3
( xi x j ) 2
d ij2
where : g ij = exp 2 = exp
.
2
w1 g11
w = g
2 21
w3 g 31
g12
g 22
g 32
g13
g 23
g 33
150
The figure on the left shows the application of the PNN on four well logs,
using six attributes, and the figure on the right shows the application of the
RBF network.
Last updated: January, 2009
151
The figure on the left shows the validation of the PNN on four well logs, and
the figure on the right shows the validation of the RBF network.
Last updated: January, 2009
152
Practical RBFN
As in the PNN, is the most important parameter in the RBF network and
needs to be optimized. Optimization is done using cross-validation, in
which each well is left out of the training process and predicted, one at a
time.
Unlike the PNN, is not allowed to vary for each attribute in the RBF
network.
153
Disadvantages:
(1) Unlike the PNN, in which sigma is allowed to vary for each attribute, the
RBF network is optimized for a single value of sigma.
(2) For small values of sigma, the fitting function can have large swings
between points.
Last updated: January, 2009
154
155
156
This well contains a porosity log, called Porosity, along with the other
logs. Restart the EMERGE program and select Start New Project. Call
the project porosity as shown:
157
The steps for loading the data into this project are identical to the
previous exercise, so they will be summarized briefly here. Click on Wells
/ Read From Database and fill in the three pages as shown below:
158
Now click on OK. Once again this menu appears, because there is more
than one P-wave log within the database:
159
160
161
162
Click Ok on this
menu.
163
164
The EMERGE main window now shows the extracted seismic trace and
the composite impedance trace from the external attribute (Inversion
Result):
165
One option in Emerge, which can sometimes improve results is to filter the
target logs to the seismic bandwidth.
The logic for doing this comes
from the fact that the target log
is usually much higher
frequency than the seismic or
attributes. We hope that by
filtering the target logs, we can
create an easier problem for
Emerge. We will do that in this
exercise.
First, we need to determine the
frequency content of the input
data. To do that, go to the
window containing the seismic
volumes, and click on Process /
Utility / Amplitude Spectrum:
Last updated: January, 2009
166
167
168
A new window appears, showing the filtered logs over the original logs. By
zooming, you can see that the highest frequency detail has now been
removed:
169
The data is now loaded and ready for analysis. The first step is to examine
the single-attribute transforms. To do this, click on Attribute / Create Single
Attribute List to get this menu:
Note that we are
choosing to test nonlinear transforms applied
to both the target
(porosity) and the
external attribute
(Inversion Result). This
is suggested by the
expected negative
correlation between
these two data sets.
170
171
172
173
174
175
To see the application, select the fourth row of the Multi-attribute Table
(Y_Coordinate) and click on Apply / Training Result:
176
Now apply this result to the SEGY data. Go to the Seismic Display window
and click on Process / Apply EMERGE. Fill in the first menu as shown
below:
177
178
179
(End of Exercise 6)
180
181
Note that there may be two reasons to leave a well out of the training:
(1) The well-to-seismic tie is poor.
(2) You may wish to use the well for blind well testing or validation later.
Last updated: January, 2009
182
This is usually
recommended since stepwise regression is the best
way to determine which
attributes to use.
Last updated: January, 2009
183
184
185
These parameters control the option to cascade the Neural Network with
the trend from the multi-attribute transform. This option exists because
Neural Networks usually work best with stationary data containing no long
period trend.
Sometimes it is best to remove the trend from the target data and use the
Neural Network to predict the residual data which is left after trend
removal.
In this option, the following steps are followed:
(1) The multi-attribute transform is used to predict the target logs.
(2) The predicted logs are smoothed using a running average.
(3) The smoothed predicted logs are subtracted from the original logs.
(4) The Neural Network is then trained on the residual or difference.
Last updated: January, 2009
186
The only way to tell if this option is helpful is to create Neural Networks both
ways and look at the training and validation errors.
188
189
The question at the top of this page asks if we wish to use one of the previously
calculated multi-attribute transforms to structure the Neural Network. Usually, the
answer to this is Yes. This is because the multi-attribute selection process has
determined which attributes are best for predicting the target porosity log. By
selecting the fourth transform (Y-Coordinate), we are constructing a Neural Network
with precisely the same attributes as those used in that transform. To see the details
of that multi-attribute transform, click on Show Transform History. With the menu
filled in as shown above, click on Next >> to get the network parameters page:
We will start by creating a
Probabilistic Neural Network, as
shown. For this network, we will not
cascade with the trend from the multiattribute transform. We will do this
later and the process will be explained
then. Finally, by choosing the type of
analysis as Mapping, we are
specifying that we wish to predict
numerical values for the porosity and
not classification types.
Last updated: January, 2009
190
191
192
Note that the correlation of 0.87 is much higher than that achieved with
multi-attribute regression. This is usually the case with Neural Networks
because of the non-linear nature of the operator. Note also that the Neural
Network has been applied only within the training windows. This is done
for two reasons:
(1) The application time for the
Neural Network can be very long if
applied to the entire window.
(2) The Neural Network is not very
good at extrapolating beyond the
bounds of the training data. For
this reason, it is expected to be
less valid outside the training
windows than the multi-linear
regression.
After zooming, the result looks like
this:
Last updated: January, 2009
193
Now we would like to see how the network performs in Validation Mode. This
means that we will hide wells and use the trained network to predict their
values.
To start this, click on Neural / Validate Neural Network. The following menu
appears:
194
Click on Next >> to select the network which has been trained. The next
page appears:
Since all the wells were used for training, only the first selection is appropriate.
This means that each of the training wells will be hidden in turn and predicted
using the remaining wells. Click on OK to start this process.
Last updated: January, 2009
195
196
Another possibility for improving the PNN result is to use the trend from the
multi-linear regression calculation. This is sometimes useful because Neural
Networks operate best on data with stationary statistics, i.e., data sets
without a significant long period trend.
To evaluate this option, we will create a new network. Click on Neural / Train
Neural Network. When the menu appears, click on Next >> to accept the
default new network name:
197
198
In this mode, the first calculation that the network performs is the multi-linear
regression with the same four attributes. The predicted log from that
calculation is then smoothed with a smoother length given on the Neural
Network training menu. The PNN Neural Network is then used to predict the
residual, which is the high-frequency component of the logs which is not
contained within the smooth trend. The final predicted log is obtained by
adding the trend from the multi-linear regression and the predicted residual
from the Neural Network.
199
200
Now click on Neural / Display Neural Network List to get a list of all
networks created so far:
As an example of the
capabilities of this list, click on
the name Network_1 and click
the Cross Plot button. The
Cross Plot window appears:
201
202
203
If horizons had been entered into the project, these could have been used to
guide the Neural Network application window. Since there are no horizons,
we will apply the network from 1000 to 1200 ms. When the menu has been
filled
as 2009
shown above, click on Next >> and OK to start the process.
Last
updated:in
January,
204
205
On the menu, go to
the Insert tab, and
change all the items
shown. Note that we
are inserting the
actual Porosity logs
into the display.
Click on the Color
Key tab.
206
207
Now display Inline 95. As before, the high porosity channel can be seen at
the well location:
208
By scrolling through the 3D volume, you can check how well the porosity
volume ties the original porosity logs. For instance, this shows Xline 26,
which actually ties 3 wells.
209
This ends the porosity exercise. On the EMERGE main window, click
Project > Save to save the project.
Do not exit the project. In the next exercise, we will next use the same
seismic and inversion data but with a different target log.
(End of Exercise 7)
210
Case Study:
Using EMERGE to predict Vclay from
Simultaneous Inversion attributes
212
Objective
213
Petrophysical Analysis
Petrophysical analysis and modeling: log and core data from 42 wells.
Core, density and P- and S-wave velocity logs: available in most wells.
Standard processes:
log editing
normalization and invasion correction
reservoir parameter interpretation: clay volume (Vclay), porosity and
water saturation (Sw)
214
Petrophysical Analysis
DEPTH
M
DEPTH_FT
ft
GR
GAPI 100
SP
-190 MV
10
CALI
100
MM
200
RES_D
0.2 OHMM 2000
RHOB_RAW
1.65 g/cc 2.65
VP_FINAL
4000 ft/s10000
AI
10000
g/cc-f/s
20000
PHIE_FIN
1
v/v
0
RES_M
0.2 OHMM 2000
NPSS
0.6 V/V
VS_FINAL
2000 ft/s 5000
PR
v/v
BVW_TMS
DEC
RES_S
0.2 OHMM 2000
VPVS
v/v
VCL_FIN
v/v
PEF
B/E
VS_FLAG
5
20
0.5
G
GPa
SI
0 ft/s-g/c
10000
QUAL_VS
0
120
VCL_FIN
RHOB_RAW NPHI
20
400
130
450
140
150
500
160
170
550
180
McM
600
190
200
210
650
Mud Plug
700
220
230
Density
Vs,
Vp
IP
IS
750
240
800
250
215
Petrophysical Analysis
216
Petrophysical Analysis
LR = Ip2 2Is2
Sands
MR = 2Is2
Shales
217
Petrophysical Analysis
Sands
Shales
218
Petrophysical Analysis
219
Simultaneous Inversion
Resulting inversion volumes: Zp, Zs, Density, Vp/Vs,
Lambda-Rho and Mu-Rho.
Zp
Zs
Density
Lambda-Rho
220
221
222
Reflectivity
Last updated: January, 2009
Volume of Clay
223
224
Conclusions
EMERGE is a powerful tool for predicting log properties from seismic
attributes.
While EMERGE has been used for a number of years, recent new success has
come from using pre-stack and simultaneous inversion results as attributes.
This case study has shown the successful prediction of a Vclay volume from
simultaneous inversion results.
225
Log
Class A
Seismic Attributes
X
Y
x2
x1 y1
y2
x3
x0
Class B
x5
x4
x6
y3
p1
p
p3 2
y0
y4
y6
y5
p4
p
p6 5
226
p1
p2
p3
d1
d3
d2
p0
(xi x0 )2 + ( yi y0 )2
Class A
d4
Class B
d5
d6
p5
p4
p6
X
Last updated: January, 2009
227
As with the mapping option, PNN classification does not use distance on its
own, but applies an exponential weighting function to the distance (called the
Parzen Estimator). For the two classes, we can write:
g A ( p0 ) = e
d12
+e
d 22
+e
d 32
, and g B ( p0 ) = e
d 42
+e
d 52
+e
d 62
PA
g A ( p0 )
g B ( p0 )
, and PB
=
g A ( p0 ) + g B ( p0 )
g A ( p0 ) + g B ( p0 )
The decision is then simple. If PA > PB, the point p0 is in Class A and if PA <
PB, the point p0 is in Class B.
Last updated: January, 2009
228
Classification
229
These are the menu items which control the use of Classification:
If the target logs have been classified previously, they must still be read
into EMERGE as numerical values, where the numbers represent the
classes.
Last updated: January, 2009
230
231
Discriminant Analysis
Discriminant
Line
232
Discriminant Analysis
Because discriminant analysis assumes a linear separation between
clusters, it can fail if the real separation is non-linear:
Attribute 1
Attribute 2
Last updated: January, 2009
233
Discriminant Analysis
Advantages:
(1) Both training and application times are much faster than any Neural
Network.
(2) The algorithm is very robust, with little tendency to over-train. This
means that cross-validation errors are usually comparable to training
errors.
Disadvantages:
(1) Only works in Classification mode.
(2) Assumes linear separation between classes.
234
235
236
To create the
classification log, the
original porosity log
has been divided
into 3 zones.
The objective is to
predict the locations
of the high porosity
zones, and the
probability of
occurrence.
Zone 1 : Porosity < 5%.
Probably shales.
Now we will start a new Emerge project but this time we will use the
existing porosity project to speed up the data loading.
(If Emerge is closed, open it by clicking on the Emerge button and
open the previous porosity project.)
238
239
240
The software then pops up a warning message about losing all the training
work. Since we have already saved the old porosity project, then this is not a
problem. Click Yes.
241
242
243
244
245
246
When the training has finished, the result looks like this. The Fractional
Classification Error means that 19% of the input samples were misclassified, i.e., predicted wrong.
In this plot, when blue overlays red, the classification is correct. Where we
see red lines, that indicates mis-classification.
Last updated: January, 2009
247
248
249
250
251
The resulting plot shows the high porosity clean sands in red:
252
253
The resulting plot looks like this. Once again we can improve the display
by changing the color key.
254
255
The resulting plot shows a very high probability (>90%) of clean, highporosity sand at the channel location
(End of Exercise 8)
Last updated: January, 2009
256
257
258
259
260
Application of 4 attributes
261
Cross-validation result
(blind well test)
262
Pwave alone cannot estimate shallow gas sands and carbonate velocity correctly
263
Neutron log improves fit for shallow gas sands and carbonates
264
Effect of density
267
To change to the new database containing the data for this example,
click on Database / Open. On the Directory Chooser menu, select the
logs database, as shown below and click OK:
268
As you can see, there are four wells in this database. To see the log curves
in the first well, select the B_Yates_11 well name and click on Display Well.
This display will appear:
Move the scroll bar,
and you will see the
eight logs that are in
this well, including a
sonic log (DLT).
Another way of
examining the logs
within a well is to click
on the arrow to the
left of the well name in
the GEOVIEW Well
Explorer window.
269
270
We see that this well, B_Yates_11, contains nine logs, including a sonic and
its depth-time relationship. Examine the other wells and you will find that one
other well, B_Yates_18D, has a sonic log, while two of the wells, B_Yates_13
and B_Yates_15, do not contain sonic logs. The objective of this exercise is
to predict sonic logs using the other log curves.
Restart the EMERGE program by clicking the EMERGE button on the
GEOVIEW main window and select Start New Project. Call this new project
logs, as shown below:
271
To load the analysis logs from the database, click on Wells / Read From
Database.
Unlike the
previous
exercises, we will
use logs as the
external attributes.
Click on
Predicting Logs
From Other Logs.
272
273
274
275
After you click OK, the EMERGE main window will look like this:
By moving the
scroll bar, you can
see each of the
four wells and their
associated logs.
You will also notice
that two of the
wells do not
contain target logs.
276
Now click on
Display / Crossplot.
Fill in the menu as
shown below:
277
Now go once again to the EMERGE main window and select Display /
Crossplot. This time, select RILD as the attribute:
278
279
280
Notice that the options Test Non-Linear Transforms of Target and Test
Non-Linear Transforms of External Attributes are checked. This means
that for each of the selected External Attributes, Caliper, Gamma Ray, etc.,
EMERGE will create a series of new attributes by applying a set of nonlinear transforms.
Click on Ok on this menu
and the following table
will appear:
281
282
This display shows all four predicted sonic logs in red. Two of the wells,
which contain target logs, show those logs in black. By using the View /
Zoom feature, you can see how well the target sonic log has been predicted
using the RILM curve:
283
284
285
286
On this display, click on File / Export to Database. This will allow you to
send the predicted logs back to the GEOVIEW database, where they can be
used just like any other log. By default, the sonic log from EMERGE will be
exported to every well in the database.
287
This completes the analysis of the third data set. To close the
EMERGE program, click on File / Exit on the EMERGE main window.
When you see the question: Do you want to save the project?, click
on Yes.
Last updated: January, 2009
(End of Exercise 8)
288
289
290
291
292
293