You are on page 1of 95

13-1

Seismic Inversion and AVO


applied to Lithologic Prediction

Part 13 Neural Network
applications

13-2
Introduction
In this part of the course, we will look at three different
case studies that illustrate neural network applications.
The first case study discusses the basics of the multi-
layer feedforward neural network (MLFN) using a
simple AVO classification problem.
The second case study discusses the prediction of
porosity and water saturation from model data using the
MLFN and PNN neural network approaches.
The third case study compares the Radial Basis
Function Neural Network (RBFN) with the Generalized
Regression Neural Network (GRNN).
These case studies should help illustrate the methods
that were discussed in the last two sections.
13-3
Neural Networks and AVO
In this first case study, we will discuss the
application of neural networks to a simple AVO
problem.
The AVO problem to be solved will be a Class 3
gas sand anomaly.
The type of neural networks that we will use to
solve the problem is the multi-layer perceptron
(MLP).
One of our key goals will be to show why a multi-
layer perceptron can solve this problem, whereas a
single layer perceptron cannot solve it.
This paper was published in The Leading Edge.
13-4
Shale
Shale
Sand Wet
3
S
p
cm / g 1 . 2
s / m 1000 V
s / m 2000 V
=
=
=

3
S
P
cm / g 2 . 2
s / m 1167 V
s / m 2334 V
=
=
=

shale overlying as same


Shale
Shale
Sand Gas
3
S
p
cm / g 1 . 2
s / m 1000 V
s / m 2000 V
=
=
=

3
S
P
cm / g 0 . 2
s / m 1046 V
s / m 1720 V
=
=
=

shale overlying as same


Gas and Wet Models
(a) Wet sand model.
(b) Gas sand model.
13-5
The Aki-Richards equation
The Aki-Richards linearized equation for the P-wave
reflectivity as a function of angle, as modified by
Wiggens, can be written as the sum of three terms:
u u u u
2 2 2
PP
sin tan C sin B A ) ( R + + =
where:
p
P
2
P
S
S
S
2
P
S
p
P
o
PP
p
P
V
V
2
1
C
V
V
2
V
V
V
V
4
V
V
2
1
B
) 0 ( R
V
V
2
1
A
A

A A A

A A
=
(

=
~
(
(

+ =
In our case, we will use angles less than 30
o
, which
means that the third term can be dropped.
13-6
Sand / Shale Parameters
Type V
P
(m/s)

V
S
(m/s)

(g/cc) V
P
/V
S
o
Shale 2000 1000 2.1 2 0.33
Wet SS 2334 1167 2.2 2 0.33
Gas SS 1720 1046 2.0 1.64 0.21
The above table shows the parameters from the gas and
wet sand models, including the V
P
/V
S
and Poissons ratio o.
Since the V
P
/V
S
ratio for the shale and wet sand is 2, it can
be shown that B = - A for the top and base of the wet
model. Also, for these parameters, B = A for the top and
base of the gas model.
13-7
Why does B = -A?
Assuming that V
P
/V
S
= 2 in the full Aki-Richards equation:
Also, again if V
P
/V
S
= 2, then:
(

+
(

+ =
=

A A

A A

A A A
S
S
P
P
S
S
p
P
V
V
V
V
2
1
2
1
V
V
V
V
2
1
B
S
S
P
P
S
P
S P
V
V
V
V
2
V
V
V 2 V
A A
A
A
= = =
Combining (1) and (2) we find that, if V
P
/V
S
= 2, then:
A B =
(1)
(2)
13-8
Model AVO Curves
-0.15
-0.10
-0.05
0.00
0.05
0.10
0.15
0 5 10 15 20 25 30
Angle (degrees)
A
m
p
l
i
t
u
d
e
Top Gas Base Gas Top Wet Base Wet
The AVO curves for the top and base of each model are shown
above. Notice that the gas sand shows an increase in amplitude,
whereas the wet sand shows a decrease in amplitude.
13-9
Rutherford / Williams Classification
Rutherford and Williams (1989) derived the following
classification scheme for AVO anomalies, with further
modifications by Ross and Kinman (1995) and Castagna
(1997). The acoustic impedance changes refer to the
anomalous layer:

Class 1: Large increase in acoustic impedance.
Class 2: Near-zero impedance contrast.
Class 2p: Same as 2, with polarity change.
Class 3: Large decrease in acoustic impedance.
Class 4: Very large decrease in acoustic impedance
coupled with small Poissons ratio change.

Note that our model is a Class 3 anomaly gas sand.
13-10
The plot above shows the synthetic seismic responses for the two
models, at a depth of 500 m, a thickness of 20 m, and using a 20 Hz
Ricker wavelet. Note again the increase in amplitude for the gas sand
and the decrease in amplitude for the wet sand.
Gas Sand Wet Sand
Synthetic seismic responses
13-11
Interface Values
Interface A(x10) B(x10)
Top Gas -1 -1
Base Gas +1 +1
Top Wet +1 -1
Base Wet -1 +1
The parameters used in the two models are realistic, but
obviously designed for this problem. In addition to A=-B for
the wet sand and A=B for the gas sand, all of the reflection
coefficients are +/- 0.1. This leads to the table shown above,
after scaling all values by a factor of 10. This is similar to the
exclusive OR (XOR) problem of Boolean Algebra.
13-12
Wet Trend
Top Wet
Base Gas
Base Wet
Top Gas
+1
-1
-1
+1
A
B
A vs B crossplot of model sands
The plot above shows the intercept vs gradient (A vs B) crossplot for
the two sands, where the wet sands define the wet trend and the gas
sands are in the other two quadrants, indicative of a class 3 sand.
13-13
Gulf of Mexico example
Figure (a) shows a relative
amplitude seismic line over
a Gulf of Mexico gas sand
bright spot, (b) shows a
crossplot of the A and B
attributes from the line and
(c ) shows the position on
the line of the ellipses from
(b), where gray=wet trend,
yellow and blue=gas sand.
(a)
(b)
(c)
13-14
An introduction to neural networks
An artificial neural network (ANN) is a mathematical
algorithm that can be taught to solve a problem that
would normally require human intervention.
Generally, we classify neural networks into two broad
groups, based on how they learn:
Supervised methods (e.g. MLP, PNN).
Unsupervised methods (e.g. the Kohonen self-organizing map).
There are two key geoscience problems that neural
networks can solve:
Classification (e.g. gas versus water).
Prediction (e.g. prediction of porosity from seismic attributes).
Our problem today is a supervised fluid classification
problem.
13-15
( ) x f
b
y
x
1
w
2
w
N
w
2
a
N
a
1
a
The single-layer perceptron
The plot above shows a schematic of a single-layer perceptron (SLP).
The SLP accepts N inputs, applies a weight to each, and then sums the
result along with a bias, b. The sum is then passed through a non-linear
function, f(x), to produce the resulting output.
Perceptron (P)

13-16
Hyperbolic Tangent Function
-1
-0.5
0
0.5
1
-3 -2 -1 0 1 2 3
x
f
(
x
)
Common neural network functions
(a) Hyperbolic tangent
function (tanh(x)).
(b) Symmetric step function.
x x
x x
e e
e e
) x ( f

> +
<
=
0 x , 1
0 x , 1
) x ( f
13-17

1
w
2
w
x
b
y
A
B
Single-layer perceptron applied to AVO
The flowchart above shows the application of the single-layer perceptron
to the AVO problem. Notice that there are two inputs, the intercept (A)
and gradient (B), and that we are using the symmetric step function. The
interpretation of the output will be: +1 = Gas Sand, and 1 = Wet Sand.
13-18
Single-layer perceptron with two inputs
Analyzing this simple case will give us an intuitive
understanding of the single-layer perceptron. Solving for
x, we get:
b B w A w x
2 1
+ + =
This is simply a straight line in the 2D plane. If we set x=0,
we find that this line crosses the A and B axes at:
.
w
b
B and ,
w
b
A
2 1

=
Note that the line x = 0 marks the boundary, or decision
plane, between f(x) = +1 and f(x) = -1. This is illustrated
on the next slide.
13-19
Decision Boundary
1
w
b

2
w
b

vector weight w =
0
( ) 1 x f + =
( ) 1 x f =
( ) 0 ) x ( f =
A
B
Decision boundary for the two-input
single-layer perceptron
The problem
with the SLP is
that it can only
separate linearly
separable points,
using a line in
2D space or a
hyperplane in
N-dimensional
space.
13-20
1 +
1 +
Gas Base
Gas Top
A
B
1
1
1 +
Gas Base
Gas Top
A
B
1 +
1
1
Decision boundaries for the AVO model
(a) If we use a weight vector
pointing to the lower left quadrant,
we can find the top of the gas sand.
(b) If we use a weight vector
pointing to the upper right quadrant,
we can find the top of the gas sand.
But we cant solve the whole problem with a single-layer perceptron!
13-21

-1
-1
-1
x
1
y
1
A
B

+1
+1
-1
x
2
y
2
A
B
Solving for the weights and biases
(a) Above are shown the simplest weights and bias for the top
of gas sand, since w
1
= w
2
= w < 0, and A = B = -b / w = -1.
(b) Above are shown the simplest weights and bias for the base
of gas sand, since w
1
= w
2
= w > 0, and A = B = -b / w = +1.
13-22
Verifying the weights and biases
Here are the calculations using the weights and biases
shown in the previous slide, showing that they give the
correct answers. Recall that +1 = Gas and 1 = Wet.
Sand A B x
1
y
1
x
2
y
2
Top Gas -1 -1 +1 +1 -3 -1
Base Wet -1 +1 -1 -1 -1 -1
Top Wet +1 -1 -1 -1 -1 -1
Base Gas +1 +1 -3 -1 +1 +1
13-23
The mathematics of a single-layer
perceptron
In the general case, for M single-layer perceptrons, each with N
inputs, we can write the equation as follows:
) W ( f b a y + =
.
b
b
b
and ,
a
a
a
,
w w w
w w w
w w w
W ,
y
y
y
M
2
1
N
2
1
MN 2 M 1 M
N 2 22 21
N 1 12 11
M
2
1
(
(
(
(

=
(
(
(
(

=
(
(
(
(

=
(
(
(
(

b a y
where:
) b ( f
T
+ = a w y
| | . value scalar a b and , w w w
N 2 1
T
= = w
where:
Note that, for a single perceptron, the mathematics becomes:
13-24
The multi-layer perceptron
The limitation of the single-layer perceptron, that it
can only solve linearly separable problems, can be
overcome by adding extra layers to the perceptron.
Each new layer of perceptrons accepts the outputs
from the previous layer as if they were new input
values.
A two-layer perceptron with N inputs and M
perceptrons is shown in the next slide. The first layer
is called the hidden layer and the second layer is
called the output layer.
We will then look at the mathematics of the multi-layer
perceptron, and see how it can be used to solve our
AVO problem.
13-25
1
a
2
a
) 1 (
M
p
) 2 (
p
) 2 (
y
N
a
) 1 (
11
w
) 1 (
21
w
) 1 (
1 M
w
) 1 (
12
w
) 1 (
22
w
) 1 (
N 1
w
) 1 (
N 2
w
) 1 (
MN
w
) 1 (
2 M
w
) 1 (
M
y
) 1 (
2
y
) 1 (
1
y
) 2 (
1
w
) 2 (
3
w
) 2 (
2
w
) 1 (
1
b
) 1 (
2
b
) 1 (
M
b

) 1 (
2
p
) 1 (
1
p
Multi-layer perceptron flowchart
Hidden layer
Output layer (can
have multiple
perceptrons).
13-26
) 1 (
2
p
) 2 (
p
) 1 (
2
p
) 1 (
11
w
) 1 (
12
w
) 1 (
22
w
) 1 (
21
w
) 1 (
2
b
A
B
) 1 (
1
b
) 1 (
2
y
) 1 (
1
y
) 2 (
2
w
) 2 (
1
w
) 2 (
b
) 2 (
y
Multi-layer perceptron with two inputs
The flowchart above shows the application of the multi-layer perceptron
to the AVO problem. Again, there are two inputs, the intercept (A) and
gradient (B), and we will use the symmetric step function everywhere.
13-27
An intuitive development of the
second layer weights
A key question is how we can derive the weights for
the full multi-layer solution.
Recall that the solutions for the top and base of the
gas sand using the weights derived for the single-
layer perceptrons were:
y
1
T
= (+1, -1, -1, -1)
y
2
T
= (-1, -1, -1, +1)
If we use the weights that produced these outputs as
our first layer weights, then these values become the
input for the second layer. The figure on the next
slide shows how we can then derive the weights for
the second layer.
13-28
2
y
1 +
Gas Base
Gas Top
1
y
1 +
Wet Top
Wet Base
&
Boundary Decision
1
1
Decision boundary for the second layer of
the multi-layer perceptron
Notice that the
outputs of the first
layer, when cross-
plotted, become a
linearly separable
problem. Thus, the
weights for the
second layer can
be solved by
finding the new
decision boundary.
13-29
) 1 (
2
p
) 2 (
p
) 1 (
1
p
1
1
1 +
1 +
1
A
B
1
) 1 (
2
y
) 1 (
1
y
1 +
1 +
) 2 (
y
1 +
The full solution to the AVO problem
Above are shown the simplest second layer weights
and bias for the gas sand, since w
1
= w
2
= w > 0, and
A = B = -b / w = -1.
13-30
Verifying the weights and biases
Here are the actual calculations using the weights and
biases shown in the previous slide, showing that they are
correct. In fact, the calculation is correct before applying
f(x), the symmetric step function.


Sand A B y
1
(1)

y
2
(1)


x
(2)
y
(2)
Top Gas -1 -1 +1 -1 +1 +1
Base Wet -1 +1 -1 -1 -1 -1
Top Wet +1 -1 -1 -1 -1 -1
Base Gas +1 +1 -1 +1 +1 +1
13-31
Top Wet
Base Wet
B
Final solution applied to original model
The plot above shows the final solution applied to the initial AVO model.
Note that the effect of a second layer of weights is to create a non-
linear separation on the original crossplot.
Top Gas
-1
-1
+1 Base Gas
A
+1
13-32
Mathematics of the multi-layer perceptron
Mathematically, the two-layer perceptron can be written as follows:
) ) W ( f W ( f
) 2 ( ) 1 ( ) 1 ( ) 1 ( ) 2 ( ) 2 ( ) 2 (
b b a y + + =
.
b
b
b
and ,
w w w
w w w
w w w
W
) j (
M
) j (
2
) j (
1
) j (
) j (
N M
) j (
2 M
) j (
1 M
) j (
N 2
) j (
22
) j (
21
) j (
N 1
) j (
12
) j (
11
) j (
) j ( ) j ( ) j ( ) j (
(
(
(
(
(

=
(
(
(
(
(

b
This can obviously be generalized to L layers, where the j
th
layer
weights and biases can be written:
However, it has been shown that two layers is usually sufficient in
most cases.
13-33
Postscript
In this development of the multi-layer perceptron, we
have derived the layer weights and biases intuitively.

In most problems, this would be impractical.

The actual method of solving for the weights is as follows:
Initialize using small random weights.
Change the weights so as to minimize the error
between the computed and desired output values,
starting with the weights in the output layer.
Backpropagate the error minimization for all layers.
Iterate until an acceptable error is found.
13-34
Case Study 1 Conclusions
In this section, we have shown how to use a multi-
layer perceptron to delineate a Class 3 AVO
anomaly.
First, we demonstrated that a single-layer perceptron
could only solve for the linearly separable case.
We then showed that, by adding a second layer, we
could transform a non-linear problem into a linear
problem, hence finding a solution to the full Class 3
anomaly.
Obviously, most realistic problems are more complex
than this. But the mechanics are the same.
13-35
Case Study 2 The AVO Modelling
Volume
In this paper, we will be looking at a new approach to
AVO modelling, called the AVO modelling volume.
The method involves creating a volume of modelled
CDP gathers in which two parameters are varied, one
in the in-line direction, and the other in the cross-line
direction.
This modelled volume is then processed using
standard AVO analysis techniques, such as
Intercept/Gradient analysis or R
P
/R
S
inversion.
The final result can be displayed as a data slice, or
used for more quantitative analysis.
13-36
Overview
We will first discuss the basics of the method.
We will then show three model examples:
Model 1 is a volume in which P-wave velocity is varied
against S-wave velocity.
Model 2 is a volume in which P-wave velocity is varied
against layer thickness.
Model 3 is a volume in which water saturation is varied
against porosity.
Finally, we will use a statistical multi-attribute
approach to look for relationships between the
seismic and the model. The derived relationships can
then be applied to modelled seismic data to predict a
parameter volume.
13-37
The AVO Modelling Volume Method
The AVO modelling volume concept is straightforward
but powerful.
We simply vary two parameters incrementally in each of
the in-line and cross-line directions, and then observe
the results on our AVO models.
Possible parameters that can be varied are:
Basic elastic constants such as V
P
, V
S
, density, Poissons ratio,
lambda, shear modulus, bulk modulus, or bed thickness.
Intrinsic parameters such as porosity, water saturation, and
permeability.
Extrinsic parameters such as pressure and temperature.
13-38
The AVO Modelling Volume Method
Modelled
logs
Parameter 1
P
a
r
a
m
e
t
e
r

2

AVO
Synthetic
Parameter 1 Parameter 1
Processed
AVO
Attribute
(a) Build a set of
logs (V
P
, V
S
, ) in
each bin.
(b) Build AVO
synthetics in
each bin.
(c) Process each
synthetic gather
and display.
13-39
Model 1 - V
P
vs V
S

In Model 1, we varied V
P
from 2000 to 2500 m/sec in
increments of 100 m/s, and V
S
from 1000 to 1250 m/s
in increments of 50 m/s, creating 36 modelled gathers
in a 6x6 grid. One of the logs is shown in the next
slide. Note that the density was held constant.
The Poissons ratios for each grid cell is shown in the
slide following the next. Notice the diagonal patterns.
We used primaries-only modelling with a 25 Hz Ricker
wavelet, and included NMO stretch. The modelled
gathers are shown on the third slide.
13-40
Model 1 - V
P
vs V
S
This figure shows an example of one of the 36 logs
used in the AVO modelling.
13-41
Poissons Ratio for V
P
vs V
S

V
S
(m/s)
V
P

(m/s)
This table shows the value of Poissons ratio for all the
V
P
and V
S
combinations in our 6x6 grid. Note the
bottom left to top right diagonals of constant value.
1000 1050 1100 1150 1200 1250
2500 0.40 0.39 0.38 0.37 0.35 0.33
2400 0.39 0.38 0.37 0.35 0.33 0.31
2300 0.38 0.37 0.35 0.33 0.31 0.29
2200 0.37 0.35 0.33 0.31 0.29 0.26
2100 0.35 0.33 0.31 0.29 0.26 0.23
2000 0.33 0.31 0.28 0.25 0.22 0.18
13-42
V
P
vs V
S
Modelled Cube
13-43
The Aki-Richards equation
The Aki-Richards linearized approximation to the
Zoeppritz equation can be written as the sum of 3 terms:
u u u u
2 2 2
sin tan C sin B A ) ( R + + =
where:
,
V
V
2
V
V
V
V
4
V
V
2
1
B
2
P
S
S
S
2
P
S
p
P

A A A
(

=
,
/ V / V
V / V
D ,
) 1 ( 1
2 1
) D 1 ( 2 D A B
P P
P P
2
A A
A
o
o A
o
o
+
=

+
(

+ =
,
V
V
2
1
R A
p
P
P
(
(

+ = =

A A
or (using Shueys approximation):
.
V
V
2
1
C : and
p
P
A
=
13-44
Approximate Aki-Richards
Assuming that V
P
/V
S
= 2 in the full Aki-Richards equation, it
can be shown that:
(

+ =

A A
S
S
S
V
V
2
1
R : where
( ) B A
2
1
R
S
=
Assuming that o

= 1/3 in Shueys equation (which is
identical to V
P
/V
S
= 2), it can be shown that:
( ) B A
9
4
+ = o A
S P
R 2 R B =
P
R
4
9
B = o A
13-45
V
P
vs V
S
Modelled Cube
Intercept (A) Gradient (B)
V
P

(m/s)
2000
2500
1000 1250
V
S
(m/s)
1000 1250
V
S
(m/s)
This slide shows slices from the A and B volumes at a
time of 444 ms. The intercept is responding to R
P
and
the gradient roughly to Ao.
13-46
V
P
vs V
S
Modelled Cube
(A + B) / 2 (A - B) / 2
V
P

(m/s)
2000
2500
1000 1250
V
S
(m/s)
1000 1250
V
S
(m/s)
This slide shows slices from the combined volumes at a
time of 444 ms. The sum is responding exactly to Ao
and the difference roughly to R
S
.
13-47
Model 2 - V
P
vs Thickness
In our second model, we varied P-wave velocity against
layer thickness, again on the middle layer of a 3-layer
model. Density was constant for all three layers, and
Poissons ratio was also held constant at 1/3.
We varied V
P
from 2000 to 2500 m/sec in increments of
50 m/s, and layer thickness from 5 to 55 m in
increments of 5 m, creating 121 modelled gathers in an
11x11 grid.
We used primaries-only modelling with a 25 Hz Ricker
wavelet, and included NMO stretch.
13-48
Parameters for Model 2
This slide shows the parameters in the top two layers
for the second model. Notice that by holding o and
constant, the intercept and gradient are simply the
negative of each other, and R
P
= R
S
.
Vp1 Vp2 Vs1 Vs2

1=

2
o
1=
o
2 A = Rp Rs B
2250 2000 1125 1000 2.2 0.333 -0.059 -0.059 0.059
2250 2050 1125 1025 2.2 0.333 -0.047 -0.047 0.047
2250 2100 1125 1050 2.2 0.333 -0.034 -0.034 0.034
2250 2150 1125 1075 2.2 0.333 -0.023 -0.023 0.023
2250 2200 1125 1100 2.2 0.333 -0.011 -0.011 0.011
2250 2250 1125 1125 2.2 0.333 0.000 0.000 0.000
2250 2300 1125 1150 2.2 0.333 0.011 0.011 -0.011
2250 2350 1125 1175 2.2 0.333 0.022 0.022 -0.022
2250 2400 1125 1200 2.2 0.333 0.032 0.032 -0.032
2250 2450 1125 1225 2.2 0.333 0.043 0.043 -0.043
2250 2500 1125 1250 2.2 0.333 0.053 0.053 -0.053
13-49
Model 2 - V
P
vs Thickness
This slide shows the stacked responses from the model.
The intercept and gradient were picked along Horizon 1.
13-50
V
P
vs Thickness - AVO Attributes
Intercept (A) Gradient (B)
V
P
(m/s)
2000 2500
T
h
i
c
k
n
e
s
s

(
m
)

5
55
V
P
(m/s)
2500 2000
This slide shows the intercept and gradient slices along
Horizon 1. Notice that tuning effects are different for the
intercept than they are for the gradient.
13-51
V
P
vs Thickness - AVO Attributes
(A + B) / 2 (A - B) / 2
V
P
(m/s)
2500 2000
V
P
(m/s)
2500 2000
T
h
i
c
k
n
e
s
s

(
m
)

5
55
This slide shows the sum and difference slices. The
estimate of R
S
((A-B)/2) is very close, but Ao ((A+B)/2) is in
error due to neglecting the third term and NMO stretch.
13-52
Model 3 - Porosity vs Water
Saturation
In our third model, we varied water saturation (from 0%
to 100% with a 10% increment) against porosity (from
23% to 33% with a 1% increment), in the middle layer
of a 3-layer model, with layer thickness of 50 m.
The Gassmann-Gregory approach was used for the
calculation of V
P
, V
S
, and density in the modelled layer.
A brief overview of this method will be given in the next
two slides, followed by plots of the parameters.
We used primaries-only modelling with a 25 Hz Ricker
wavelet, and included NMO stretch.
13-53
The Gassmann-Gregory Method

Velocity is a function of the saturated bulk modulus, K
sat
,
shear modulus,
sat
, and density,
sat
:
sat
sat sat
P
3
4
K
V

+
=
sat
sat
s
V

=
Density is a function of water saturation and porosity:
| | | ) S 1 ( S ) 1 (
w hc w w m sat
+ + =
dry sat
=
Shear modulus is assumed not to change with saturation:
13-54
Saturated bulk modulus is given by Gassmanns formula:
2
m
dry
m fl
2
m
dry
dry sat
K
K
K
1
K
)
K
K
1 (
K K

+ =
| |
hc
w
w
w
fl
K
S 1
K
S
K
1
+ =
The Gassmann-Gregory Method

where:
m dry P
K
1
K
1
K
=
|
Gregorys method assumes that we can determine K
dry
by
estimating the dry rock Poissons ratio o
dry
, and solving a
quadratic formula.
Finally, porosity effects are modelled by estimating the pore
bulk modulus K
P
:
13-55
P-wave velocity vs Porosity and S
W
(a) Contour plot of V
P
vs
porosity (|) and water
saturation (S
W
).
|
23%
33%
S
W

0% 100%
(b) Surface plot of V
P
vs
porosity (|) and water
saturation (S
W
).
S
W

|
13-56
S-wave velocity vs Porosity and S
W
(a) Contour plot of V
S
vs
porosity (|) and water
saturation (S
W
).
(b) Surface plot of V
S
vs
porosity (|) and water
saturation (S
W
).
|
S
W

|
23%
33%
S
W

0% 100%
13-57
Density vs Porosity and S
W
(a) Contour plot of vs
porosity (|) and water
saturation (S
W
).
(b) Surface plot of vs
porosity (|) and water
saturation (S
W
).
|
S
W

|
23%
33%
S
W

0% 100%
13-58
Poissons Ratio vs Porosity and S
W
(a) Contour plot of o vs
porosity (|) and water
saturation (S
W
).
(b) Surface plot of o vs
porosity (|) and water
saturation (S
W
).
|
S
W

|
23%
33%
S
W

0% 100%
13-59
Gathers from the Modelled Volume

(a) In-line 1,
S
W
= 0 %.
Notice the
negative
gradient on
the top event
(i.e. trough
increasing).
(b) In-line 11,
S
W
= 100 %.
Also notice
the negative
gradient on
the top event
(i.e. peak
decreasing).
|
33% 23%
13-60
Stack of the Modelled Volume

(a) In-line 1, S
W
= 0% (b) In-line 3, S
W
= 20% (c) In-line 5, S
W
= 40%
(d) In-line 7, S
W
= 60% (e) In-line 9, S
W
= 80% (f) In-line 11, S
W
= 100%
Here is every second in-line from the stacked volume. Notice the subtle
change in amplitudes except for the 100% saturated case, in which the
polarity changes. The AVO attribute slices are shown in the next 2 slides.
13-61
Porosity versus S
W
- AVO Attributes
Intercept (A) Gradient (B)
S
W

|
0% 100%
23%
S
W

0% 100%
33%
Here are the intercept and gradient attribute slices at 450 ms from
the modelled volume. Notice that there is no correlation between
the attributes and either porosity or water saturation, although
100% S
W
is anomalous on both.
13-62
(A + B) / 2 (A - B) / 2
S
W

0% 100%
S
W

0% 100%
|
23%
33%
Porosity versus S
W
- AVO Attributes
Here are the combined AVO attribute slices at 450 ms from the
modelled volume. There is no correlation for the sum with either
porosity or water saturation, but the difference shows some
correlation with porosity.
13-63
The next step...
So far, we have only been looking qualitatively at the
modelling results. This is useful from a learning
standpoint, but does not allow us to quantitatively
calibrate the model with the seismic data.
For the calibration, we used a statistical multi-
attribute procedure to derive relationships between
the processed AVO attributes and the original
modelled logs.
This was done on the third model example: porosity
vs water saturation, as it was the most difficult to
interpret from the attributes.
13-64
This involves closing the loop...
Processed
AVO
attribute
volumes
Modelled
log volume
AVO
synthetic
volume
Predicted
reservoir
parameter
volume
13-65
Multi-attribute Parameter Prediction
In the following slides, we will be using a multi-
attribute reservoir parameter prediction program
developed jointly by Mobil (now ExxonMobil) and
Hampson-Russell, called EMERGE.
The program is first trained at selected well locations
for the chosen parameter (either porosity or water
saturation, in this case).
The training involves finding weights for either multi-
linear prediction, or neural network prediction (using
either the MLFN or PNN method).
Cross-validation is used to determine how many
attributes should be used.
The selected weights and attributes are then applied
to the whole volume.
13-66
Multi-attribute Parameter Prediction
We distinguish between internal attributes derived
from the stack (such as amplitude envelope,
instantaneous phase, etc), and external attributes
derived from other software.
Our external attributes included: A,B, R
P
, R
S
, I
P
and
I
S
(inverted R
P
and R
S
).
S
W

100%
23%
0%
33%
|
The figure on the right
illustrates the positions
of the 5 wells used in
the training process.
28%
50%
13-67
Porosity Prediction - Validation
To the right is shown
the derived set of
5 attributes used to
predict porosity, which
included both P and S
impedance.
The figure to the right
shows the total error in
black and the validation
error in red, where the
target well is left out.
Note that only the first
three attributes show
decreasing error.
13-68
The predicted porosity logs using the MLFN neural
network. Since the algorithm was trained over the
zone of interest, it has done a good job there but not
outside of the zone.
Predicting the porosity logs
13-69
Slice through the predicted porosity
This is a slice through the MLFN predicted porosity. It has
done a reasonable job, but is still coupled to the strong
water saturation effect.
|
23%
33%
Side values
show true
porosity
Colour
bar shows
predicted
porosity
13-70
Saturation Prediction - Validation
The derived set of
5 attributes used to
predict porosity. The
first two attributes
are derived from the
gradient.
The figure to the right
shows the total error in
black and the validation
error in red. Again, only
the first three attributes
show decreasing
validation error.
13-71
The predicted water saturation logs using the MLFN
neural network. Again, the fit is much better through the
zone of interest, than outside this zone.
Predicting the water saturation logs
13-72
Predicted Water Saturation Slice
S
W

0% 100%
Here is the MLFN prediction result for the water saturation. The result
is better for water saturation than it was for porosity, probably because
of the strong water saturation effect on the gathers. It could be
improved with more wells.
Bottom
values
show true
water
saturation
Colour bar
shows
predicted
water
saturation
13-73
Case Study 2 - Conclusions
In this section, the AVO modelling volume technique
was applied to three examples: V
P
vs V
S
, V
P
vs layer
thickness, and porosity vs water saturation.
The first two examples gave us a good understanding
of AVO effects, but were straightforward. The third
example was more complex, and was used as an
input to a multi-attribute prediction algorithm.
By using 5 wells from the volume, we were able to
get a reasonable fit to both porosity and water
saturation, although it is felt that using more wells
could have produced better results.
13-74
Case Study 3 - Radial Basis Function
Neural Network
In this final case study, we will apply the radial basis
function neural network (RBFN) to the prediction of
log properties from seismic attributes.
We will compare this method with the generalized
regression neural network (GRNN).
We will first discuss the theory of both the GRNN
and RBFN, showing how the GRNN can be seen as
a simplified version of the RBFN.
We will then illustrate the behaviour of both the
RBFN and GRNN using the Blackfoot channel sand
dataset.
13-75
Historical notes
The first seismic application of the RBFN method was by
Ronen et al. (Seismic-guided estimation of log properties,
Parts 1, 2, and 3,TLE, 1994) in which the authors applied
the method to the prediction of reservoir parameter maps
using seismic attributes.
Hampson et al. (Use of multiattribute transforms to predict
log properties from seismic data: Geophysics, 2001) used
the generalized regression neural network (GRNN) for the
prediction of reservoir parameter volumes from seismic
attributes.
The motivation for this research came from trying to
understand the relationship between RBFN and GRNN,
and in applying RBFN for the creation of seismically-
derived reservoir volumes.
13-76


Log Seismic Attributes
s
i1 t
i
y
k
t
j
s
i2
s
j1
s
j2
x
k1
x
k2
x
k
s
j
s
i
?
Here is the basic concept using two attributes. We wish to find
a relationship between attribute vectors s
i
and s
j
and training
samples t
i
and t
j
and then apply this relationship to attribute
vectors x
k
to produce computed log values y
k
.
The Basic Concept
13-77
Basis functions
Both the GRNN and RBFN
use Gaussian basis functions
of distance in attribute space.
Using the three points from
the previous slide, we can
visualize the distances as
shown on the right.
Mathematically, the basis function is written:
parameter. smoothness a and
, d , d where
d
exp or
d
exp
j i ij j k kj
2
2
ij
ij
2
2
jk
jk
=
= =
(
(

=
(
(

=
o
o
|
o
|
s x s x
s
j

s
i

x
k

s
i2

s
j2

s
i1
s
j1
x
k1

x
k2

d
ij

d
ik

d
jk

Attribute 1
A
t
t
r
i
b
u
t
e

2

13-78
Theory of GRNN and RBFN
In the RBFN and GRNN, the computation of the predicted
values can be written in the following general way:
. M , , 2 , 1 k , w ) ( y
N
1 j
kj j k
= =

=
| x
For the RBFN, weights are pre-computed from the training
data using the following linear equations:
. N , , 2 , 1 i , w ) ( t
N
1 j
ij j i
= =

=
| s

=
=
N
1 j
kj j j
/ t w |
In the GRNN we compute the predicted values on the
fly from the training values, as follows:
13-79
Solving for the RBFN weights
. and ,
w
w
,
t
t
where
,
NN
N 1
1 N
11
N
1
N
1
(
(
(

=
(
(
(

=
(
(
(

=
=
|
|
|
|

u
u
w t
w t
Note that the RBFN equations can be written in
matrix form as:
| |
factor. ng prewhiteni a where
,
1
=
+ =

t w I u
This has the following solution:
13-80
Relationship between GRNN and
RBFN
Note that
matrix. inverted the of values row the are and
RBFN, the for t t t w
, GRNN the for / t w where
, M , , 2 , 1 k , w ) ( y
ji
N jN 2 2 j 1 1 j j
N
1 j
kj j j
N
1 j
kj j k


|
|
+ + + =
=
= =

=
=

x
After scaling, RBFN is thus equivalent to GRNN when:
i j , 0
ji
= =
In other words, u in the previous slide becomes the
identity matrix, which happens when either or
o approaches infinity.
, 0 s s
j i
~
13-81
Finding the seismic attributes
So far, we have discussed only the mathematics of the
GRNN and RBFN methods, but not the mechanics of
finding the optimum attributes to use as input.
We will perform the training using the well logs from the
wells that tie the seismic dataset.
We will then use the method described by Hampson et al.
(2001), which consists of the following two steps:
(1) Find the best set of attributes by a procedure called step-wise
regression, in which we compute combinations or attributes that
have the lowest least-squares error with the reservoir parameter
to be predicted. A convolution operator can also be used.
(2) Then find the L attributes that are statistically significant using
cross-validation, in which we successively leave out wells and
compute the least-squares errors for their prediction.
13-82
Optimizing the parameters
One final point concerns the optimization of the
parameters for the GRNN and RBFN methods.
For both methods, the key parameter is o, or the
smoothing parameter.
For GRNN, Masters (Advanced Algorithms for Neural Networks,
1995) has given an elegant method for letting sigma vary as a
function of attribute number.
For RBFN, computation of multiple sigmas is difficult because of
the matrix inversion step used in the optimization.
In the RBFN method, another key parameter is the
prewhitening . Although it has been shown theoretically
that there is no need for prewhitening, practical experience
has shown that a prewhitening of up to 10% are needed
for large numbers of training values.
13-83
Blackfoot channel sand case study
We will now compare the RBFN and GRNN methods using the
Blackfoot channel sand case study. The survey was recorded in
October, 1995, and the objective was a Glauconitic channel
within the Lower Cretaceous Mannville formation. The map
below shows the twelve wells in the area, and the seismic line
used in the study:

Inline 95
13-84
On the left is shown
inline 95 with the sonic
log from well 08-08
spliced in at its correct
location. Figure (a)
shows the stacked
seismic dataset, and
figure (b) shows a
model-based inversion.

The channel is clearly
indicated on the
inverted section.
(a)
(b)
Blackfoot channel sand case study
13-85
Case Study #1
RBFN was compared against GRNN using the
following parameters:
# of wells = 12,
# of attributes = 4,
operator length = 7.
Our objective will be to predict P-wave velocity by
training the attributes using the P-wave sonic in
each well. Note that the inverted impedance will
be one of our attributes.
We will first look at the results of the multiattribute
analysis described earlier.
We will then look at the comparison between the
RBFN and GRNN algorithms.
13-86
Results of the multiattribute
training, where (a) shows
the best 7 attributes, and (b)
shows the training with all
wells in black, and the
validation in red. Note from
the validation curve that only
the first four attributes will be
used by the RBFN and
GRNN algorithms.
(a)
(b)
Results of training
13-87
Application at wells
(a)
(b)
Results of the neural
network training at four of
the twelve wells, where
(a) shows the GRNN
algorithm, which has a
correlation coefficient =
0.87 and avg. error =
179 m/s, and (b) shows
the RBFN algorithm,
which has a correlation
coefficient = 0.82 and
avg. error = 196 m/s.

The results are virtually
identical, but GRNN has
a slight edge.
13-88
Validation at the wells
(a)
(b)
Results of the neural
network validation at four
of the twelve wells, where
(a) shows the GRNN
algorithm, which has a
correlation coefficient =
0.59 and avg. error = 273
m/s, and (b) shows the
RBFN algorithm, which
has a correlation
coefficient = 0.53 and
avg. error = 290 m/s.

For the validation, it
would appear that GRNN
is slightly better than
RBFN.
13-89
Application to the seismic
Results of the neural
network training at
inline 95 from the 3-D
survey, using 12 wells
and a 7 point operator
in the training, where
(a) shows the GRNN
algorithm, and (b)
shows the RBFN
algorithm.

Notice the extra
stratigraphic detail in
the RBFN result.
(a)
(b)
13-90
Case Study #2
RBFN was then compared against GRNN using
the following reduced set of parameters:
# of wells = 3,
# of attributes = 5,
operator length = 1.
Again, our objective will be to predict P-wave
velocity by training the attributes using the P-wave
sonic in each well. Also, the inverted impedance
will again be one of our attributes.
We will first look at the results of the multiattribute
analysis described earlier.
We will then look at the comparison between the
RBFN and GRNN algorithms.
13-91
Results of the multiattribute
training, where (a) shows
the best 5 attributes, and (b)
shows the training with all
wells in black, and the
validation in red. Note from
the validation curve that only
the first three attributes will
be used by the RBFN and
GRNN algorithms.
(a)
(b)
Results of training
13-92
Results of the neural
network training at all
three of the wells, where
(a) shows the GRNN
algorithm, which has a
correlation coefficient =
0.76 and avg. error = 234
m/s, and (b) shows the
RBFN algorithm, which
has a correlation
coefficient = 0.76 and
avg. error = 231 m/s.

For a small training set, it
appears that the RBFN is
now slightly superior to
the GRNN.
(a)
(b)
Application at wells
13-93
Validation at the wells
Results of the neural
network validation at all
three of the wells, where
(a) shows the GRNN
algorithm, which has a
correlation coefficient =
0.67 and avg. error =
264 m/s, and (b) shows
the RBFN algorithm,
which has a correlation
coefficient = 0.68 and
avg. error = 260 m/s.

Again, the RBFN is
slightly superior to the
GRNN.
(a)
(b)
13-94
Application to the seismic
(a)
(b)
Results of the neural
network training at
inline 95 from the 3-D
survey, using 3 wells
and a 1 point operator
in the training, where
(a) shows the GRNN
algorithm, and (b)
shows the RBFN
algorithm.

Notice the improved
stratigraphic detail in
the RBFN result over
the GRNN result.
13-95
Conclusions
In this study, we have compared the radial basis
function neural network (RBFN) to the generalized
regression neural network (GRNN) for the prediction of
reservoir parameters from seismic attributes.
For the 12 well case, the GRNN was slightly better
than the RBFN for training and validation at the wells,
but the application to the seismic was better using
RBFN.
For the 3 well case, the RBFN was slightly better than
the GRNN for training and validation at the wells, but
the application to the seismic was much better for
RBFN.
It would appear that the RBFN is the superior method
as the size of the training dataset decreases.
However, more comparisons need to be done.

You might also like