AI Inversion Course 4
AI Inversion Course 4
AI Inversion Course 4
INVERSION
Introduction to Inversion
Earth Seismic
Input: Model Response
Modeling Inversion
Process: Algorithm Algorithm
Seismic Earth
Output:
Response Model
2
•Inversion is the reverse of the forward modeling process.
3
Seismic Inversion methods can be devided into the categories shown below:
4
5
•All inversion algorithms suffer from the “non-uniqueness” problem.
•This means that there is more than one possible geological model
consistent with the seismic data.
•This means that the final result always depends on the “other
information” as well as the seismic data.
6
The Convolutional Model
The seismic trace is modeled as follows:
9
The reflection coefficient for the ith interface is defined as:
V - Vi
i 1 i 1
ri
i
V Vi
i 1 i 1 i
where:
i density of layer i
Vi velocity of layer i
This equation is true only for vertical incidence rays. This means that
AVO effects are assumed to be negligible.
The forward model requires both velocity and density. Often, density is
not available. STRATA contains the option to calculate the density using
Gardner’s equation:
0.23 *V 0.25
When loading logs into STRATA, the coefficients of this equation may
be modified using the menu.
a *V b
The final inversion result from STRATA is actually a series of
impedance traces. To get a velocity result, STRATA uses the
generalized Gardner’s equation:
If a real density was used in the forward model, this procedure will use
the inverse formula shown above with a and b derived from the real
density and velocity.
12
INVERSION SCHEME
Seismic Data
Data Quality
Amplitude Velocity
Preserved Analysis
Accurate wavelet
14
Amplitude Preservation :
“True Amplitude Recovery” (~ Relative
amplitude is preserved) using
Deterministic amplitude corrections:
• Spherical divergence
• Absorption
• Transmission loss
15
Improvement of vertical resolution :
s(t) = w(t) * r(t) time domain convolutional model
S(f) = W(f) . R(f) frequency domain multiplication
16
Improvement of horizontal
resolution :
Migration :
• locating reflectors back to their “true” location.
• beaming up reflection amplitude back to
where it belongs.
17
Noise Attenuation :
• Random noise is reduced by the stacking
process.
• Coherent noise, such as multiples should be
eliminated: F-K filtering, Radon, Inverse Velocity
stacking.
18
Simplified
Pre-inversion
processing flow
19
Incorporating the low frequency
component :
20
21
Recursive Inversion
23
Issues in Recursive Inversion
The samples in the seismic trace are actually reflection coefficients. This means
that there is no seismic wavelet.
Since the equation is applied recursively from top to bottom of the trace, the effect
of errors is cumulative. This means that the error at the bottom may be much
greater than the error at the top.
The process is called bandlimited because the final impedance traces are defined
within the same frequency band as the input seismic data.
24
Model Based (Blocky) Inversion
Model based inversion follows from the convolutional model:
Assuming that:
The reflectivity is defined as that sequence which “fits” the data best.
That is, if we can find a reflectivity which convolves with the wavelet to
give a good approximation to the seismic trace, we assume that this is
the right answer.
The wavelet:
26
Step 1: Block the initial guess impedance with a uniform block size.
Step 2: Form a synthetic trace by convolving the blocky impedance with the
known wavelet:
Step 4: Modify both the amplitudes and thickness of the blocks to improve the fit:
28
(2) Non-uniqueness.
For a given wavelet, all these results fit the trace about equally well:
29
Sparse Spike Inversion
Sparse Spike Inversion assumes that the actual reflectivity can be
thought of as a series of large spikes embedded in a background of
small spikes:
Sparse Spike Inversion assumes that only the large spikes are
meaningful. It finds the location of the large spikes by examining the
seismic trace.
30
Sparse Spike Inversion builds up the reflectivity
sequence one spike at a time. Spikes are added until
the trace is modeled accurately enough:
Model-Based Inversion
32
The Initial Guess Model
The initial guess model for each trace consists of an impedance log, usually
derived by multiplying a real sonic log by a real density log. The impedance log
model must be measured in 2-way travel time. The original logs are measured
in depth. A critical step is depth-to-time conversion:
33
The depth-to-time conversion is made using a depth-time
table which maps each depth to the two-way travel time
from the datum (surface) to that depth and back:
34
The depth-time table is usually calculated from the sonic
log velocities using this equation:
i
dj
ti 2* where: ti = time down to layer i
j 1 Vj dj = thickness of layer j
V j = velocity of layer j
Note: The time to an event depends on all the velocities above that layer,
including the first velocity to the surface, V1. That velocity is unknown
and is usually approximated by extrapolating the first measured velocity
back to the surface:
35
If the well is deviated, it must be corrected to vertical
and the correction made from KB to datum:
38
Check Shot Corrections
The interpolation of points on the drift curve uses one of three options:
Linear: Honors the points exactly with straight line segments between
40
(2) Change the velocities for layers
between the first and last check shot
depth only.
41
(3) Do not change the velocities in the
sonic log.
42
Depending on the interpolation option used, the sonic log
changes may be drastic:
43
Log Correlation
Log correlation should be applied after the check shot correction, and is
ideally a small change.
Log correlation changes the depth-time curve in exactly the same way as
a check shot correction.
44
The log correlation
window looks like this:
45
The extracted wavelet will
look like this:
46
Now correlate the points on the log as shown here, and
then click on the Stretch button at the bottom of the log
correlation window.
47
The default parameters use
Spline interpolation between
points on the drift curve.
48
Change the Type of Interpolation to Linear and click on
Apply. Note the change in the shape of the drift curve.
49
Now change the Type of Interpolation to Polynomial with
the parameters shown below and click on Apply.
50
Change the menu as shown below and click on Apply. Note
that the option to Apply all changes adds a ramp to the top
of the sonic log.
51
Finally, change the menu as
shown below and click on
Apply. Then click on Ok on
the Check Shot window to
accept these parameters.
52
The log correlation
window now looks
like this. Note that
we have achieved
an 85% correlation.
To see the
parameters for this
calculation, click
on the Parameters
button.
53
The Initial Guess Model
Interpolating the Log
54
Picking a single event guides the interpolation of the log:
Note: A single picked event simply produces a bulk time shift on the log
for each trace. This is equivalent to applying a check shot correction
with a single point.
55
Picking a different event causes a different shift to be applied:
56
Picking two or more events is equivalent to applying a
variable check-shot at each trace. The impedances between the two
picked events are stretched / squeezed.
57
A pinch-out is handled by forcing the two picked events to converge:
58
Handling Faults
59
Interpolating the Log
60
Assume that there are two input logs, L1 and L2. We
wish to calculate the output log, Lout.
The weights vary inversely as the distance from the target point to each
of the input logs:
1 d1
2
w1
1 d 12 1 d 2 2
In general:
Lout wi * Li
i
where: -2
wi d i
d
-2
j
j
61
Using picked events with multiple logs forces the inverse
distance interpolation to be guided by the picked events:
62
Note the difference between interpolation with and
without picked events:
63
The next slide shows the same comparison for the
wedge model.
64
Original Model and Inversion Inversion with horizon removed
65
Wavelet Extraction
The Convolutional Model is used as
the basis for all inversion:
The intercept of the line is the constant phase rotation which best
characterizes this wavelet.
NMO stretch
Processing artifacts
STRATA assumes that the wavelet is constant with time and space:
Time invariant: This means that the inversion is optimized for
a limited time window.
Space invariant: This assumes that the data has been
processed optimally to remove spatial variations in the
wavelet.
69
In general, a variety of methods can be used for
wavelet extraction.
(1) Estimate amplitude spectrum using the the seismic data alone. The
phase is assumed known from some other source.
autocorrelation
maximum entropy spectral analysis
cross spectral analysis
in STRATA: statistical wavelet extraction uses the
autocorrelation
(2) Estimate both amplitude and phase spectra from the seismic data
alone.
marine signatures
VSP analysis
in STRATA: read in the external wavelet as an ASCII file
(4) Estimate both amplitude and phase spectra using both seismic
and well log measurements.
73
Model Based Inversion Parameters
The menu for model-based inversion:
74
Inversion option
where:
T = the seismic trace
W= the wavelet
r = the final reflectivity
M = the initial guess model impedance
H = the integration operator which convolves with the
final reflectivity to produce the final impedance
* = convolution 75
The objective function has two parts:
Minimizing the first part, (T - W*r), forces a solution which models the
seismic trace. Minimizing the second part, (M - H*r), forces a solution
which models the initial guess impedance using the specified block size.
This parameter controls the resolution of the final result. The initial
guess model is blocked to a series of uniform blocks with this size:
The final inversion result may change the size of the blocks, but the
number of blocks is still the same. This means that some blocks get
bigger and some get smaller, while the average is kept constant.
Using a small block size (2 ms) will increase the resolution, but the
increased detail may be coming from the initial guess.
Using a small block size will always improve the fit between the input
trace and the final synthetic trace. 78
Number of Iterations
There is never any harm in having more iterations - it only affects the
run-time.
79
Maximum-Likelihood Sparse Spike
Inversion Parameters
The menu for sparse spike inversion:
80
Maximum Number of Spikes
81
Bandlimited Inversion Parameters
The menu for bandlimited inversion:
82
The resulting initial guess model for various settings
of the Constraint High-Cut Frequency:
83
Scaling Parameters
After the main inversion menu is set, the following pages appear,
controlling the scaling of the data:
84
Why is Scaling an Issue?
Normally, when a wavelet is extracted, only its shape is known; not its
absolute amplitude. Inversion requires that the absolute amplitude be
known as well.
For either option, the window used to determine the scaler may be
different from that used in the actual inversion. For some data sets,
especially sparse models, the automatic scaling may not be ideal. In that
case, you may override with a manual adjustment, which multiplies the
automatic scaling result:
Scaling too
low:
Scaling too
high:
87
Error Plot
The Error Plot shows the difference between the actual traces and the
synthetic traces calculated using the inversion impedance result:
Ideally, the Error Plot should show no coherent energy, and should
have a low over-all amplitude.
88
Low frequency component in the error – probably
caused by using the wrong wavelet:
89
Error localized to one side of line – probably caused by
not picking enough events:
90
Coherent error throughout data set – probably caused by:
91
This is a summary of the three inversion tests. The
biggest improvement occurs when the block size is
reduced from 6 ms to 4 ms. Increasing the number of iterations also
has a lesser effect.
6 ms - 5 4 ms - 5
iterations iterations
4 ms - 10
iterations
92