Stochastic vs Deterministic Pre-stack
Inversion Methods
Brian Russell
Introduction
Seismic reservoir analysis techniques utilize the fact that seismic amplitudes
contain information about the geological properties of the reservoir.
The mathematics behind this observation was developed in the early 1900s,
but its application to exploration seismic data did not start until the 1970s.
We classify these methods into two categories: methods that analyze only the
amplitudes, and methods that invert the amplitudes to reservoir properties.
Newer methods analyze pre-stack data, where the analysis of the amplitudes
without inversion is called Amplitude versus Offset, or AVO.
Pre-stack inversion has many forms, where the major division is between
deterministic and stochastic, or geostatistical, methods.
In this talk I will discuss these methods and look at their assumptions and
limitations.
A suggested workflow
Well Log Data Seismic Data
VP, VP, VS, Post-stack Gathers, only Gathers with
only offsets azimuths
Modeling for
VS Post-stack AVO & Pre- AVAz / Fracture
inversion stack Identification
inversion
Build rock physics
model
Integrate using multivariate or Bayesian statistics
Seismic Inversion Methods
Joint PP/PS
Model
Inversion
Based
Post-stack Recursive
Sparse 4D Inversion
spike
Inversion Colored
methods Azimuthal
Inversion
Elastic
Pre-stack Impedance
LMR Stochastic /
Geostatistical
Simultaneous Inversion
The basic model for inversion
The zero offset, or
stacked, seismic trace
can be modeled as
the convolution of the
acoustic impedance
(AI) reflectivity with
Wavelet
the wavelet.
As shown in the next W
slide, this is the basis
for post-stack Acoustic Reflectivity Seismic
inversion. Impedance
AI = VP RAI = AI W S = W * RAI
2 AI
Post-stack inversion
Post-stack seismic Inversion, developed in the 1970s, reverses the forward
modeling procedure, allowing us to derive the impedance from the reflectivity:
Inverse
Wavelet
Impedance Reflectivity Seismic
6
Qualitative AVO
In the 1980s, geophysicists observed that the amplitudes in a seismic gather
could be written in linearized form using the amplitude versus offset (AVO)
equation, a reformulation of the Aki-Richards linearized solution to the
Zoeppritz equations:
RP ( ) = RAI + G sin 2 + C tan 2 sin 2 , where :
2 2
VP VP VS VS VS VP
RAI = + , G= 4 2 , and C = .
2VP 2 2VP VP VS VP 2VP
Note that this has added two extra terms to the zero-offset case, a gradient
term G and a curvature term C, often referred to as A, B and C, where the
term A is called the intercept.
This formed the basis to what I refer to as qualitative AVO.
Intercept and gradient analysis
The amplitudes are extracted at all
times, two of which are shown here:
Offset or Angle
+RAI
+G
sin2
Time -G
-RAI
The AVO equation predicts a linear
relationship between these amplitudes
and sin2. Regression curves are
calculated to give RAI and G values for
each time sample.
Using the angle gathers for inversion
Fatti et al. (1994) re-formulated this equation to show that the pre-stack seismic
data is a function of the acoustic impedance reflectivity (RAI), shear impedance
reflectivity (RSI) and density reflectivity (RD) term:
RP ( ) = aRAI + bRSI + cRD ,
AI VS SI
where RAI = , RSI = + = , SI = VS , RD = ,
2 AI 2VS 2 2 SI 2
2 2
VS VS
a = 1 + tan , b = 8 sin , and c = 4 sin 2 tan 2 .
2 2
VP VP
Independent pre-stack inversion
Independent pre-stack inversion is Angle
implemented by first extracting the 1 N
reflectivity components, and then 600
inverting them separately. t
To estimate the reflectivities, the Time
650
amplitudes at each time t in an N- (ms)
trace angle gather are picked as Generalized inverse
shown here, to give RP(1) RP(N):
We can then solve for the RAI 1 RP (1 )
weight
reflectivities at each time sample
R SI = matrix
using least-squares inversion.
RD RP ( N )
Finally, these estimates are inverted
using a post-stack type scheme.
Reflectivities Observations
Simultaneous Pre-stack Seismic inversion
Pre-stack inversion is also based on an extension of the Fatti formulation
of the Aki-Richards equation:
S ( ) = c1W ( ) DLP + c2W ( ) DLS + c3W ( ) DLD , where :
S( ) = seismic trace at angle , LP , LS , LD = logarithms of Z P , Z P , and ,
W( ) = the extracted wavelet at angle , and D is the derivative operation.
As in our discussion of AVO and independent inversion, this can again
be set up as a least-squares problem:
model parameters = generalized inverse x observations
As we discussed earlier, there are two main types of pre-stack inversion,
deterministic and stochastic.
Deterministic vs Stochastic Inversion
First of all, let us define the fundamental difference between deterministic
and stochastic inversion:
In deterministic inversion we produce what we consider to be a single
best solution.
In stochastic inversion we produce many possible solutions, all
plausible, which average to the deterministic solution.
The advantage of deterministic inversion is that we get the best least-
squares solution to our problem.
The advantages of stochastic inversion are its higher frequency nature and
the calculation of uncertainty.
Deterministic pre-stack inversion example
On the next slide, I will show an example of deterministic pre-stack inversion.
A Gulf Coast dataset (shown on the left of the slide) was inverted for P-
impedance, S-impedance and density (which are shown on the right).
The inverted volumes were transformed to Vshale, porosity and Sw (also
shown on the right of the slide).
Our assumption is that each inverted or transformed result is the correct
answer.
However, this will not allow us to obtain uncertainty estimates from of the
rock properties.
Deterministic pre-stack inversion example
Seismic Inverted
Amplitude Acoustic
Map Impedance
Deterministic pre-stack inversion example
Seismic Inverted
Inverted
Amplitude Acoustic
Shear
Map Impedance
Impedance
Deterministic pre-stack inversion example
Seismic Inverted
Inverted
Amplitude Acoustic
Inverted
Shear
Map Impedance
Density
Impedance
Deterministic pre-stack inversion example
Derived
Vshale Map
Seismic Inverted
Inverted
Derived
Amplitude Acoustic
Inverted
Shear
Vshale
Map Impedance
Density
Impedance
Deterministic pre-stack inversion example
Derived
Vshale Map
Seismic Inverted
Inverted
Derived
Derived
Amplitude Acoustic
Inverted
Shear
Vshale
Porosity
Map Impedance
Density
Impedance
Deterministic pre-stack inversion example
Derived
Vshale Map
Seismic Inverted
Inverted
Derived
Derived
Derived
Amplitude Acoustic
Inverted
Shear
Vshale
Porosity
Sw
Map Impedance
Density
Impedance
Stochastic inversion
In stochastic inversion, the least-squares inversion method is extended by
formulating the problem using a Gaussian or Log Gaussian posterior probability
density function, or pdf (Tarantola, 1987).
This allows us to sample various scenarios from the pdf using the Monte Carlo
(MC) or Markov Chain Monte Carlo (MCMC) approach.
The earliest approach to stochastic inversion was by Haas and Dubrule, 1994,
in which Sequential Gaussian Simulation (SGS) is used.
Buland and Omre (2003) developed a fast approach to stochastic linearized
inversion which utilized a Gaussian pdf.
The GeoSI method that I will discuss today combines both a Gaussian pdf and
the SGS approach (Doyen, Williamson et al., 2007)
My colleague Ali Tehrani discussed the Jason StatMod approach yesterday.
Geostatistical inversion (Haas and Dubrulle)
Populate model with AI data at wells
Define a random path through all (x,y) trace
locations
(x,y)
Best simulated At each trace location perform a local
synthetic trace
optimization
Generate a large number of trial AI
sequences using SGS with spatial and
vertical variograms.
Compute reflectivity series and convolve
Actual
wavelet
with extracted wavelet.
Compute misfit against observed
seismic trace
seismic.
Retain best matching AI ( >0.8).
Go to next trace
Adapted from
Dubrule, 2003 AI simulations
Variogram models
Vertical (temporal) Horizontal variogram map Anisotropic variograms
variogram showing anisotropy in principal directions
Here are the variograms computed by Haas and Dubrulle (1994),
showing the vertical, or temporal change, and the horizontal change
including anisotropy.
Bayesian stochastic inversion
Although geostatistical stochastic inversion produces reasonable
results, it has two limitations:
It is quite slow.
It has difficulty in converging to an answer.
Buland and Omre (2003) introduced a new type of stochastic inversion
which was based on multivariate Bayesian statistics.
To illustrate the concept of Bayesian statistics, I will first consider the
standard least-squares regression problem.
We will then look at the general theory proposed by Buland and Omre.
We will then extend this method by combining it with SGS.
Least-squares regression
Consider a regression fit
to 16 measured porosity
values (i) plotted against
seismic impedance (zi),
shown by the red line in
the plot. This can be
written:
i = a + bzi
The regression line is the least-squares fit between porosity and impedance
and is considered the right answer, even though only one point falls on it.
Statistical interpretation
In the statistical
interpretation of this p()
crossplot, each variable Joint pdf p(,z)
(porosity and impedance)
is given as a Gaussian
probability distribution
function (pdf) defined by its z
mean () and variance (). p(z)
z
The joint pdf p(,z) is the probability of and z occurring, and is defined by
the variances and means, as well as the covariance between and z.
Bayesian regression
Bayesian statistics tells us
| z
that the conditional conditional pdf
probability of given z, or p( | z)
the posterior, equals the
| z
joint probability divided by
the probability of z, or the
prior.
p ( , z )
p ( | z ) =
p( z )
The conditional mean |z is the least-squares fit, and the conditional
variance | z gives us the scatter in this fit. Note it is narrower than p().
Bayesian stochastic inversion
Generalizing the previous example to inversion, Buland and Omre (2003)
showed that:
m|d = Cm|d (G T Cd1d + Cm1 m )
where:
m|d = conditional mean, Cd = data covariance ,
Cm = model covariance , m = model mean (prior),
and : Cm|d = (G T Cd1G + Cm1 ) 1 = conditional covariance.
This equation reduces to the least-squares solution if we assume that m
= 0, and Cd = d2I:
m = m|d = (G T G + d2Cm1 ) 1 G T d
GeoSI
The GeoSI method, as implemented by CGG and ported to the Hampson-
Russell suite of software, involves the following steps:
Build a stratigraphic grid using horizons, well logs and layer-based kriging.
Bring in partial angle stacks and wavelets.
Compute the Bayesian posterior distribution by combining the model,
seismic data and well logs.
Create multiple P and S-impedance realizations using the SGS technique.
Compute the mean and standard deviations from the impedance
realizations.
These steps are shown diagrammatically on the next two slides.
Building the stratigraphic grid
Stratigraphic grid
Ip
Horizons in time
Layer-based
Kriging
Is
Stratigraphic
layering style
Low-pass
filtering
Low-frequency prior model
in stratigraphic grid
Well logs in time
(Vp, Vs, Density) R. Moyen and J. Frelet
Stochastic Inversion Workflow
n Ip-Is
realisations
Partial angle stack seismic cubes
Well logs Bayesian
(Vp, Vs, Density) stochastic
Well uncertainty inversion
AI Horizontal
Ip-Is prior mean & & vertical
standard deviation variograms Posterior mean &
time in stratigraphic grid standard deviation Ip-Is
R. Moyen and J. Frelet
Bandwidth components
For all inversion methods, the prior model is constructed by
interpolating filtered logs, and controls low frequencies.
For both deterministic and stochastic inversion, the seismic
amplitudes control intermediate frequencies within the seismic
bandwidth.
In stochastic inversion, the vertical variogram model controls
the high frequencies.
Prior model Seismic Variogram
model
Spectrum
Power
Frequency (Hz)
Adapted from Moyen and Frelet
Offshore West Africa example
Elastic inversion (Ip-Is)
3 seismic angle stacks
16-30-40
120,000 traces
Time window of 200 ms
132 layers in grid
500 realisations (59 Gb total)
3 wells with Vp, Vs and density logs
Computations on standard workstation
Courtesy of R. Moyen and J. Frelet
Ip-Is Prior Model
P Impedance
ms
200
5200 7200
m/s x g/cm3
S Impedance
2200 4200
Courtesy of R. Moyen and J. Frelet m/s x g/cm3
Ip-Is Realisations
4
3
2
1
P Impedance
ms
200
5200 7200
m/s x g/cm3
S Impedance
2200 4200
Courtesy of R. Moyen and J. Frelet m/s x g/cm3
Ip Posterior Mean and Standard Deviation
P Impedance
mean
ms
200
5200 7200
m/s x g/cm3
P Impedance
std. dev.
Ip (km/s . g/cm3)
Courtesy of R. Moyen and J. Frelet 3.5 5.5
Vp/Vs Mean vs Realisations
1.5 Vp/Vs 2.5
Mean
Realisation
1.5 2.0 2.5
Posterior mean
Sand/shale
Cutoff
1.5 2.0 2.5
Realisation
Inversion Results Vp/Vs Ratio
Mean of 500
realisations
1.5 2.0 2.5
Vp/Vs Courtesy of R. Moyen and J. Frelet
Using the realizations
One of the key questions about stochastic inversion is: what do we do with all
the realizations?
In other words, wouldnt a single answer (i.e. the deterministic solution) be
better?
The answer is that with multiple realizations we can generate a number of new
results, such as:
Seismic lithology prediction.
Facies classification.
Volumetric uncertainty analysis.
Petrophysical property analysis.
These concepts are illustrated in the next few slides.
Stochastic Lithology Prediction
N sand / shale
simulations
Is
Ip
Sand probability cube
Histogram of sand volume
N realisations of Ip, Is
Courtesy of R. Moyen and J. Frelet
Facies Discrimination
0.5
Poissons
ratio
1.5 2.0 2.5 VSH
0 1
Ip/Is
P Impedance
Courtesy of R. Moyen and J. Frelet 0 3000 11000
Individual Realisations
Courtesy of R. Moyen and J. Frelet
Histogram of sand volume
For each realization, we can compute the sand volume from the number
of cells with sand.
This can then be arranged in histogram format, and the probability
percentiles can be computed.
A percentile is computed from the total area under the probability
curve.
Note that the percentile maps do not indicate a higher probability of
sand, only where the map falls within the distribution.
As shown by an earlier slide, the percentile values will in general be
larger than mean computation.
These concepts are illustrated in the next few slides.
Histogram of Sand Volume from Realizations
40 P50
Number of P90
realisations P10
0
Sand volume
P10 P50 P90
Ranked lithology simulations
Sand Volume from Realisations and Mean
40 P50
Sand volume from
Number of
inversion mean P10 P90
realisations
Sand
0 volume
Connected Sand Geo-bodies
Geobodies connected Color-code:
geobody volume
to at least one well (only largest are displayed)
Courtesy of R. Moyen and J. Frelet
Facies Probability from Stochastic Inversion
Sand probability
0.3 1
Large volume but
small probability Smaller volume but
Courtesy of R. Moyen and J. Frelet high probability
Stochastic Petrophysical Modelling
Multiple Ip & Is models Multiple Vsh and models
Ip
Statistical petro-elastic
calibration
Courtesy of R. Moyen and J. Frelet
Geostatistical inversion vs other modelling techniques
Geostatistical reservoir modeling
Interpolate between the wells
Plausible details
Accurate near wells
Not elsewhere
Deterministic seismic inversion
Optimize P-Impedance to minimize synthetic-to-seismic misfit
Accurate within seismic bandwidth
Unrealistically smooth
Only one possibility
StatMod/GeoSI geostatistical seismic inversion
Subsumes geostatistical modeling and deterministic inversion
Does both, simultaneously and in a statistically rigorous way
Multiple plausible realizations at high detail (e.g. 1ms 25m)
Yet also coherent interpretations of the seismic up to the km scale
Conclusions
Stochastic inversion is a natural extension of deterministic inversion
(mean of realizations deterministic inversion)
But it can provide extra information, such as:
Lithology probability
Facies distribution
Volumetrics
Petrophysical parameters
Our case study focussed on a channel sand play from West Africa.
Thank You
