Professional Documents
Culture Documents
Seismic Inversion
All Presentations & Materials are copyright of Piet Gerritsma 2006-20098 PetroEdge. All rights reserved
Frequency range of currently used acoustical imaging applications
Typical Scales of Reservoir Investigation
Simulation 40 - 200 m
Seismic 4 - 40 m
Logs 2 - 40 cm
Lab Fractions of cm
V 2000 m/s
= V T = e.g. = 50 m
f 40 Hz
Scales of Geological Reservoir Heterogeneity
Interwell
Well Area Well
Determined
Field Wide
From Well Logs,
Seismic Lines, 100's
Statistical m
Modeling,
etc.
1-10 km
Interwell
Reservoir 10's
Sandstone m
100's m
1-10's
Well-Bore
10-100's
m
10-100's mm
mm
Unaided Eye
Hand Lens or
Petrographic or Binocular Microscope
Scanning Electron
Microscope (modified from Weber, 1986)
Einstein: Make your model as simple as possible but no simpler.
t
Z Z Z
SIGNAL SHAPE:
w(t ) t
Thin bed configuration gives rise to a composite seismic response due to time overlap and
interference of the reflected seismic pulses.
FB June 2004, 47-70; Seismic inversion methods and some of their constraints Organised
Acoustic impedances and reflection coefficients and inversion
Discrete Inversion layer i
i +1Vi +1 iVi Z i +1 Z i i ,Vi
ri = = interface i
i +1Vi +1 + iVi Z i +1 + Z i i +1 , Vi +1
2 Z i +1 layer i + 1
1 + ri =
Z i +1 + Z i
2Z i
1 ri =
Z i +1 + Z i
Recursive Inversion:
Z i +1 1 + ri 1 + ri n 1
1 + ri Problems:
=
Z i 1 ri
Z i +1 = Z i
1 ri
Z = Z
n 1 1. Frequency bandlimiting
i =1 1 ri 2. Noise
Continuous Inversion
Z (t + dt ) Z (t ) 1 dZ (t ) 1 d [ln Z (t )]
r (t ) = = =
Z (t + dt ) + Z (t ) 2 Z (t ) 2
Note the time variable!
dt
t
Z (t ) = Z (0 ) exp 2 r ( )d valid if r (t ) 0.3
0
1 dZ (t )
t
r (t ) = Z (t ) = 2 Z (0 ) r ( )d valid if r (t ) 0.1
2 Z (t ) 0 Organised
Reflectivity section and inverted acoustic impedance section
ds (t )
j 2f S ( f )
s(t ) S ( f ) t dt
s( )d S( f )
1
j 2f
Refl. AI Seismic section P-Impedance section
3500
2,1 107
Comparison of seismic data and inverted P-Impedance for the target interval.
FB May 2011, 103-114 Organised
TLE March 2000,
242-256
Fig. 4. Interpretation is made easier by reference to the impedance inversion model. This is clearly
demonstrated with this example of a "hidden" channel. A seismic section is shown in wiggle-trace format.
The yellow event on the left and right is interpreted based on well control and is a known unconformity. A
number of places exist where the yellow horizon could drop down onto a lower event in the attempt to tie
the horizons. The correct answer appears to be a matter of interpretative judgment and knowledge of the
regional play concepts. Organised
TLE March 2000,
242-256
Fig. 5. The interpretation problem is resolved by looking at the acoustic impedance inversion section. By
following the top of the high impedance layer (red/yellow), there is no conflict between the geologic well
picks. There is a low acoustic impedance channel that has subsequently been interpreted as an incised
valley. Organised
TLE March 2000, 242-256
Fig. 6. The completed interpretation from Figure 5 has been transferred back to the seismic data. The
interpretation from the seismic data alone was clearly problematic. The change in layer impedances
allowed the inverted data to image the unconformity clearly.
Organised
Reflectivity
Acoustic impedance
Fig. 1. The objective is to interpret the yellow reservoir level horizon. The location is known around well
control. Clearly, interpreting this event between well locations is quite difficult from seismic data (top)
due to cycle skipping. In contrast, on the equivalent impedance section (bottom) interpreting the event is
straightforward.
TLE August 2000, 878-881; Quantitative Reservoir Characterisation by v. Riel Organised
Fig. 1. Some of the many advantages of acoustic
impedance over seismic data are illustrated in this
figure. (a) A simple model of a low-impedance
wedge in a high-impedance background. (b) The
synthetic seismic data generated by convolution of a
Ricker wavelet with the reflection coefficients from
this model are shown with the traditional wiggle
trace and (c) as color amplitude with wiggles
overlain. From the seismic data it would be simple,
in this case, to interpret the general structure of the
model. However, because of the effects of the side
lobes of the wavelet and the effects of tuning, it is
difficult to know whether there are any internal
structure or lateral variations in the properties of the
wedge. (d) The inversion of the seismic data. It is
now a simple matter to interpret the boundaries of
the wedge. It is also possible to examine the internal
structure of the wedge in terms of absolute physical
properties. Even though in the real world the
situation is usually more complicated than this
simple wedge, analogous interpretative advantages
may be achieved through acoustic impedance
inversion.
TLE March 2000, 242-256 Organised
TLE March 2000, 242-256
Fig. 2. Impedance inversion models contain more information than seismic data because they have a
broader frequency content. (b) Synthetic seismic data set based on the acoustic model in (d) and the
wavelet in (a). There are three transitions50 ms, 135 ms, and 230 ms. Each interface represents the same
change in absolute AI units but in varying gradational degrees, representing varying dominant frequencies.
The synthetic seismic identifies the sharp interface at the top 50 ms (80 Hz), sees the top of the event at 135
ms (15 Hz), but it is not clear that it is a gradational coarsening upward sequence and fails to recognize the
most gradual interface at 230 ms (6 Hz). Compare the seismic responses to that of the inverted traces in (c).
The inverted traces are shown in color with a black overlay. There is a significant difference in the
properties of the rocks at 150 ms and 230 ms. These differences are not clear from the synthetic seismic
data because the low-frequency information is missing. On the other hand the impedance inversion model
contains this necessary information. Organised
TLE March 2000, 242-256
Fig. 3. The band-limited nature of seismic data is often considered in terms of the high frequencies and
consequent lack of resolution. However, the low frequencies missing from the seismic data are extremely
important if quantitative interpretation is required. This is illustrated by filtering a simple impedance layer
model to three different frequency ranges (a) 1080 Hz, (b) 10500 Hz, and (c) 080 Hz. The inclusion of
the high frequencies (b) allows us to interpret the location of the layer boundaries more accurately, but it is
the inclusion of the low frequencies (c) that allows us to obtain absolute values for use in the quantitative
interpretation of the rock properties.
Organised
Broadband seismic the importance of low frequencies
Butterworth band-pass
filter responses with
different low-cut fre-
quencies.
The increasing low-cut
frequency results in
increasing sidelobe
amplitudes.
Figure 1.
High-cut 32 Hz and 12 dB/octave low cut at 2 (green), 4 (red), 8 (blue), and 16 Hz (yellow).
Geophysics March- April, 2013, WA3-WA14; Broadband seismic data the importance of low frequencies, by ten Kroode et al
Broadband seismic the importance of low frequencies
Higher frequencies sharper wavelet (central peak)
Lower frequencies reduced side lobes
Lower frequencies:
1. Higher-resolution seismic images
2. Better penetration
3. Better suitability for seismic inversion
Lower frequencies:
1. Extending the bandwidth of the recorded seismic signal at the low end of the spectrum
leads to less sidelobes on the wavelet.
2. Low frequencies scatter less and suffer less from absorption in the earth, so they penetrate
deeper.
3. At low frequencies, the number of local minima in least-squares misfit functions used in
FWI is smaller, so we have a better chance to converge to the true model.
Low frequencies enlarge the basin of attraction for FWI.
4. A richer low-frequency content may eliminate the need for well-derived velocities to fill
the traditional bandgap in impedance inversion.
Geophysics March- April, 2013, WA3-WA14; Broadband seismic data the importance of low frequencies, by ten Kroode et al
Broadband seismic the importance of low frequencies
s(t ) S ( f )
ds(t )
j 2f S ( f )
dt
s( )d j 2f S ( f )
t 1
Fig. 4
Wedge model showing the interference of a positive and equal strength negative reflection at various
time intervals. (a) A 1-45-Hz Butterworth filter with a 2.5 dB/octave emphasis on the low frequencies
is used as wavelet. The time interval in ms is indicated by the label. (b) Same as (a), but with a 7-Hz
low cut and an additional 6dB/octave emphasis of the low frequencies. (c) Same as (b), but with a 1-Hz
low cut on the Butterworth filter.
Geophysics March- April, 2013, WA3-WA14; Broadband seismic data the importance of low frequencies, by ten Kroode et al
Broadband seismic the importance of low frequencies
Note
Amplitude and phase fidelity
at the low frequencies are im-
portant to achieve an accurate
interpretation. In our experi-
ence, phase fidelity of the
low frequencies is generally
poor in real data and needs
further improvement.
Note
Emphasizing the low fre-
quencies present in the
data can only be done if
the S/N is good enough
at the low end.
Figure 8.
The low-impedance formation (red) is obscured by the side lobe of a high-impedance layer (blue) in 2D
nonbroadband seismic data (top left). 2D broadband data successfully resolve the low-impedance formation
(top right). Synthetic using nonbroadband seismic wavelet (bottom left) and synthetic using a broadband
wavelet (bottom right) confirms the preceding interpretation.
Geophysics March- April, 2013, WA3-WA14; Broadband seismic data the importance of low frequencies, by ten Kroode et al
Broadband seismic the importance of low frequencies
1.5 5 Hz
2 5 Hz
4.5 5 Hz
Fig.19
Velocity models obtained from FWI in the v0+kz starting model and using frequency bands 1.5-5 (top),
2-5 (middle), and 4.5-5 Hz (bottom).
Geophysics March- April, 2013, WA3-WA14; Broadband seismic data the importance of low frequencies, by ten Kroode et al
Broadband seismic the importance of low frequencies
Inversion for velocity and impedance
Impedance inversion is usually performed after imaging as the first step toward
quantitative interpretation. Typical marine seismic data have frequencies starting
at 5-8 Hz, whereas the model obtained from seismic reflection tomography has
frequencies up to 2 Hz. The resulting bandgap is usually filled in by using well
information: low-passed impedance logs between well locations are interpolated
using horizons interpreted on migrated seismic as a guide. If wells are sparse
and/or the subsurface complex, such low-frequency models will, however, be
inaccurate. It would obviously be preferable to not have to rely on well-based
information and to have good S/N on the frequencies down to ~ 2 Hz in the data.
Geophysics March- April, 2013, WA3-WA14; Broadband seismic data the importance of low frequencies, by ten Kroode et al
Signal resolution versus noise boosting
f
Area of noise boosting
Organised
The impact of missing information and noise on inversion
This illustration shows the reconstruction of the earths spectrum by reservoir property inversion. The
extent of the earths spectrum that can be recovered (blue curve) is limited by the frequency range in
which the seismic amplitude (green) is above the noise amplitude (red). The zero-frequency information
can be estimated from imaging velocities and the higher-frequency information can be obtained from
the seismic reflection data. There is a gap (purple) in the seismic information at the lower frequencies.
Obtaining low-frequency information is particularly challenging because of increased noise at low
frequencies.
TLE Feb 2008, 163-165 Organised
The role of a smoothly varying background model
The P-wave velocity profile ( z ) has been converted to vertical two-way travel time and
the Fourier spectrum ~ ( f ) is calculated. On the same frequency scale the bandwidth of
the source pulse (blue bar) and the bandwidth of the two-way time converted background
velocity model (red bar) are indicated.
Organised
The challenge to obtain full-bandwidth acoustic impedance volumes
Multiple data sets are used to build full-bandwidth acoustic impedance volumes. The data contain different
useful frequency bandwidths. By combining the seismic processing, log, and inverted seismic data together,
a full-bandwidth impedance volume can be created. The inversion of the seismic data alone will create a
band-limited impedance volume that does not contain any low frequencies.
TLE Jan 2007, 74-87; Uncertainties in low ac. imp. models. Organised
Combining FWI and WE-MVA
The sequential imaging process is slowly being undermined by three long-standing trends in the industry:
1. Acquisition of lower-frequency data
2. Imaging under complex overburden which requires higher-resolution velocity models to focus and
correctly position reflectors
3. Acquisition of longer-offset data.
Simplified 1D graphical representation of separation of scales in seismic imaging (black line) and how
current industry trends are narrowing the gap between the estimation of long wavelengths and short
wavelengths (blue and green lines).
TlE Sep 2013, Special Section: FWI; 1074-1080, Tomographic FWI (TFWI) by combining FWI and WEMVA, by Biondi and Almomin.
Broadband seismic for Reservoir Characterization
The model illustrates that broadband bandwidth extension results in impedance response close to the
original acoustic model.
Wedge model showing seismic responses for wavelets of different bandwidth. The four top right images
illustrate the seismic responses for the various wavelets and the four bottom images illustrate the impe-
dance inversions based on the modeled seismic. The color palette for the impedance sections are identi-
cal. Each wedge tapers from 30 ms at the edge.
FB Sep 2012, 68-75
Acoustic impedance from Broadband seismic (only)
This approach works based on the fact that the seismic data is very rich on the low frequencies.
Comparison of two acoustic impedance results based on pre-stack inversion using two different initial low
frequency models (LFM), dependent on the bandwidth of the seismic data input. Top figure corresponds
to the conventional acquisition results and the bottom corresponds to the dual-sensor streamer inversion
results. FB Sep 2012, 68-75
The challenge to obtain full-bandwidth acoustic impedance volumes
Seismic processing velocities along with well log data are used in various combinations to
build low-frequency impedance models. Different types of processing velocities may be
used, such as stacking, time migration, depth migration, and tomographic velocities.
However, processing velocities are not necessarily true rock seismic velocities. The
velocities are generally calibrated to well control through seismic-to-synthetic ties so that
accurate gross time-depth relations exist in the calibrated velocity models. The calibrated
velocities are then used to build the low-frequency impedance model, although they
generally only span a limited portion (e.g., 02 Hz) of the missing low-frequency spectrum
(e.g., 05 Hz). The remaining portion will be filled in with a model.
TLE Jan 2007, 74-87; Uncertainties in low ac. imp. models. Organised
Methods for the recording of low-frequency seismic
1. From a filtered log after the conversion from depth to two-way traveltime ~ stretching
to fit the seismic data in the vicinity of the well.
Organised
Frequency components of a sonic log
Organised
Reflection coefficients from logs averaging interval
The logging tool response itself applies a high-cut filter. For a receiver spacing of 0.61 m, the transfer
function of the tool can be approximated by a sinc function with the first zero-crosssing at 1.64/m.
(Note that this is approximately the Nyquist value for sampling the log at regular 0.3 m intervals.)
This sinc function has already fallen to 90% of its peak value at 0.41/m, and for a velocity of 2000m/s
this corresponds to 820 Hz, a value well within the Nyquist range for 0.5 ms block-averaging. By
taking as fine a block-averaging interval as 1 ms twt, and calculating reflection coefficients at this
rate (followed by anti-aliasing filtering and thinning to the required rate) aliasing effects are kept
relatively low. Since the process of calculating reflection coefficients is non-linear, it is difficult to
predict aliasing, but experiments on three logs suggest a noticeable effect, although it is difficult to
separate this from inaccuracies caused by the block-averaging per se. Due to the action of the logging
tool, it is pointless to reduce the block-averaging interval much below 1 ms twt.
Choosing the averaging interval when calculating primary reflection coefficients from well logs, by Walden and Hosken; GPr 1988, 799-824
Acoustic impedance from reflection seismograms
N
x(t ) = r (t ) w(t ) r (t ) = rk (t k )
k =1
k = k vk
k 1+ r
k +1 k 1 + rk
rk = k +1 = k = 1 j
k +1 + k
1 rk j =1 1 r j
Geophysics Oct. 1983, 1318-1337; Recovery of the acoustic impedance from reflection seismigrams, by Oldenburg, Scheure and Levy.
Acoustic impedance from reflection seismograms
1 d [ln (t )]
r (t ) =
2 dt
(t ) = (0 ) exp 2 r ( )d
if r (t ) 0.3
t
0
where (0 ) = 1 is the surface impedance.
With the exception of a possible bias, this equation shows that the logarithms of the
normalized acoustic impedance estimated from averages of the reflectivity function
are in fact averages of the true logarithm of the impedance (if R0 = 0 ). Low- and high-
frequency information about the acosutic impedance is not directly derivable from a
band-limited seismogram.
Geophysics Oct. 1983, 1318-1337; Recovery of the acoustic impedance from reflection seismigrams, by Oldenburg, Scheure and Levy.
Acoustic impedance from reflection seismograms
w0(t) W0(f).R(f)
w0(t)*r(t)
(a) The acoustic impedance, (b) the reflection coefficients, (c) the amplitude spectrum of the reflection
coefficients, (d) the seismic wavelet, (e) seismogram, (f) the amplitude spectrum of the wavelet, (g) ave-
raging function (or interpreters wavelet) from a high-resolution appraisal deconvolution, (h) the decon-
volved seismogram, (i) amplitude spectrum of (h), and (j) estimated acoustic impedance.
Geophysics Oct. 1983, 1318-1337; Recovery of the acoustic impedance from reflection seismigrams, by Oldenburg, Scheure and Levy.
Hadamard: A mathematical well-posed problem should have properties were:
1. A solution exists
www.dugeo.com Organised
www.dugeo.com
QI workflow for exploration and appraisal (2B)
Seismic Well logs Geological
prior
Wavelet with
uncertainty Depth dependent
elastic rock physics
Simultaneous model with
SPIKE uncertainty
inversion Low frequency model
Depth dependent
multivariate rock
Rock property property PDFs
volumes uncertainty
QI workflow for exploration and appraisal. The comprehensive workflow assimilates geology, geophysics, petrophysics,
rock physics and uncertainty. Simultaneous inversion products are quantitatively integrated with statistical rock physics
models to make probabilistic, volume based fluid and lithology predictions using a Bayesian framework. Interpretation
of the probability volume outputs can be easily digested by the decision makers, maximising business impact.
Seismic Inversion Techniques
Signal: S ( ) = AS ( )e j S ( )
Noise: N ( ) = AN ( )e j N ( )
Filter: F ( ) = AF ( )e j F ( )
Desired output: D( ) = AD ( )e j D ( )
Minimize: E{F ( )} = [D ( ) F ( )( S ( ) + N ( ))] [D ( ) F ( )(S ( ) + N ( ))]
d
The filter that minimizes this expression has amplitude and phase spectrum:
F ( ) = D ( ) S ( )
AS ( )AD ( )
AF ( ) =
AS2 ( ) + AN2 ( )
Least-squares filtering in the frequency domain with different objectives
AS ( )AD ( )
F ( ) = AF ( )e j F ( ) with AF ( ) = and F ( ) = D ( ) S ( )
AS2 ( ) + AN2 ( )
input data: X ( ) = S ( ) + N ( ) = W ( ) R( ) + N ( )
desired output: D( ) = 1 or D( ) = R( )
AS ( ) AW ( )
AF ( ) = AF ( ) =
lls inverse filter:
AS2 ( ) + AN2 ( ) or AW2 ( ) + AN2 ( ) AR2 ( )
F ( ) = S ( ) F ( ) = W ( )
Linear least-squares reflectivity estimation in the presence of noise
s( t ) = w ( t ) r ( t ) + n ( t )
Both r (t ) and n(t ) are assumed to be random, stationary and independent of each other.
n(t ) is normal and zero mean.
The linear least-squares sense estimation of r (t ) solved by means of the Wiener deconvolution
filter has spectral characteristics:
W ( )
H ( ) =
W + Rn ( )
( ) Rr ( )
2
This filter ensures the estimation of r (t ) with minimum variance that can be expressed as:
Rr ( ) Rn ( )
= 2
2
d
W ( ) Rr ( ) + Rn ( )
r 2
0
EAGE 2000, B-37
Estimation of the Signal Spectrum and Noise Spectrum
trace i trace j CMP
seismic section
t
xi (t ) = s (t ) + ni (t ) x j (t ) = s (t ) + n j (t )
2. Calculate the autocorrelation of the traces: Rxi xi (t ) = Rss (t ) + Rsni (t ) + Rni s (t ) + Rni ni (t )
3. Calculate the crosscorrelation of pairs of traces: Rxi x j (t ) = Rss (t ) + Rsni (t ) + Rn j s (t ) + Rni n j (t )
{
{
}
E Rxi xi (t ) = Rss (t ) + Rnn (t )
}
E Rxi x j (t ) = Rss (t ) } Rss (t ) and Rnn (t )
Linear least-squares reflectivity estimation in the presence of noise
The model:
R ( f ) = X(f )
1
W ( f ) S( f ) + N( f )
2 2
or
W ( f )
R ( f ) = X ( f ) = zero-phase deconvolved
W ( f ) + N( f ) R( f )
2 2 2 band-limited reflectivity
An empirical observation tells us that all earth reflection coefficient series have spectra
that exhibit a similar trend that can be described as f . The term is a positive
constant. may vary from one field to another but tends to remain reasonably constant
within any one field.
Relative Impedance and Coloured Inversion Impedance
Wavelet: w(t )
R( f ) f f
Well/Seismic Tie
Reflectivity spectrum: AI-spectrum
{
Phase rotation
Relative Impedance ~ 1. Unconstrained transform Trace integration
Recursive inversion
Original Seismic
Zero-phase deconvolved
Relative Impedance and Coloured Inversion Impedance
+1.0
-1.0
Relative Impedance
+1.0
-1.0
Coloured Inversion
SEG 2000, 1572-1575, Fast-track coloured inversion.
Least-squares filtering: Wiener filtering
input filter actual output
at ft ct = f t at
_
+ et = d t ct
error
dt
desired output
0 at n
m 0
ft
0 ct = f t at n+m
0 dt n+m
filter: f t = f 0 ,L, f m m
ft 0
0 ct = f t at n+m
actual output: ct = c0 , L, cn + m
dt n+m
desired output: d t = d 0 , L , d n + m 0
m+ n m+ n m+ n 2
m
Objective is the minimization of E = E ( f 0 ,L, f m ) = et2 = (d t ct ) = d t f s at s
2
t =0 t =0 t =0 s =0
Minimization of E leads to the normal equations:
E m+ n m+n
= 2 d t f s at s ( at j ) = 2 et ( at j ) = 0 for
m
j = 0, L , m
f j t =0 s =0 t =0
orthogonal
This set of equations can be written as:
autocorrelation of input
r0 r1 L rm f 0 g 0 m+ n
r r O M f g r r rj s = at s at j = rs j
1 0 1 = 1 or [R] f = g with t =0
M O O r1 M M m+ n
g j = d t at j
rm L r1 r0 f m g mr t =0
T
r
A A f AT d crosscorr. of input and desired output
Convolution in matrix notation
at = (a0 , a1 ,L , an )
bt = (b0 , b1 ,L , bm )
m n
ci = ai bi = ai k bk = ak bi k
k =0 k =0
(m + 1) columns (n + 1) columns
a0 0 0 a0b0 b0 0 0 b0 a0
a b
a0 a b + a b 1 b0 b a + b a
1 1 0 0 1 1 0 0 1
O O b0 O O a0
O O 0 b1 O O 0 a1
an O a0 = anb0 + L + a0bn bm O b0 = bm a0 + L + b0 am
0 an a1 0 bm b1
b a
O
n
O m
O anbm 1 + an 1bm O bm an 1 + bm 1an
0
0
0 an a b
n m 0 bm b a
m n
(n + m + 1, m + 1)
r r(n + m + 1) (m + n + 1, n + 1)
r r
(m + n + 1)
[A].b = c [B].a = c
Organised
Linear least-squares estimation expression for the error
r
d r r r
e = d Af
r r
c = Af
( ) ( )
rT r r r T r r rT r rT r rT r rT r rT r rT r
E = e e = d c d c = d d d c c d + c c = d d c c
c
r r r r r
c = Af e = d c
Least-squares filtering/estimation
input: at = a0 , L, an
filter: f t = f 0 ,L, f m
actual output: ct = c0 , L, cn + m
desired output: d t = d 0 , L , d n + m
m+ n m+ n m+n 2
m
Objective is the minimization of: E = E ( f 0 ,L, f m ) = e = (d t ct ) = d t f s at s
2 2
t
t =0 t =0 t =0 s =0
r rT r r rT r r
E ( f ) = e e = (d [ A] f ) (d [ A] f )
(m + 1) columns
a0 0 L 0
a a O
The solution of this minimization can be formulated as: 1 0
O 0
a0
r
( T 1 T
)r
f = [A] [A] [A] d with [A] =
n
a a n 1
0 an O
[A]T [A] = [R] is the autocorrelation of the input signal M O O a n 1
r 0 L 0 a n
[A] d is the crosscorrelation of the input signal and the desired output signal
T
(Weighted) linear least mean square (llms) filtering
n+m n+m 2
m
n+m 2
llms filtering E ( f 0 ,L, f m ) = [d t ct ] = d t f s at s = et
2
t =0 t =0 s =0 t =0
( ) ( )
r rT r r rT r
E ( f 0 , L, f m ) = d [A] f d [ A] f = e e
minimization of E leads r r
to the normal equations: [A] [A] f = [A] d
T T
( )
r r
f = [A] [ A] [ A] d
T 1 T
equation for the filter:
t =0 t =0 s =0 t =0
( ) ( )
r rT r r rT r
E = d [A] f [W ] d [ A] f = e [W ] e
minimization of E leads r r
to the normal equations: [A] [W ][A] f = [A] [W ] d
T T
( )
r r
f = [ A] [W ][A] [A] [W ] d
T 1 T
equation for the filter:
Singular Value Decomposition: SVD Aki and Richards
r r 11 1m 1
P. 677-699
Axr = dr 1
UV T x = d r r
A . x=d
A T A = V 2V T m
n
( A A)
T 1
= V 2V T n1 nm
( AA )uri = i2 uri
data space
T 11 1m 1
11 1n 11
1
( A T A )vri = i2 vri
U
U
.
V T . xr = dr
model space
mm m
with i = 1, L , p n1 nn n
n1 nm
p +1 = L = m = 0
data eigenvectors model eigenvectors
U T U = UU T
= I n ; V T V = VV T
= Im
r r
p 0 V pT =
= (U p , U 0 )
U u , L , u (n, p )
A = UV T r1 r
T p p
= L (n , n p )
0 V0 U 0 u p +1 , , u n
(n, m ) 0 r
V p = v1 , L , v p
r
(m , p )
((n, p ), (n, n p )) (n, m ) r r
V 0 = v p +1 , L , v m (m , m p )
A = U p pV pT
Singular Value Decomposition: SVD
r r
Axr = dr
UV T x = d
The right singular matrix V maps the physical parameters, as defined for a specific problem,
to the active parameters that can be estimated effectively (i.e. independently). The active
parameters are a weighted combination of the physical parameters, where the weights are
r
determined by the columns of V .
We can rewrite the above equation U d = V x
T
[ ]
T r
By using the active parameter interpretation of the columns of the right singular matrix V ,
the left singular matrix U maps the physical data samples, that comprise the seismic input,
contribute to the solution. Thersingular matrix
( Tr
) ( )
to the active data samples that effectively
T
relates the active parameter space V x to the active data space U d .
This is consistent with the interpretation of the singular values in terms of resolution (e.g. an
active parameter corresponding with a large singular value = well resolved from the data).
r r r
2 xix j
e = d A x The Jacobian matrix of first derivatives is J = A
The boundary for the -difference region is found as
r r r
2 = (L ( x ) L ( x )) = ( x x ) H ( x x )
T
8 20 0.40 0.91
AT A = has eigenvalues 1, 2 29.883,0.117 and eigenvectors
5 13 0.91 0.40
A = U V T
A = Up p V pT
UU T = U T U = I 4 VV T = V T V = I 2
AAT = UV T VU T = U2U T
AT A = VU T UV T = V2V T Organised
Aki &Richards
Singular Value Decomposition: SVD - Example P. 682,683
1 1 0
p 0V pT
A = 0 0 1 ( )
A = U p ,U 0
0V0T
= U p V
p p
T
0 0 1 0
r 2r data space model space
AA ui = i ui
T
r r
AT Avi = 2i vi
2 0 0 2 2 0 0
AAT = 0 1 1 0 1 2 ( )(
1 = 2 2 2 2 2 = 0) 12 = 22 = 2 ; 3 = 0
2
0 1 1 0 1 1 2
r
1 1 0 0 1 0 2 0 0 0 0 1 v1
r
A = 0 0 1 = 1 2 0,1 2 0 2 0 1 2 1 2 0 v2
r
0 0 1 1 2 0 1 2 0 0 0 1 2 1 2 0 v3
A = UV T r r r
u1 u2 u3
A = U p pV pT
1 1 0 0 1
2 0 0 0 1
A = 0 0 1 = 1 2 0
2 1 2 1 2 0
0 0 1 1 2 0
0
Organised
Least-squares solution expressed in terms of SVD
1 T r
( ) ( )
r rT r
x = (A A) A d = AL d
r 1 r r
Ax d Ax d
T
minimizes
with A = U V
T
A 1
L = ( A A) A = V U
T 1 T 1 T
or, instead [(
V + I
2
)
1
]
U T
r 1
r 1 T
r
x = AL d = V U d
or
( )
r 1 r rT r 1 r rT r
( )
x = v1 u1 d + v2 u2 d + L +
1 2
1 r rT r
vm u m d
m
( )
r 1 r 2 r m r
x = v1 + v2 + L + vm
1 2 m
r
x=
+
2
1
( )
r rT r
v1 u1 d +
2
+
2
( )
r rT r
v2 u 2 d + L +
+
m
2
( )
r rT r
vm u m d
1 2 m
Least-squares solution expressed in terms of SVD
r r
Ax = d
with A = U V T = U p p V pT
1
A = A A
L ( T
)
1
AT = V1U T = V p p1U Tp
1. Unconstrained least-squares
1 T r r r
r r
min A x d
2 r
( )1 1 T
x = A A A .d = AL .d = V p pU p d
T
( )
r 1 r rT r 1 r rT r
x = v1 u1 d + v2 u2 d + L +
1 2
( )
1 r rT r
p
vp up d ( )
r 1 r 2 r p r
x = v1 + v2 + L + vp
1 2 p
2. Marquardt-Levenberg method; constrained least-squares with damping factor
r r
min A x d
2 r 2
+ x
2
[ 1 T
]
x = V ( + I ) U d
r r
r j T r
x = V diag 2 U d
+
j
Least-squares inversion (estimation) with prior information
r
(
r r T 1 r r
) (
r r T 1 r r
)
E (m ) = Am d Cd Am d + (m m0 ) Cm (m m0 )
)( )
r
(
m = AT Cd1 A + C 1 1
m
T 1 1 r
A Cd d + Cm m0
Organised
Factors affecting efficiency and reliability of an inversion algorithm
1. A solution exists
Mathematical description
Ill-posed (under-determined with a large null space
Ill-conditioned
Unstable (small errors in data cause large variations in the model)
Model parameterization dependence
= the data
T
(
m = A A A d ) SEG 2002
Regularization of the inversion problem
1. Regularization of the optimization method
2. Regularization of the model parameters = modification of the parameterization
b) ( A A)
T 1
AT V ( ) (( ) ) (U ( ) )
k k 1 k T
= eigenvalue truncation
b) e.g. add a term that encourages the velocity gradient to follow reflector
shape.
c) Modify the parameterization by using grid cells of variable size (data
driven (re)parameterization) based on the resolution matrix.
Thesis Cox 2004, P. 60
Regularization ralgorithms for seismic inverse problems
r
The forward model: Am = d
r r r r
The data residual: r (m ) = Am d
Data regularization
r 1 rp
1. L p norm: F (r ) = r where 1 p 2
p
When p=2, we have a least-squares problem; the solution tends to be sensitive to outliers/spiky data
residuals. The L1 norm (p=1) overcomes this problem by overweighting data terms with small
residuals and underweighting the large spiky data residuals.
r
(
2. Hybrid L1 L2 : F (r ) = 1 + r
r2 r2
r0 ) p 2
1
r
( r r
3. Cauchy distribution: F (r ) = ln 1 + r 2 r02 )
SEG 2002
Resolution matrix and Covariance matrix
r r
The problem: Gm = d
r 1
r r r
A particular solution: m p = G p d
1
m p = G p Gm
( )
r ~ ~ 1 2 1 ~ 1 2 ~ 1
= G p dd G p d G p G p = d G p G p
1
Thus once the operator G p1 for a particular solution is known, the resolution and the
error in the solution are easily obtained. Aki & Richards, P. 677
Resolution matrix and Covariance matrix
11 L 1M d1
r r ~ r r M 1 M
m
The problem: G m = d UV m = d M M =
M M M
1 1 ~
G = V U mM
N1 L NM dN
1
( )
~ 1 ~ 1 ~
The generalized inverse matrix: G = GG G = VM M U M
g
U 0 exists
U0 V0 does not exist;
M N
r
d r r
d Gm p
r
Gm p Up
r 1
r 1 r 1 ~
r
A particular solution: m p = G p d = G p G m = V p p U p d with p M
1 ~
The resolution matrix: G G = V pV p
p
2 1 ~ 1 2 ~
The covariance matrix: d G p G p = dV p p V p
2
( )( )
rT ~ T ~ r
Minimum energy: Emin = d U pU p I U pU p I d Aki and Richards, P. 677
Least-squares solutions for overdetermined and underdetermined systems
r r
t = A u
N,1 N,M M,1
1. Overdetermined system N M ; R( A) = M
r
(
u = AT A + 2 I )
1 r
A t
T
damped least-squares solution
2. Underdetermined system N M ; R( A) M
r
(
u = AT AAT + 2 I )1 r
t damped least-squares solution
r
where all u are perpendicular to the null space of A .
f
Area of noise boosting
Organised
Model Covariance and Resolution matrices
A common way of describing uncertainty in linear inverse problems is by means of model covariance and resolution
matrices (see the Figure). In essence the model covariance matrix shows how errors in the data propagate into errors
in the estimated model parameters. It is a square symmetric matrix with as many rows as columns as there are unknowns
in the problem. Off-diagonal entries indicate how errors are correlated between pairs of model parameters. For example
if the value is positive then the errors in the two corresponding parameters will tend to be positively correlated (e.g. when
data errors force one parameter too high (low) they will also tend to force the other too high (low)). A negative entry
indicates the converse. Significant off-diagonal entries indicate strong trade-offs in the variables. Diagonal entries give
information on the variance of corresponding parameters, i.e. how data noise affects them individually.
It is important to note that the uncertainty is described by the whole matrix and not just the diagonal entries. In practice
off-diagonal entries are often ignored which means that the influence of data noise can be severely underestimated, even
in linear inverse problems. As the number of unknowns increases model covariance matrices quickly become difficult to
calculate both due to the computational burden and also because the number of entries grows quadratically with the
number of unknowns.
The right side of the Figure shows six rows of the resolution matrix. These describe how independently each parameter
can be recovered from the data. Each row corresponds to a particular parameter and the height of the columns gives the
relative leakage or blurring of one parameter into another. For example, parameter 7 (first panel) is relatively well
resolved since its own column (shown by the arrow beneath) is much higher than the others, whereas parameter 9 (third
panel) is poorly resolved because its column is virtually indistinguishable from the other columns. Unlike the model
covariance matrix the resolution matrix does not depend on the errors in the data but instead reflects the uncertainty
arising from the nature of the forward problem as well as the amount and distribution of the available data.
An example of a model covariance matrix (left) and six rows of a model resolution matrix (right) for an inverse problem
with 24 parameters. On the left, blue indicates positive covariance values and red indicates negative ones. On the right,
the number in the box identifies the parameter index. Each column height represents the resolution value. The arrow
shows where the row crosses the diagonal of the resolution matrix. For rows corresponding to well-resolved parameters
we expect to see a dominant column at the position of the corresponding arrow.
R R R
R = R R R
R R R
Broadband wavelet:
individual layers are
well resolved
ambiguities
in the same
subsurface layer
Resolution matrices for the determination of three elastic parameters, using p = 0.0,0.1 and 0.2 s/km;
zero-phase wavelet, corner frequencies 5-12 Hz and 55-65 Hz. All cases have a similar sensitivity to noise.
(
Angle of incidence range of P-waves in the first layer = 0o 16o ) GPr Jan. 1991, 61-76
Resolution matrix for , , for a 12 layer case:II
(n ) = 10 ms, each
z = 10 m (Vp=2000m/s)
Resolution matrices for the determination of three elastic parameters, using p = 0.0,0.25 and 0.5 s/km;
zero-phase wavelet, corner frequencies 5-12 Hz and 55-65 Hz. All cases have a similar sensitivity to noise.
(
Angle of incidence range of P-waves in the first layer = 0o 44o ) GPr Jan. 1991, 61-76
Relative contributions to the reflection coefficientfrom Richards and Frasier
GPr Jan. 1991, 61-76
= =
= =
RPP ( j )
= =
1 2 2
1 4 2 sin j
2
+
1
2
(
1 + tan 2 j
)
2 2
4 2 sin j
Contribution of r , r and r to the
PP ( ) and PS ( )
reflection coefficients.
Organised
Relative contributions to the reflection coefficientfrom Richards and Frasier
GPr Jan. 1991, 61-76
= =
= =
RPP ( j )
= =
1 2 2
1 4 2 sin j
2
+
1
2
(
1 + tan 2 j )
2 2
4 2 sin j
Contribution of r , r and r to the
PP ( ) and PS ( )
reflection coefficients.
Organised
Relative contributions to the reflection coefficient from Richards and Frasier
Aki & Richards, P. 154
sin i sin j
p= =
}
RSS (i ) =
~ top of first layer
1
2
(1 4 p2 2 )
1 2 2
4 p
2
2 cos j
cos j
RSP (i) = RPS (i )
cos i
Contribution of r and r to the SS ( ) and SP ( ) reflection coefficients.
GPr Jan. 1991, 61-76; Quantifying seismic reflection data, by van Rijssen and Herman. Organised
Inversion with the propagation algorithm
Propagation can resolve far beyond the classical acceptation of resolution (i.e. 4 ).
This is because of two main reasons:
1. Propagation uses a (exact) layer parameterization. By this means, it reduces dramatically
the number of parameters to be resolved and it introduces the low and high frequencies
missing in the seismic bandwidth.
2. Propagation searches for lateral continuity in the data. This is introduced by using a
local search algorithm which uses for an initial guess the solution found at the previous
position.
The propagation algorithm ensures the continuity of the parameters along each layer (when
it is consistent with the seismic data) and therefore the results are geologically and dyna-
mically consistent.
It allows separate inversion for deltarho/rho and deltaVp/Vp without assuming any relation-
ship between the two parameters.
A layered definition of the model imposes a square shape for the elastic perturbations thus
broadening the bandwidth of the inverted signal.
Example
1. sample-by-sample: Vp, Vs, density 150 parameters
Time window of 200 ms; 4 ms sampling
2. e.g. 12 layers: Vp, Vs, density, thickness 48 parameters
TLE Nov 2013, 1318-1326; Breaking the limit of seismic resolution: A synthetic example, by Thore.
Elastic stratigraphic inversion
i1 i1
i
1 , 1 , 1
2
2
2 i
2
i
2 , 2 , 2 i2 +
+
+ i +
2
2 2 2
i
i+
2
2 2
1
2
(
RPP ( p) = 1 4 p
2 2
+
1
)
2 cos i
2
4p
We assume that i1 , i2 , j1 and j2 are real and that none of these angles is near 90o .
Then it follows that i = i2 ii = tan i ( ) and j = j2 ji = tan j ( ).
Correct to first order in the jumps , , . Aki and Richards
Ch.5; P. 153
Simplifications of the Zoeppritz equations Aki and Richards, Ch. 5
2 2
RPP (i )
1
2
(
1 4 p2 2 ) +
1
2 cos 2 i
4 p
Richards and Frasier: p=
sin i
=
sin j
1 2 2
RSS ( j )
1
(
1 4 p2 2 )
2 cos j
4 p
2
2
p 2 cos i cos j 2 cos i cos j
2 2
R PS (i ) 1 2 p
2 2
+ 2 4 p 4
2 cos j
cos j
RSP = R
cos i PS
1 1 2 2 1
R PP (i )
2
+ +
2
2 2
+2 sin i +
2
tan 2 i sin 2 i [ ]
[ ]
R PP (i ) R PP (0) + G. sin 2 i + F . tan 2 i sin 2 i = RPP (0) + G sin 2 i + F tan 2 i sin 2 i
2 2 1 1 Z 1 1 zero-offset reflectivities
R PP (0) = = = d ln Z +
2 2 + 1 1 2Z 2 2
1 I P 1
Fatti: = +
1 2 2 1 2 IP 2
RPP (i) + 2 2 2 + sin i + tan 2 i 1 I S 1
2 2 = +
2 IS 2
I P 2 2 I S 1 2 2 2
RPP (i ) (1 + tan i )
1 2
4 2 sin i tan i 2 2 sin i if i 35o 1.5 2.0
2 IP IS 2
Simplifications of the Zoeppritz equations
1 1
1 , 1 , 1
2
2
2
2
2 , 2 , 2 2 +
+
+ +
2
2 2 2
+
2
1 1 2 2 1
RPP ( )
2
+ +
2
2 2
+ 2 sin +
2
[
tan2 sin2 ]
RPP() A+ B.sin2
Organised
Intercept (A) and Gradient (B) Estimation with Stabilization
r r r
I. The model: d = Gm + n RPP ( ) = A + B sin
2
RPP (1 ) 1 sin 1 n1
2
in matrix notation:
A
M = M M . + M
R ( ) 1 sin 2 B n
PP M M M
II. The stabilized least-squares solution:
1
1 T r A 1 sin 2 r
r
( )
m = G G G d =
T
G d
T
B sin sin +
2 4
sin 2 ,
1
r r r
d = Gm + r
r
(
m = G G G d
T
)
1 T
r r
r = d Gm
r r
2 = rT r
(
cov(m ) = 2 G T G )
1
GPr 1999, 663-705 Geophysics 2001, 1284-1293
TLE Dec 2007 1508-1515
Acoustic impedance and Shear impedance
1 I P I
( )
2
RPP (i ) 1 + tan 2 i 4 2 S sin 2 i (~ Fatti)
2 IP IS
Volumes of P-impedance and S-impedance are obtained through blocky model-based inversion
and the and volumes (for isotropic materials) are obtained using the relationships:
= I P2 2 I S2 I P = = ( + 2 )
I S = =
= I S2
The inconvenience of the linearized reflection coefficient is that is depends on perturbations in three
parameters (e.g. the differences in P-wave velocity , the S-wave velocity and the density
across the interface) and requires prior information about Poissons ratio. Well control is needed to
provide such prior information.
1 I P 2 I S 2 2
RPP (i ) ( )
1 + tan i 4 2
2
IS
1
sin i tan i 2 2 sin i
2 2
2 IP 2
Organised
SEG 1997; AVO 2.7
Impedances & LMR = lambda, mu, rho
2 = ( + 2 ) = ( + (4 3) )
2 =
I P2 = ( ) = ( + 2 )
2
I S2 = ( ) =
2
= 2 2 2
= 2
= I P2 2 I S2
= I S2
( + 2 ) 2 2 1
RPP ( ) 1 + tan
1
4
2
(( + 2 )
)
2 2 sin + 1 tan 2
4
( ) ~ Goodway
I P 2 2 I S 1 2 2 2
(
RPP ( ) 1 + tan 2
)
8 2 sin tan 2 2 sin
~ Fatti
2I P 2I S 2 Goodway
P-reflectivity S-reflectivity the last term, in density only, cancels for most
ratios around 2 and small angles
Inversion: P-impedance S-impedance
1 2 2 1 2 2
RPP ( ) = 4 2 sin + 1 4 2 sin
2 cos
2
2
2
( )
2 2
RPP ( ) = 1 + tan
1 I I 1
2 P
4 2 sin
2 S
tan 2 2 sin
2
2 IP IS 2
I P = 1 I P 1 1 I S 1
= + = +
I S = 2 IP 2 2 IS 2
Parameterization of AVO behaviour
1 2 2 1 2 2
RPP ( ) = 4 2 sin + 1 4 2 sin
2 cos
2
2
2
RPP ( ) = (1 + tan )
1 I 2
I 1 2
2 P
4 2 sin
2 S
tan 2 2 sin
2
2 IP IS 2
1 1 2 1 K 2 1 1 2 1 1 1
RPP ( ) =
2
+ 2 sin +
4 3 cos K 3 cos 2 4 cos
2 2 2 2
1 1 2 1 2 1 1 2 1 1 1
RPP ( ) =
2
+ 2
2 sin +
4 2 cos 2 cos 2 4 cos
2 2 2
Two approaches:
- Relative contrasts - Simultaneously for all times in a
given window
- Single interface
- Explicit use of the wavelet
- Independently, on a sample-by-sample basis (temporal (vertical) correlation)
Geophysics Jan-Feb 2003, 185-198; Bayesian linearized AVO inversion, by Buland and More.
Workflow for AVO processing and Inversion
Logs CMP Gathers
Pre-Conditioning Mute
Mud-Invasion Correction
Multiple Attenuation
Trim Statics
Mute
Edited Logs
Wavelets Stack Cube Offset Scaling
Angle Gathers
Horizons
P Cube
AVO Attributes Transform
G Cube
Initial Model Building Initial Model Horizons Wavelets Angle Gathers
= ( + 2 )
= Zp Cube Zs Cube
I P =
I S =
LambdaRho MuRho Transform
= I 2 I
2
P
2
S
= I S2
LambdaRho MuRho
Cube Cube
Horizons Wavelet
Major interpretation
in time
Geological
Model Seismic Stack or A-trace (intercept)
Single volume in time
Low frequencies
Inversion
P-Impedance
Porosity Lithology
www.petrologic.de Organised
Post-stack Inversion
- Reduction of wavelet effects, tuning and side lobes. Hence enhanced resolution
of the sub-surface layers.
www.petrologic.de Organised
Pre-stack Inversion
- Sonic, density logs
Well Data - Markers (opt.)
- Deviation (opt.)
- Check shots (opt.)
Wavelet Seismic
Horizons
Major interpretation Wavelet Seismic
in time
Wavelet Seismic
Geological
Model Angle or offset gathers in time
Low frequencies
Inversion
Porosity
P-Impedance
Density
Fluid fill
www.petrologic.de Organised
Pre-stack Inversion
- Reduction of wavelet effects, tuning and side lobes. Hence enhanced resolution of the
sub-surface layers.
- The output components are physical rock properties. They can be directly compared to
well log measurements and allow a physically motivated link to reservoir properties.
- As amplitude variations for different angles/offsets are modeled during the inversion and
not averaged in the stacking process usually the P-impedance inversion is improved.
www.petrologic.de Organised
Post-stack stratigraphic inversion workflow
A three-step methodology:
Inversion parameters
The inversion has been carried out within a 650 ms long time window starting at 750 ms,
with the following inversion parameters:
1. Seismic standard deviation (estimated from seismic dataset) s = 29.8 %.
2. Standard deviation of the uncertainties on P-wave impedances Ip = 300 g/cm3 m/s.
3. Correlation length x = 2000 m.
4. Correlation length y = 1000 m.
Inversion results are obtained after 30 iterations.
A priori AI
model with
low frequency
constraints
derived from
stacking vels.
Optimal AI cubes
obtained from the
1094 and 2006
datasets.
The AI log derived
from well 15/9-13
is shown for com-
parison.
Comparison of seismic amplitude section (left) with acoustic impedance results obtained by
waveform inversion. In the right-hand side section, one out of three seismic traces of the
left is superimposed on the inversion results.
2I P 2I S
P-wave reflectivity
Global
Acoustic
Joint Inversion
Impedance
S-wave reflectivity
Rock properties and
Reservoir properties
Wavelets Model-driven
Constrained Shear
Simulated annealing Impedance
Macro-models
Convolutional model
I P (kt ), I S (kt ); k = 1, L , n
Low-frequency Ac.Imp.
E = W1 k =1 + k =1
n n
S obs
P (kt ) S obs
S (kt )
k =1 k =1
n pri n
I P (kt ) I P (kt ) I Spri (kt ) I Smod (kt )
mod
+ W2 k =1 Constraints compensate
+ k =1
n n for low-frequency compo-
I pri
P (k t ) I S
pri
(kt ) nents which are missing
k =1 k =1 in the band-limited data.
n
(I S (kt ) I P (kt )) pri (I S (kt ) I P (kt ))mod
+ W3 k =1 ~ background lithology trend
n
I S (kt ) I P (kt ) pri
k =1
Before inversion, both P- and S-seismic amplitudes must be scaled to the level such that they
match the synthetic seismic data generated from well logs.
1 d [ln I (t )]
S (t ) = W (t ) R (t ) and R(t ) = FB Oct. 2001; 557-566
2 dt
Simultaneous inversion acoustic impedance and shear impedance
I P 2 2 I S 1 2 2 2
( 2
)
RPP ( ) = 1 + tan 8 2 sin
tan 2 2 sin
FB June 2004, 47-70
2I P 2I S 2
I P 2 2 I S
(
RPP ( ) = 1 + tan
2
)
2I P
8 2 sin
2I S
Modeled impedances were used to derive
reflectivities, which were convolved with
the wavelet and compared with the real
seismic. Differences were minimized by
a simulated annealing technique.
Organised
A B C Lithology
and
hydrocarbon mapping
IP PP and PS Inversion
IS
GPr March 2010, 297-306
Acoustic impedance (top, Sw inserted), shear impedance (middle, volume of shale inserted) and density
(bottom, effective porosity inserted) sections after prestack joint inversion. Units of density and impedance
are g/cm3 and g/cm3xm/s. Color inserts are corresponding log data. Low acoustic impedance and low-
density zone observed below about 1600 ms at Well-C is due to gas-charged Balder shales. Well-A has pay
sands and reaches the seismic line at about termination depth position. It is displayed for reference only.
AVO inversion
The output volumes from AVO represent P-waves, S-waves and density; however, they still relate only to
relative impedance contrasts at layer boundaries (reflectivities). Changing reflectivities to rock-layer
properties requires conversion from the observed relative measurements to absolute values. The inversion
method is the model-based method. In this approach, an initial smooth-impedance model is derived from
the well data and interpreted seismic horizons. The model constrains the acceptable absolute impedance
range as well as supplying a low-frequency trend from well logs which is lacking in seismic data.
The P-impedance (Ip) and S-impedance (Is) volumes can be transformed into additional rock property
volumes: = I 2
s
= I p2 2 I s2
((m/s)(g/cc))
90
100
150
200
220
Figure 6. Profile through the P-impedance model used for the model-based inversion of the P-wave
reflectivity.
TLE Nov 2012, 1362-1372
((m/s)(g/cc))
Ip
((m/s)(g/cc))
Is
Figure 7. Inversion results for (a) P-impedance, (b) S-impedance, and (c) density. At well locations, gamma-
ray log curves are displayed along with appropriate color attributes where available (without shear logs at
these wells, S-impedance cannot be calculated).
TLE Nov 2012, 1362-1372
= I s2 AVO inversion
= I p2 2 I s2
Figure 8. LMR transformation results. (a) lambda rho and (b) mu rho. At well locations, gamma-ray log
curves are displayed (without shear logs at these wells, lambda rho and mu rho cannot be calculated).
TLE Nov 2012, 1362-1372
Rock Properties from Seismic Data
Rock Properties from Seismic AVO Inversion Seismic Data
T VP
I VS VP
M
E
CDP
VS = /
T
I Angle Gathers
M
E
CDP
Reservoir Properties from Seismic Data
Rock Properties from Seismic Sand Probability Section
T VP
I VS T
Top Frigg
M I
E M
CDP E
Top Sele
T
I
M
CDPs
E
Sand Probability
CDP 0.0 1.0
Simultaneous Inversion
In Simultaneous Inversion, we
analyze fully-processed CDP
gathers to create volumes of ZP, ZP
ZS, and Density:
ZS
Density
Pre-stack Inversion
Pre-stack Inversion has multiple attribute outputs.
ZP
ZS
Density
Vp/Vs
Synthetic
Angle gathers: Vshale log inserted
The simultaneous AVO inversion workflow has given proven results in field studies around
the world. The workflow has at its core a constrained sparse spike inversion (CSSI)
algorithm. The elastic models generated from constrained sparse spike inversion are
absolute and therefore require low-frequency trend models to constrain the lowest
frequencies not present in the seismic data and not obtained through the sparsity criterion.
An unconstrained reflectivity inversion of the seismic data is first carried out in the
reflectivity domain. The resultant reflection coefficients are merged with the reflectivity
from the low-frequency model and integrated with impedance. A fully constrained inversion
of the impedance is then carried out to improve the fit to the input seismic data. The
frequencies at which the unconstrained reflectivity and low-frequency models are merged
are selected to ensure that the maximum information is coming from the seismic data, but to
avoid any gaps in the amplitude spectrum.
Constrained
Sparse
Spike
Inversion
Simultaneous AVO inversion workflow used to convert seismic angle stacks to rock property cubes of P-
impedance, Vp/Vs, and density.
Organised
Elastic inversion
Three term inversion of
prestack time-migrated Background model
used as as input See next slide
data; angles 60o
Forward modeling to
match real seismic
Genetic algorithm:
Fig. 1. Cascaded seismic and rock physics inversion workflow. Seismic angle stacks are first inverted for
P and S impedances (seismic inversion). We then construct a volume that describes the fluid distribution
in the reservoir. The P and S impedances along with the fluid-fill volume are then simultaneously
inverted (the rock physics inversion) to produce porosity and shale volume.
Organised
Rock physics model
TLE July 2005, 732-736
Fig. 2. Schematic depiction of the rock physics model. Clay and sand grains are mixed together in
proportion to the specified VShale to form a solid rock. Clay and sand pores are then added to the solid using
differential effective media theory. The sand pores have a large aspect ratio and are dry when they are
added. The clay pores have a low aspect ratio and are filled with brine when they are added. In the last
step, fluids are added to the sand pores using Gassmann fluid substitution.
Organised
TLE July 2009, 772-781
Data conditioning prior to inversion
Fig. 1. Raw (left) and conditioned (right) gather at the well location. Reservoir zone is indicated by the red
lines at 3.33.35 s. Circled zone shows an area of significant improvement after stretch removal.
Horizontal axis is offset (ft) and vertical axis is time (s).
Organised
Data conditioning prior to inversion TLE July 2009, 772-781
1. SNR enhancement
2. NMO stretch removal
3. Gather flattening
Fig. 2. Conditioned gather (left) and the two-term Shuey AVO-fit metrics (right). First metric trace is for
the raw data, second is with SNR enhancement and stretch removal, third also includes alignment. The
metric traces are scaled from 0 (on the left of the trace) to 1 (on the right). A value of zero means no fit; 1
means perfect fit. Red numbers are rms amplitudes of the trace segment indicated by the horizontal red
lines. Organised
TLE July 2009, 772-781
Data conditioning prior to inversion
Angle dependent wavelet extraction from lls shaping filter that when convolved with the reflectivity from
the well best matches corresponding seismic trace.
Fig. 5. Raw (lower) and conditioned wavelets (upper) for the four angle stacks. First panel is wavelet
shape, second is amplitude spectra, and third is phase spectra. Note that the far wavelet in the raw data
becomes unstable at about 30 Hz, corresponding to a significant drop in frequency.
Organised
TLE Dec. 2009, 1492-1497
Inversion and Reservoir Modeling
Fig. 1. Workflow diagram. The blue arrows represent a linear flow, the order in which processes need to
be applied for the first result. The green arrows represent iterations within the workflow.
Organised
TLE Dec. 2009, 1492-1497
Inversion and Reservoir Modeling
Criteria for quantitative use of seismic data:
1. A seismic image colocated with the well data (i.e. in the correct 3D space).
2. Amplitude variations consistent with the Earth in two-way time (TWT), spatially and
across angles of incidence.
3. Amplitude variations that can differentiate lithologies and fluids (i.e., sufficient elastic
property contrasts between lithologies and fluids.
4. Completely flat CMP gathers prior to stack.
5. Good seismic-to-well ties.
6. Good temporal resolution relative to reservoir thickness.
Organised
TLE Dec. 2009, 1492-1497
Inversion and Reservoir Modeling
Fig. 2. CMP gathers (a) before and (b) after seismic conditioning. (d) The 2428 angle stack after seismic
conditioning. (c) The corresponding, previously available, angle stack.
Organised
Inversion
and
Reservoir Modeling
TLE Dec. 2009, 1492-1497
Petro
Elastic
Model
Fig. 6. (a) Schematic showing the fundamental difference between the elastic and petro-elastic inversions.
(b) Petro-elastic well-tie analysis. The black curves are the well logs and the blue curves are the elastic
properties computed from the PEMs. On the far right panel, the gray trace is the real seismic and the blue
trace is the synthetic seismic created from the PEM-derived elastic properties. The X-sand is yellow.
TLE July 2010, 842-846
Multiparameter Simultaneous AVO Inversion
The process requires seismic horizons to be interpreted to provide structural guidance,
and a low-frequency model to be estimated from this structural information, Z p , Z s and
calibrated logs (from available boreholes), and additional constraints.
AVO modeling and inversion are highly sensitive to flattening or alignment of events
near to far offsets.
Results of simultaneous inversion showed a good match between AI, SI and measured log data.
FB May 2012, 69-75 Organised
Lithology and hydrocarbon mapping: PP and PS Inversion
2
2
(
RPP ( ) 1 + tan 2
) I
2I P
P
V
8 sin
S 2 I S
2I S
2 V S
tan 4 sin 2
2
VP VP
VP tan 2VS2 4VS2 I S
RPS ( , )
2VS 4VS
1 + 2 sin cos cos 2 sin cos cos
2 2
VS V P V P 2 V P V P 2 I S
where (IP ) VP and (IS ) VS
= + = +
IP VP IS VS
B sin sin +
2 4
sin 2 ,
1
r r r
d = Gm + r
r
(
m = G G G d
T
)
1 T
r r
r = d Gm
r r
2 = rT r
( )
GPr 1999
cov(m ) = G G
1 Geophysics 2001
2 T
P. 663-705 P. 1284-1293 Organised
PP, PSv and SvSv ~ the value of additional data sets
r 1 T r
r r
d = Gm + n m = G G G d
T
( )
PP data: RPP (1 )
A11 (1 ) A12 (1 ) A13 (1 ) RI P
M = M M M RI S
R ( ) A ( ) A ( ) A ( ) R
PP n 11 n1 12 n 13 n
r r r rT r
n = d Gm n = n n
2
Cov(m ) = G G 2
n ( T
)1
Organised
Joint PP and PS seismic inversion
Fig. 2. Blocked well logs for a producing oil well at Blackfoot Field.
TLE Sep 2001, 1048-1052; Joint PP and PS seismic inversion (Crewes) Organised
Joint PP and PS seismic inversion
Fig. 3. PP (blue) and PS (red) synthetic seismograms for the well logs of Figure 2. (The seismograms were
made from the unblocked logs.) In each seismogram, the three traces on the right are three repetitions of
the stacked trace. The PP seismogram has a 1080 Hz zero-phase wavelet while the PS seismogram has a
1040 Hz zero-phase wavelet.
TLE Sep 2001, 1048-1052; Joint PP and PS seismic inversion (Crewes) Organised
Joint PP and PS seismic inversion
TLE Sep 2001, 1048-1052; Joint PP and PS seismic inversion (Crewes) Organised
Joint PP and PS seismic inversion
Fig. 4. The results of a comparison of the PP inversion and the simultaneous PP and PS inversion for a
simple synthetic consisting of a single reflector with reflection amplitudes taken from the exact Zoeppritz
equations. Percent error is shown for three different signal-to-noise (S/N) ratios.
TLE Sep 2001, 1048-1052; Joint PP and PS seismic inversion (Crewes) Organised
Joint PP/PS Inversion
Acquisition: 2D line, 3C Sercel DSU3 (digital sensor unit) MEMS accelerometer.
PP and PS 1D elastic isotropic joint inversion using simulated annealing reconciles the
VP VS from event registration with VP VS from AVO inversion from RPP , GPP and GPS
from three angle PSTM PP-gathers and PS-gathers.
Inversion results:
VP VS , fluid factors and Poissons ratio in the target zone.
Acquisition: offshore Brazil; using 3C-2D seismic data from an OBC survey.
Eseismic = penalty for differences between seismic data and synthetic seismic accumulated over all
partial stacks and is related to the S/N of each PP and PS seismic angle stack.
E prior = penalty for deviation of the estimated layer properties from the LFMs accumulated over all
layer properties, and guides the solution towards/against the a priori model.
Ehor = penalty for horizontal changes accumulated over all layer properties, and controls the
horizontal continuity of the results.
Evert = penalty for vertical changes accumulated over all layer properties, and controls the
threshold for reflection coefficients.
Organised
Joint PP/PS Inversion
Fig. 6. The required inputs and standard outputs of simultaneous inversion of multicomponent
data.
TLE Oct 2010, 1220-1226 Organised
Joint PP/PS Inversion
Fig. 7. Well 1 density inversion result for joint PP and PS (left) and PP only (right).
Fig. 8. (left) Simultaneous inversion of PP and PS data. (right) Inversion of PP data only.
TLE Oct 2010, 1220-1226 Organised
Geophysical reservoir characterisation
Interpreted Horizons
Partial Stack
Well Logs
AVO Analysis
Attribute Analysis
DHI Analysis
EI Inversion
AI Inversion
GEOPHYSICAL RESERVOIR
MODEL
Organised
Bayes Theorem
P (B A) P( A)
P( A B ) P(B ) = P(B A) P( A) P(A B ) =
P (B )
P(B ) = P (B A1 ) P( A1 ) + L + P (B An ) P( An )
P(B Ai ) P( Ai ) P (B Ai ) P( Ai )
P(Ai B ) = =
P(B A1 ) + L + P(B An ) P (B )
then
Geophysics Jan-Feb 2001; 55-69; A Bayes tour of inversion: A tutorial, by Ulrych, Sacchi and Woodbury.
Organised
TLE May 2010, 596-601
Bayes Theorem
= Prior probability
Likelihood =
Fig. 5. Predictive power (in Bayesian jargon, the posterior probability) as a function of test reliability and
the base rate of occurrence (also called the prior probability of the event of phenomenon in question). The
position of the scenario in the exploration problem is shown by the white square.
Organised
Bayes Theorem - Example
Imagine you are working in a newly accessible and unexplored area of an otherwise
mature basin. Statistics show that on average 10% of structures are filled with gas and
the rest is dry. Fortunately, you have some seismic analysis technology that allows
you to predict the presence of gas with 80% reliability. In other words, four out of five
gas-filled structures test positive with the technique, and when it is applied to water-
filled structures, it gives a negative result four times out of five.
You acquire some undrilled acreage, map some structures, and perform the seismic
analysis. One of the structures tests positive. Assuming this is the only information
you have, what is the probability that it is gas-filled?
Negative Positive
Water 72/100 18/100
Gas 2/100 8/100
The possible scenarios in the exploration problem. Water scenarios make up 90/100 cases;
gas just 10/100. The high rate of false positive confounds the reliability for the minority
scenarios.
( )
r r
p md =
(
p )
r r
d m p (m
r
)
() r
Bayesian inversion:
pd
constant
Special case
r
The means and covariances of m are assumed to be given; the a priori pdf is:
r 1 r rT r r
p (m ) exp (m ) Cm1 (m )
2
The noise is assumed to have zero mean and covariance matrix Cn ; then the likelihood
function can be written as:
( )
rr
( )
1 r r r T 1 r r r
(
p d m exp d g (m ) Cn d g (m )
2
)
rr
( ) r
We want to maximize p d m p (m )
( )
r
= ( A Cn A + Cm ) A C n d + C m
r r r 1 r
For g (m ) = Am , the explicit solution is: m T 1 1 1 T 1
r 1 r rT r r
p(m ) = exp (m ) C 1 (m )
1
Gaussian or normal pdf:
(2 )N 2 C 1 2 2
1 m 2
p(m ) =
1
N =1 exp
2 2
N =2 p(m1 , m2 ) =
1 1 m 2 m m m
2
= exp 2 12 +
( )
1 1 1 1 2 2 2 2
2 1 2 1 122 2 1 12 1 1 2 2
2
lim F (m ) = 0
m
dF (m )
The probability density function (pdf): p (m ) =
dm
as a consequence F (m ) = p(u )du and p(m )dm = 1
m
N stochastic variables M 1 , L , M N
r
The joint cdf: F (m ) = P(M 1 m1 L M N mN )
r
r r N F (m )
The joint pdf of M : p (m ) =
m1 L mN
( )
r r
P M A = L p (m )dm1 L dmN
A
GPr 1988 P. 878-898
GPr 1988 P. 899-918 Organised
GPr 1988 P. 878-898
Probability Theory GPr 1988 P. 899-918
r r
Given two sets of variables M and D , the marginal pdf gives the information on one
set of variables regardless of the other set. It is obtained by integrating
r the joint pdf r
()
r
over the disregarded set of variables; e.g. the marginal pdf p d for D disregarding M
is given by:
()r
( )
r r
p d = L p d , m dm1 L dmN
() r
( )
r r r
or, shorthand p d = p d , m dm
The conditional pdf gives the information on one set of variables, when values for the
other set of variables have become available. Intuitively it should be proportional torthe
( )
r r
joint pdf with the rvalues for the other set filled in. The conditional pdf p m d for M
given values for D is defined as:
( ) (( ))
r r
( ) ( )
r r
r r p m, d r r
p md = r ; note, that p m d p d m
pd
( ) ( )() ( )
r r r r r r r r
p m, d = p m d p d = p d m p (m ) Organised
Probability Theory
r r
Two sets of variables M and D are independent when their joint pdf can be written
as the product of their individual (marginal) pdfs r
( )
r r r
p m, d = p (m ) p d ()
( ) p (d m ) = p (d )
r r r r r r
It follows that in this case p m d = p(m ) ;
i = 1, L , N or, = E (m ) = m p(m )dm
[
r r r rT r
] r r rT r r
C = E (m )(m ) = (m )(m ) p (m )dm
[ ]
which is shorthand for cij = E (mi i )(m j j )T =
(mi i )(m j j ) p (m )dm
T r r
The diagonal elements of this matrix C are the variances i2 = cii = E (mi i )
2
[ ]
of variables M i .
- The mode m of the posteriori pdf corresponds with the maximum of the a posteriori
pdf and is referred to as Maximum A posteriori Pdf (MAP) estimator. It represents
the most likely set of values for the parameters given the measured data, the theoretical
information and prior information.
From a Bayesian point of view the MAP estimator is preferred over the mean and median estimators, as it
represents the most likely set of values for the parameters given all measurements and information. More-
over, from a practical (numerical) point of view the MAP estimator is to be preferred too, since the calcu-
lation of the mean and the median requires the evaluation of the a posteriori pdf through the whole para-
meter space. Finding the maximum of the a posteriori pdf turns the inverse problem into an
optimization problem. Organised
Point estimates and probability density functions
Point estimation
r r
MAP estimator: maximum of the a posteriori density function p m d ( )
The MAP estimator can be interpreted as yielding the most likely values of the parameters
given data and a priori information.
( )
r r
If p m d is symmetrical and unimodal, the mean coincides with the mode and the least-
mean-squared estimator is equivalent to the MAP estimator.
a posteriori pdf
likelihood function
a priori pdf
( ) ( ) ( )
r r r r r r r
( )()
r r r
p m, d = p d , m = p d m p(m ) = p m d p d
( )
p md =
( )
r r r
r r p d m p (m )
( )
r r r
p d m p(m )
r r r r
d = g (m ) + n
()pd
r
( )
r r constant
p m d = the a posteriori pdf, the solution of the inverse problem
( )
r r r r r
p d m measured data = forward modeled data d from parameters m plus the noise n
r r
p (m ) gives information about m disregarding the measured data
r
Learning by experience allows the integration
r of information from different sources.
r d1 were used to estimate the m , using Bayess rule (see above)and that
Suppose data
new data d 2 are measured then applying Bayess rule gives:
(
r r r
p m d1 , d 2 = ) (( )
r r r
) r
p d1 , d 2 m p(m )
=
p
r r
d 2 m
r
p( ) ( )
r r
d1 m p (m
r
)
( ) ( ) ( )
r r r
p d1 , d 2 p d2 p d1
The last factor in this equation, recognized as the posterior of the first parameter estimate,
r becomer the prior knowledge for the second estimate!
( ) ( )
has r r r
d1 and d 2 are assumed to be independent, i.e. p d 2 d1 = p d 2 . Organised
GPr 1988 P. 878-898
Bayesian Estimation GPr 1988 P. 899-918
likelihood function
a posteriori pdf a priori pdf
( )
r r
p md =
(p)r r
d m p (m
r
)
() pd
r
constant
a priori pdf
lkhd function
a posteriori pdf
( )
r r
The posterior pdf p m d is obtained using the three ingredients:
r
1. The data d (including the noise) (pdf)
r r
2. Theoretical information g (m )
r r
3. Prior information about m ; pdf p (m )
r
Prior information about m can be obtained by e.g.
- Regional and local geologic information
- Tabulated values from similar prospects or from literature
- Empirical values for the individual parameters or empirical relations between different
parameters, e.g. Vp versus Vs, porosity versus depth
- Hard constraints on parameter values, usually based on physical grounds
- Experts (interpreters, petrophysicists) information.
Organised
Bayesian integration of
rock physics and Markov
random fields.
Rules and representations used in the network. Spatial representations are used for facies (Q)
and fluid saturation (S). Other variables are defined through S and Q by rock physics theory.
Parameters are extracted from mineralogical values, well data and prior knowledge.
TLE March 2002, 290-294 Organised
Bayesian integration of rock physics and Markov random fields.
Fig. 3. Network relating the reservoir variables; Facies (Q), fluid saturation (S), porosity ( ),
density ( ), VP, VS, and reservoir variables to the data; Reflection times (tx), well data (dw),
AVO seismic data (ds,x) and cap-rock properties (cr). x denotes location, and the network holds
for every location.
The Bayesian approach includes deterministic as well as stochastic inversion. The deterministic
inversion is represented with the posterior mean and is smoother than the stochastic simulation.
Simulation can capture minute details of the elastic parameters, but the solution is non-unique
and hence multiple simulations should be used.
FB Nov 2006, 69-77
FB Dec 2006, 79-84 Organised
Sequential Gaussian simulation - SGS
A valid geostatistical simulation must be
a) Conditional to the measured data.
b) Reproduce the histogram of the data.
c) Honor the spatial correlation function, such as a variogram.
Rock/Fluid
Property
Model
Update Rock
Properties
FB Nov 2006 P. 69-77
FB Dec 2006 P. 79-84 Organised
Stochastic inversion
Detail of cross-line through a well at B46 sand level showing deterministic inversion result
(top) and mean of 100 stochastic impedance realizations (lower). Sands indicated by blue/
purple colors.
Comparison of the deterministic inversion result with the next four realizations shows
how much uncertainty is smoothed out by deterministic inversion.
FB Nov 2006 P. 69-77
FB Dec 2006 P. 79-84 Organised
Stochastic inversion
Deterministic seismic inversion refers to those processes that provide only one representation
of the subsurface for a given set of inversion parameters despite the uncertainties inherent in
the input data and inversion parameter settings.
Geostatistical inversion is designed to work with uncertainties. Uncertainties in the input
data and model parameters are all incorporated within the inversion process to generate
multiple representations of the subsurface that can be combined to quantify the uncertainty
in the results.
Uncertainties
1. Noise in the seismic data
2. Limitations due to the input data, e.g. limited angular range, limited bandwidth, unknown
wavelet(s)
3. Overlap of impedance values between different lithology types due to inherent overlap
4. Component of low-frequency models between shale trend and lower-frequencies of seismic
5. Uncertainties in shale trend
6. Relationship between the elastic properties and petrophysical properties (rock physics model)
FB February 2010, 35-44 Organised
Comparison of simultaneous AVO inversion techniques
Deterministic
Simultaneous
AVO inversion
Geostatistical
Simultaneous
AVO inversion
Upper Morrow comparison of seismic and inverted data over the study area. The top panel shows the
full angle stack. The center panel shows P-impedance from deterministic simultaneous AVO inversion
of the seismic partial-angle stacks. The lower panel shows the mean P-impedance from a set of reali-
zations of a geostatistical simultaneous AVO inversion. Note the large increase in detail in the lower
panel. Wells A on the right and B on the left show the high-cut filtered P-impedance logs displayed on
top of the inversions. SEG 2010, AVO 1, 363-366
Types of Inversion techniques
Types of inversion: Types of inversion:
. Deterministic . Coloured inversion (Deterministic)
. Stochastic-elastic . Non-linear sparse spike inversion (Deterministic)
. Stochastic-petroelastic . (Model-based) simultaneous inversion (Deterministic)
. (Joint) geostatistical inversion (Stochastic-elastic)
. Bayesian inversion delivery (Stochastic-petroelastic)
Statistical RPMs (rock physics model) can be used to turn this process into a joint geostatistical
inversion. The statistical RPMs capture the relationship between the two (or three) target impe-
dances e.g. AI and SI. Once N AI realizations have been derived, incorporate these RPMs in
the stochastic inversion to SI. The RPMs will ensure that pair-wise the N AI and SI realizations
have characteristics corresponding to the wells.
Quantitatively integrates well log elastic rock properties and AVA seismic to produce calibrated
quantitative 3D volumes of rock properties.
jason.cgg.com
Deterministic inversion and Geostatistical inversion
StatMod MC and RockMod use Bayesian Inference and Markov Chain Monte Carlo (MCMC) simulation
methods to build geologic models that honor wells, geology and seismic data rigorously and provide a
quantified measure of uncertainty. StatMod MC uses one stack of seismic data as input and therefore
produces P-impedance as the elastic property output. RockMod uses multiple angle stacks and produces
P and Shear Impedance and an estimate of Density. On this page, well refer to the two products as
MCMC since they share the same algorithmic processes.
MCMC yields far greater vertical detail than available from standard seismic inversions. It integrates
high resolution well data with low resolution 3D seismic and provides high vertical detail near and
away from well control. MCMC generates reservoir models with geologically plausible shapes and
provides a clear quantification of uncertainty to assess risk. Highly-detailed petrophysical models are
generated, ready for input to reservoir flow simulation. www.fugro-jason.com/res_srvcs/index.htm
Deterministic inversion and Geostatistical inversion
www.fugro-jason.com/res_srvcs/index.htm
Factors affecting efficiency and reliability of an inversion algorithm
Model
parameters
Theory
r r
Observed
r g data d
m
Inversion
r
Forward modeling and inversion. Data d are the measured outcome of an experiment.
r r r r r r r
The standard model: d = g (m ) + n ; particular case: g (m ) = Am
(M,N) (N,1)
(M,1) (M,1)
r
d = (d1 , L , d M ) = the data, i.e. the observations (M,1)
T
r
m = (m1 , L, mN ) = the model parameters (N,1)
T
r
n = (n1 , L , nM ) = the theoretical or experimental errors
T
(M,1) Organised
Estimating uncertainties with Monte Carlo
r r r r
d = g (m ) + n
r r r r
( ) (
r r r
The likelihood function l (m ) can be written as: l (m ) = p d m = pn d g (m ) )
r r r r r
with pn (m ) the pdf of the noise n , and n independent of g (m ) .
( )
r r
For Gaussian noise with zero mean = 0 and covariance matrix Cn , we can write:
r
l (m ) =
1
( ) (
1 r r r T 1 r r r
)
exp d g (m ) Cn d g (m )
(2 ) N 2
Cn
12
2
A Monte Carlo algorithm is applied to find the global extreme and the shape of the posterior
pdf, providing parameter estimates and their associated uncertainties.
A guided Markov Chain Monte Carlo method learns during the search; this makes it
possible to estimate many more parameters than with crude Monte Carlo, in a statistically
sensible way.
Organised
Iterative trace inversion
Proposed
layer-
parameters
Layer-parameter r r'
mm
1D convolutional update
forward modeling
No
r r
g (m )
Synthetic
trace
Yes Estimated
Satisfactory layer-
fit ?
r r r parameters
r Trace from d g (m )
d migration
Iterative trace inversion loop. A Markov Chain Monte Carlo algorithm is used to
propose parameter updates in order to properly sample the parameter space.
Organised
Stochastic trace inversion
The 1D convolutional trace model: s(t ) = w(t ) r (t ) + n(t )
r (t ) = R j (t j )
N
j =1
r
n = (n(t1 ), L , n(t M ))
T
{
r
The model parameters are split up in two vectors: mR = ( R1 , L , RN )
T
reflectivities
r
m = ( 1 , L, N )
T
layer thicknesses
j =1
Organised
Simulation of a continuous random variable with a CDF F(x)
Cumulative Density Function
1
F(x) U
Draw a
uniform
number U 0.5
between
zero and one
0 -1
-3 0 X=F (U) x 3
p4 X4
p3 X3
U(0,1)
p2 X2
p1 X1
P(m d ) P (d m ) P(m ) Likelihood function which gives a measure of how good a model m is
(d )
M dm,i = theoretical data
m ,i d
2
obs ,i W 2
i obs ,i
dobs,i = experimental data
with S= = 2 i =1
and
F Wi = weight (usually unity)
obs,i = data uncertainties vector
F = number of degrees of freedom
GPr 2008 P. 357-371 Organised
Improved Monte Carlo sampling
The optimization approach uses a random generation of the model parameters but they
can either guide their search using a transition probability rule, i.e. simulated annealing
based on the Metropolis algorithm or apply genetic algorithms. These approaches are
used to reduce the number of needed simulations by concentrating the sampling on the
highest probability density region of the model parameter space.
One method that can be considered a compromise between sampling and optimization
is the importance sampling method (GPr 1996; 313-350). This method is based on opti-
mization and the sampling is concentrated on high probability regions but without limiting
the scan to a single zone.
GPr 1998; 313-350, Bayesian inference, Gibbs sampler and uncertainty estimation in seismic inversion
GPr 2008 P. 357-371 Organised
Markov Chain
Monte Carlo
(MCMC)
sampling
2 2
Markov
r ( )
r r
Chain Monte Carlo sampling of the p m d for a model consisting r 0 of two parameters
m = (m1 , m2 ) . On the left-hand side an initial collection of 25 models mi (i {1,L,25}) .
On the right-hand side the situation is sketched after a certain number of updates.
r'
The Metropolis scheme: a new model m is accepted only with the probability: r
( )
p m'
P = min 1, r
r p(m )
m = the current set of model parameters
r
m ' = the new set of model parameters Organised
Improved Monte Carlo inversion
Inversion of surface wave data suffers from solution non-uniqueness and is hence strongly biased
by the initial model. The Monte Carlo approach can handle this non-uniqueness by evidencing the
local minima but it is inefficient for high dimensionality problems. Smart sampling of the model
parameter space becomes more efficient and with respect to traditional global search methods
avoids the subjective use of control parameters that are barely related to the physical problem.
The optimization methods use a random generation of the model parameters but they can either
guide their search using a transition probability rule, i.e. simulated annealing, based on the
Metropolis algorithm or apply genetic algorithms. These approaches are used to reduce the number
of needed simulations by concentrating the sampling on the highest probability density region of
the model parameter space. One method that can be considered a compromise between sampling
and optimization is the importance sampling method. This method is based on optimization and
the sampling is concentrated on high probability density regions but without limiting the scan to a
single zone.
Pre-stack 1D vertical
depth depth-to-time
migrated data conversion
PreSDMed 2d
tv ( x1 , x2 , x3 ) =
x3
data in 0 v p ( x1 , x2 , )
vertical twtt
from seismic-to-well tie
Reservoir
layer parameters
The migration image in tv is a structural, band-limited image of (reflection angle-averaged) subsurface
reflection coefficients in which depth is represented by vertical two-way traveltime, that is in principle
suitable for inversion with 1D convolutional modeling.
Organised
Trace inversion overview
r (tv ) = R j ( ) (tv j )
N
Target area: reflector n to N
For post-stack data: = 0o
j =n
Seismic data
Depositional environment
{ , ,
Lithology
Diagenesis
Mineralogy
Burial
Compaction
Porosity IP , IS ,
Permeability
+
Fluid
Tectonics Structure
Saturation
Eustacy Thickness
Climate
Rate of deposition and subsidence
Geophysics March-April 2001, 398-412; Y2K Tutorial: Fundamentals of seismic rock physics, by Wang.
From seismic traces to reservoir properties: a Bayesian scheme
r
m = model parameters H , C , , Sw
r
d = real seismic data
posterior PDF H = reservoir thickness
likelihood function
prior PDF C = clay content
( ) r r = porosity
r
r r P d m P m ( )
P (m d ) =
Sw = water saturation
P (d )
r
( )
r r r r r
P d m d g (m )
P (B A) P( A)
posterior PDF H = reservoir thickness
P(A B ) =
P (B )
C = clay content
= porosity
Sw = water saturation
The methodology consists of six steps:
1. Identification of the reservoir in the well and the corresponding seismic data.
2. The well data provide the information to establish a rock-physics transform between porosity,
lithology (clay content), and fluid (Sw) and the elastic properties.
3. Set bounds and discrete increments for H , C , , Sw . Expected spatial variability is based on
geologic reasoning; sizes of increments are related to their significance (impact) for the seismic.
4. Construct for each sample of model space an elastic model.
5. Construct synthetic seismograms from all samples of the model space.
6. Accept or reject synthetic traces based on the correlation with the real seismic data; this gives the
likelihood function P B A .( ) ( )
P B A from e.g. 5% acceptance
80% correlation threshold
( ) {
P AB
- marginal PDFs
- most likely combination of properties
or
15% acceptance
( ) (
e.g. P H B = )
P H , C , , Sw B dCddSw
90% correlation threshold
Fig. 7. Bivariate PDFs for six two-parameter combinations. PDFs with the clay content or saturation
component exhibit high uncertainty.
TLE April 2008, 456-461 Organised
From seismic traces to reservoir properties: a Bayesian scheme
Fig. 8. The most likely combination of reservoir properties from the full conditional posterior distribution
includes thickness of 120 m, clay content of 0.2, porosity of 0.2 and saturation at 0.4. The corresponding
most likely synthetic seismograms are displayed next to the real data.
( )
prior distribution
r r
p d m, H qm (m H )
posterior distribution
r
(
r r
) s (m H ) = vector with measurement predictions
r r
p m d, H = r
pdH ( )
H is an indication for the assumptions (hypothesis) that are made.
The prior distribution contains all the information that can be obtained from the petrophysical data and
from the geological concepts. The vector with measurements contains the samples of a seismic trace
and interpreted arrival times of seismic events. The likelihood links the information that can be extrac-
( ) [
ted from the measurements to the prior. It is assumed that the noise in the seismic trace is additive and
)]
r r r
for this reason can be written as p d m ,H = q d s mH .
r
n (
From the posterior distribution, a value for the posterior probability density can be computed for each
possible model and, as such, the posterior distribution contains the full solution of the inversion problem.
This is useful to find the global maximum of the posterior distribution or compute posterior expectation
values and standard deviations of reservoir properties.
The maximum of the posterior distribution can be found with an iteration scheme comparable to a lineari-
zed inversion method. The values of the model parameters at this maximum can be an estimate of the
posterior expectation values. The second-order derivatives around the maximum can be used to compute
the standard deviations of this estimate.
TLE Dec. 2009, 1478-1484; Seismically constrained probabilistic reservoir modeling by Leguijt Organised
Seismically constrained probabilistic reservoir modeling
When the problem is strongly nonlinear, as withrseismic inversion, a better result can be obtained
by generating an ensemble of model samples, {mi : i = 1, L , N }, from the posterior distribution
using a sampling method such as the Metropolis algorithm. Expectation values of reservoir proper-
ties are estimated using the Monte Carlo summation method, 1 N r
FN = f (m ) i
N i =1
where FN is an runbiased estimator of the expectation value from which the property is defined by
the function f (m ) .
A large advantage of the probabilistic approach is that apparent discrepancies between different types
of information caused by noise in the data are automatically resolved. Conclusions are postponed until
all available information is combined. Promise is target-oriented, meaning that only a limited depth
range is modeled. This range contains the reservoir plus a few layers above and below to take seismic
interference into account. The model is defined in depth while the seismic data are in time. A link
between them is a reference horizon that specifies the reflection arrival time of one of the dept horizons
in the model.
Models are layer-based, and each layer may have a variable thickness. The physical properties of a layer
are described using rock property models. Such a rock property model is a set of physics-based equations
describing the relationship between the elastic properties (Vp, Vs and density) and reservoir properties.
TLE Dec. 2009, 1478-1484; Seismically constrained probabilistic reservoir modeling by Leguijt Organised
TLE DEC 2009, 1474-1484
Probabilistic reservoir modeling - Promise
- Seismic constrained geological modeling
- Bayess rule is used to combine different sources of information
- Different model realizations are generated using the Metropolis algorithm
- From the realizations, expectation values and other statistics of various reservoir
properties are computed with the Monte Carlo simulation method.
( )r
Bayess rule makes it possible to combinerdifferent types of information and uncertain
r
p d m, H q m H
r
( )
( )
information:
r r m
p m d, H =
r
( )
pdH
m = vector with model parameters
r
d = vector with measurements (= samples of seismic traces + interpreted arrival times of seismic
s (m H ) = vector with measurement predictions
r r events)
qm (m H ) = prior distribution
r
(r r
)
p d m, H = likelihood
(r r
)
p m d , H = posterior distribution
H is an indication for the assumptions (hypotheses) that are made
The noise is assumed to be additive ( r r
) [ r r r
]
p d m, H = qn d s (m H )
Uncertainties in the scaling and wavelet phase can be taken care of with extra model
parameters. Organised
TLE DEC 2009, 1474-1484
Probabilistic reservoir modeling - Promise
The prior distribution contains all the information that can be obtained from the petrophysical
data and geological concepts.
The vector with measurements contains the samples (target oriented trace window) and the
interpreted arrival times of the seismic events.
The likelihood links the information that can be extracted from the measurements to the prior.
Noise is assumed to be additive.
The posterior distribution gives the probability density for each possible model, and as such,
the posterior distribution contains the full solution of the inversion problem.
The maximum of the posterior distribution can be found with an iterative scheme comparable
to a linear inversion method. The values of the model parameters at this maximum can be an
estimate of the posterior expectation values. The second-order derivatives around the maxi-
mum can be used to compute standard deviations of these estimates.
When the problem is strongly non-linear, however, as with seismic inversion, a better result
r
is obtained by generating an ensemble of model samples, i , i = 1, L , N } from the posterior
{ m
distribution using a sampling method such as the Metropolis algorithm. Expectation values
of reservoir properties are estimated using the Monte Carlo summation method,
r
FN = i =1 f (mi ) N
N
where FN is an unbiased estimator of the expectation value from which the property is defi-
r
ned by the function f (m ) .
Organised
Inversion and Reservoir Modeling
Seismic Well
Location B:
Evidence for hydrocarbon presence is very high
Location C:
Probability for finding hydrocarbons is reduced
relative to the prior
Seismic data and posterior facies probabilities for locations B and C. (Bayesian Monte Carlo method
for seismic predrill prospect assessment by Kjnsberg et al.)
Organised
Background models for inversion HorizonCube
Horizon mapping can improve seismic inversion!
Fig. 1. Modes of traditional model building. Sparse horizons (black) are interpreted manually. Based on
rules that can vary by packages, infill horizons are created for model building or interpretation purposes.
TLE May 2012, 546-554
Background models for inversion HorizonCube
Fig. 4. Principle of data-driven tracking of the HorizonCube using a dip field extracted from the seismic
reflectors.
Fig. 5. The difference between (left) conventional workflow and (right) horizon mapping in regard to not
only the quality of the model but also the quality of the acoustic impedance (AI) inversion.
TLE May 2012, 546-554
Background models for inversion HorizonCube
Horizon mapping was used to extrapolate the IP and IS logs to create a stratigraphically
consistent low-frequency model which incorporates the inversion workflow.
low-frequency
Fig. 9. IP background model before (left) and after (right) HorizonCube, with the latter showing a much
better consistency with the geologic structures.
Fig. 10. IP inversion results from the initial inversion to the final result. (left to right) P-impedance
obtained from prestack data and original background model; P-impedance from preconditioned data and
original background model; and P-impedance with preconditioned data and the HorizonCube-derived
background model.
TLE May 2012, 546-554