Professional Documents
Culture Documents
net/publication/309895138
CITATIONS READS
13 258
2 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Xinming Wu on 01 March 2017.
1
Simultaneous multiple well-seismic ties 2
a) b)
Figure 1: A seismic image is displayed with 11 velocity logs in (a). These velocity logs are displayed in time using
initial time-depth functions. A 3D image-guided nearest neighbor interpolation (b) of these velocity logs is laterally
discontinuous (highlighted by red arrows) along reflectors because the logs are not tied to the seismic image.
Baan, 2014; Herrera et al., 2014; Muñoz and Hale, 2015; from the real seismic image, to interpolate a similar syn-
Cheverry et al., 2015) propose to use dynamic time warp- thetic image. These methods often produce robust and
ing (DTW) or equivalent methods to better match the laterally consistent multiple well ties. However, they re-
synthetic seismogram with the real seismogram. DTW quire constructing a synthetic image with an image-guided
was first proposed by Sakoe and Chiba (1978) in the con- interpolation or RGT volume. Computing such an inter-
text of speech recognition, and is today widely used in polation or RGT volume is computationally expensive and
geophysics for correlating seismic traces (Anderson and is sensitive to the presence of faults and unconformities.
Gaby, 1983), seismic images (Hale, 2013), well logs (Smith In addition, one potential problem of these methods is
and Waterman, 1980; Wheeler and Hale, 2014), and mag- that the interpolated initial synthetic image is often lat-
netostratigraphic data (Lallier et al., 2013). Compared to erally inconsistent (e.g., as shown in Figure 1b) because
windowed cross-correlation methods, DTW is often more wells are initially misaligned with the seismic image.
accurate in computing relative shifts between series or im- To address this problem, we propose to first laterally
ages, especially when the shifts are large and rapidly vary- correlate synthetic seismograms and the corresponding
ing. real seismograms to obtain flattened synthetic and real
Although many methods have been proposed for all the seismograms, in both of which seismic events correspond-
three steps of well-seismic ties, most of them (e.g., White, ing to the same layers are horizontally aligned. We then
1980; White and Simm, 2003; Herrera and van der Baan, vertically match the flattened synthetic seismograms with
2014; Herrera et al., 2014) tie multiple wells to the seismic the flattened real ones using only vertically-variant but
image one-by-one and often have difficulties in maintain- laterally-invariant shifts, which is simply a 1D correla-
ing lateral consistency among multiple wells. To simulta- tion problem. With this two-step correlation method, we
neously tie multiple wells to a seismic image, Muñoz and are able to efficiently compute simultaneous multiple well-
Hale (2014, 2015) first use multiple synthetic seismograms, seismic ties and e↵ectively maintain lateral consistency
computed from the wells, to interpolate a synthetic am- among multiple wells because the events corresponding to
plitude image. Because the interpolation is guided by the the same layers in the synthetic seismograms are corre-
real seismic image as discussed by Hale (2010b), the struc- lated to those corresponding to the same layers in the real
tures of the interpolated synthetic image laterally conform seismograms.
to the real seismic image. They then match the synthetic The paper is organized as follows. We first briefly dis-
image with the real seismic image using smooth dynamic cuss computation of synthetic seismograms, DTW method,
image warping (Compton and Hale, 2014). Cubizolle et al. and single well-seismic ties. We then discuss in detail how
(2015) propose a similar method to compute simultaneous to compute simultaneous multiple well-seismic ties using
multiple well ties, but use a relative geologic time (RGT) our two-step correlation method. We finally demonstrate
volume (Stark, 2004; Wu and Zhong, 2012), computed the results of well-seismic ties by interpolating a 3D ve-
Simultaneous multiple well-seismic ties 3
a) b)
Figure 2: Initial (a) synthetic seismograms (red) are independently aligned (b) with real seismograms (black).
a) b)
Figure 3: Velocity logs are independently tied (a) to the seismic image using time-depth functions computed from single
well-seismic ties. A 3D image-guided nearest neighbor interpolation (b) of these logs is laterally more continuous than
the one in Figure 1b, but is still discontinuous at some positions (highlighted by red arrows), which indicates some logs
may not be correctly tied to the seismic image.
locity model from the tied velocity logs using an image- TYING WELLS INDEPENDENTLY
guided interpolation method (Hale, 2010b). In tying a single well to a seismic image, we begin with
choosing a wavelet, computing a reflectivity series in depth,
and generating an initial time-depth function. We then
generate an initial synthetic seismogram by convolving the
Simultaneous multiple well-seismic ties 4
reflectivity series with the wavelet delayed by the initial real seismograms are directly extracted from the seismic
time-depth function. We then use smooth DTW (Comp- image at locations closest to the wells. The synthetic seis-
ton and Hale, 2014) to compute vertical shifts that match mograms computed in this way are all zero-phase, which
the synthetic seismogram with the corresponding real seis- is likely inconsistent with the corresponding real seismic
mogram, and use these shifts to update the time-depth data. Therefore, we estimate a constant-phase rotation
function. We iteratively update the time-depth function to these synthetic seismograms using a DTW method de-
and synthetic seismogram until the updates are insignif- scribed by Muñoz and Hale (2012). As shown in Figure 2a,
icant to finally obtain the single well-seismic tie, as de- these synthetic seismograms (red), computed with the ini-
scribed previously by Muñoz and Hale (2015). tial time-depth functions (equation 2), do not match the
real seismograms (black). Next, we will discuss how to up-
Synthetic seismograms date the time-depth functions one-by-one for better single
well-seismic ties using smooth DTW (Compton and Hale,
The 3D seismic image and well logs that we use in this 2014).
paper are from the freely available Teapot Dome data set
(Anderson, 2009). In this paper, we choose only the well
logs with both velocity and density logs that are signif-
Dynamic time warping
icantly long in the same depth ranges, as shown in Fig- To match a synthetic seismogram f (⌧ ) with a real seismo-
ure 1a. gram g(t), we expect to shift the former by s(⌧ ) so that
From these velocity v(z) (Figure 1a) and corresponding they are approximately the same: f (⌧ + s(⌧ )) ⇡ g(t(⌧ )).
density ⇢(z) (not shown) logs, the reflectivity series can The shifts s(⌧ ) are often large and nonlinear (Herrera
be computed as follows by assuming a layered-earth model and van der Baan, 2014), and therefore are difficult to
(Sheri↵ and Geldart, 1995): estimate using windowed cross-correlation methods. The
v(z + z)⇢(z + z) v(z)⇢(z) DTW method, first proposed by Sakoe and Chiba (1978)
r(z) = , (1) in speech recognition, is a better method to estimate non-
v(z + z)⇢(z + z) + v(z)⇢(z)
linear and rapidly varying shifts (Hale, 2013; Herrera and
where the depth sampling interval in this case is z =0.1524 van der Baan, 2014; Muñoz and Hale, 2015). DTW corre-
m. An initial time-depth function for each well can be es- sponds to solving the following constrained optimization
timated from the velocity log of that well problem
Z z X
d⇠
⌧0 (z) = ⌧min + 2 , (2) min ||g(t(⌧ )) f (⌧ + s(⌧ ))||p
zmin v(⇠) s(⌧ )
⌧
, (4)
where ⌧min = 2zmin /v(zmin ). ds
subject to "min "max
The wavelet we choose for this example is simply a zero- d⌧
phase Ricker wavelet w(⌧ ) with 35-Hz peak frequency, where p > 0. We choose p = 2 corresponding to L2 norm
which is also the peak frequency of the real seismic data. for examples in this paper. We will discuss how to specify
A more accurate wavelet might be estimated from nearby the constraint bounds of ds/d⌧ in computing the shifts.
seismic traces by using a coherency matching technique One potential problem of the common DTW method
(White, 1980; Walden and White, 1998), or by rotating (e.g., Hale, 2013; Herrera and van der Baan, 2014; Herrera
and shifting a given zero-phase wavelet so that it best et al., 2014) is that the estimated shifts are limited to in-
matches the energy spectrum of the seismic data (White tegers, which may not be sufficient to accurately correlate
and Simm, 2003). For multiple well-seismic ties, it might the synthetic and real seismograms. We use the smooth
be desirable to account for spatial and temporal changes in DTW method, proposed by Compton and Hale (2014), to
seismic wavelets between wells as discussed by Ziolkowski compute smoothly varying shifts which are often more ac-
et al. (1998). However, we use a Ricker wavelet in this curate than those from the common DTW method, as sug-
paper to avoid potential uncertainties due to the wavelet gested by Muñoz and Hale (2015). As described in detail
estimation and also to demonstrate that our method is by Compton and Hale (2014), the smooth DTW method
robust even using a single simple wavelet for constructing solves the same minimization problem (equation 4) to first
all the synthetic seismograms. compute coarsely sampled shifts and then interpolate back
After choosing a zero-phase Ricker wavelet, we then smooth shifts for all samples.
construct a synthetic seismogram for each well by convolv- After computing the vertical shifts s(⌧ ) that match the
ing the reflectivity series (r(z)) with the wavelet (w(⌧ )) synthetic seismogram with the real seismogram, we can
delayed by the initial time-depth function (⌧0 (z)): then compute an updated time-depth function ⌧1 (z) from
Z zmax
a given initial time-depth function ⌧0 (z):
f (⌧ ) = r(z)w(⌧ ⌧0 (z))dz. (3)
zmin
⌧1 (z) = ⌧0 (z) + s(⌧0 (z)). (5)
Using this simple convolution method, we compute all the
synthetic seismograms that are denoted by red and over- These shifts s(⌧ ) correspond to squeezing and stretching of
laid with the black real seismograms in Figure 2a. These the synthetic seismograms in matching the seismic traces,
Simultaneous multiple well-seismic ties 5
and such squeezing and stretching can be unreasonably image-guided nearest neighbor interpolation (Hale, 2010b)
excessive as discussed by White and Simm (2003). To of the measurements. In this method, assume we have
avoid such excessive stretching and squeezing, the shifts a set of k known well-log values V = {v1 , v2 , · · · , vk }
are constrained to satisfy some specified bounds on the (vk 2 R) that are spatially scattered at corresponding
time-derivatives ds/d⌧ of the shifts in the smooth DTW k known locations X = {x1 , x2 , · · · , xk }, we then com-
method (equation 4). As suggested by Muñoz and Hale pute an image-guided nearest neighbor interpolation of
(2015), we can relate the time-derivatives ds/d⌧ to the the known values by solving the following anisotropic eikonal
initial and updated velocity functions v0 (z) = 2dz/d⌧0 equation (Hale, 2009)
and v1 (z) = 2dz/d⌧1 , respectively. From equation 5, we
have rt(x) · D(x)rt(x) = 1, x2/ X;
d⌧1 d⌧0 ds d⌧0 t(x) = 0, x 2 X, (9)
= + , (6)
dz dz d⌧0 dz
which is equivalent to where x = (x1 , x2 , x3 ) represent 3D spatial coordinates
ds v0 (z) within the 3D seismic image (Figure 1a), t(x) is a map of
= 1. (7)
d⌧0 v1 (z) non-Euclidean distance (Hale, 2009) from x to the nearest
known sample xk . When solving the above anisotropic
This means that we can specify bounds on v0 (z)/v1 (z)
eikonal equation for the minimal-distance map t(x), it
to impose corresponding constraints on the shifts (Muñoz
is straightforward to simultaneously obtain the nearest
and Hale, 2015):
✓ ◆ ✓ ◆ neighbor interpolant (Hale, 2009). A known sample xk
v0 (z) ds v0 (z) is nearest to a point x only if the non-Euclidean distance
1 1. (8)
v1 (z) min d⌧0 v1 (z) max t(x) is less than that for any other known sample point.
The metric tensor field D(x) represents the coherence
The updated velocity v1 (z) should have a similar trend and orientation of seismic structures, and therefore of-
as the initial velocity v0 (z) (measured well-log velocity), ten provides anisotropic and spatially variant coefficients
which provides a qualitative way to determine the bounds for the eikonal equation. As discussed by Hale (2010a),
on the velocity ratio v0 (z)/v1 (z) by qualitatively evaluat- we construct the metric tensor field using structure ten-
ing the reasonableness of updated velocity comparing to sors (Van Vliet and Verbeek, 1995; Fehmers and Höcker,
the initial velocity (Muñoz and Hale, 2015). In this way, 2003) computed from the 3D seismic image. With such
we are seeking for shifts that not only optimally match the a metric tensor field, the computed nearest neighbor in-
synthetic and real seismograms but also reasonably up- terpolant conforms to structures apparent in the seismic
date the velocity. In all examples in this paper, we choose image.
(v0 (z)/v1 (z))min = 0.9 and (v0 (z)/v1 (z))max = 1.2. If the well logs are correctly aligned with the seismic
Errors and uncertainties may exist in computing the ini- image, the image-guided nearest neighbor interpolation
tial time-depth functions for constructing synthetic seis- of well-log measurements should be laterally continuous
mograms. Therefore, the updated time-depth function along the seismic reflectors. We observe that the interpo-
(equation 5) should be used to compute a new synthetic lation (Figure 3b) of the velocities after single well-seismic
seismogram, which should be again correlated to the real ties (Figure 3a) is laterally more consistent than the one
seismogram to compute new possible shifts. This means before tying (Figure 1b). However, some lateral disconti-
that the whole process should be iteratively performed nuities (highlighted by red arrows in Figure 3b) are still
until updates become negligible and a stable calibrated apparent in the interpolant, and these discontinuities do
time-depth function is obtained. not coincide with geologic faults. This indicates that er-
Using this iterative process, we update time-depth func- rors exist in the single well-seismic ties without any lateral
tions one-by-one and independently match the synthetic control. Next, we will discuss how to efficiently compute
(red) seismograms with the real (black) seismograms (Fig- simultaneous multiple well-seismic ties to improve the lat-
ure 2b) one-by-one. We observe that the synthetic seismo- eral consistency among these multiple wells.
grams vertically match the real seismograms, which indi-
cates that the smooth DTW method successfully finds the
optimal vertical match for each pair of synthetic and real TYING WELLS SIMULTANEOUSLY
seismograms independently. However, an optimal match The idea of tying multiple wells simultaneously was pre-
of a synthetic seismogram and the corresponding real seis- viously proposed by Muñoz and Hale (2015) and Cubi-
mogram does not necessarily guarantee a reliable well- zolle et al. (2015), who suggested to first extrapolate the
seismic tie, because errors and noise may exist in both synthetic seismograms away from well locations, follow-
synthetic and real seismograms. ing the real seismic reflections, to compute a synthetic
Using the updated time-depth function computed for image. Then they matched the synthetic image with the
each well, we can independently tie the velocity logs to the corresponding real seismic image to compute simultaneous
seismic image as shown in Figure 3a. One way to check multiple well-seismic ties. However, constructing a 2D or
for possible errors in well-seismic ties is to extend the well- 3D synthetic image and matching it with the real seis-
log measurements along seismic reflectors to compute an mic image significantly increases the computational cost.
Simultaneous multiple well-seismic ties 6
a) b)
a) b)
Figure 5: Real seismograms (nearest to the well locations) before (a) and after (b) flattening.
In addition, the initial synthetic image, interpolated with thetic seismograms with the flattened real seismograms
uncorrelated synthetic seismograms (or well logs), is of- using only vertically-variant but laterally-invariant shifts,
ten laterally discontinuous, like the interpolated velocity which is simply a 1D correlation problem.
image shown in Figure 1b.
We propose to first compute laterally correlated (flat-
tened) synthetic and real seismograms, in both of which
all events corresponding to the same geologic layers are
horizontally aligned. We then match the flattened syn-
Simultaneous multiple well-seismic ties 7
a) b)
Figure 6: Flattened synthetic seismograms (red) and real seismograms (black) before (a) and after (b) alignment.
a) b)
Figure 7: The aligned synthetic (red) and real (black) seismograms in RGT (Figure 6b) are mapped back to two-way
time to obtain a simultaneous match of the multiple real and synthetic seismograms (a). These simultaneously matched
synthetic seismograms (colored by red) are overlaid with those independently matched ones (colored by green) in (b).
a) b)
Figure 8: Velocity logs are simultaneously tied (a) to the seismic image using time-depth functions computed from
simultaneous multiple well-seismic ties. A 3D image-guided nearest neighbor interpolation (b) of these logs is laterally
consistent along seismic reflectors.
where uf (I, z) = ⌧ (I, z) + sf (I, ⌧ ). Therefore, the time- mograms with the flattened real seismograms to efficiently
depth functions t(I, z) are computed by compute simultaneous multiple well-seismic ties. This
two-step approach should only be applied when the lat-
t(I, z) = t(I, ug (I, z)), (13) eral correlations of the synthetic and real seismograms are
trustworthy.
where ug (I, z) = ⌧ (I, z) + sf (I, ⌧ ) + su (⌧ + sf ). We believe this methodology could also work with geo-
As in the case of single well-seismic ties, we want to logically complex settings (growth faults, rapid layer thick-
iteratively update the time-depth functions t(I, z) using ness variations, and late diagenetic units). However, the
equation 13. With the finally updated t(I, z), we can di- proposed correlation of seismic traces and synthetic traces
rectly compute synthetic seismograms that match the real would likely yield inaccurate results in these cases. To
seismograms in two-way time, as shown in Figure 7a. In improve seismic trace correlation, an alternative strategy
Figure 7b, we display all the aligned synthetic seismo- could then be to better exploit imaging information by
grams computed using single (green traces) and simul- flattening the whole seismic image (e.g., Lomask et al.,
taneous (red traces) well-seismic tying methods together 2006; Fomel, 2010; Wu and Zhong, 2012; Wu and Hale,
with the real seismic traces (black ones). We observe the 2015a,b, 2016), instead of flattening only the seismograms
alignments computed using the two methods are almost extracted at well locations. To improve well correlations,
the same for the long seismograms, but are significantly one could also replace our DTW-based synthetic trace cor-
di↵erent for the short ones, especially the 3rd, 5th, 7th, relation by expert-based manual correlation or correlation
10th, and 11th seismograms. using various logs or rules (Lallier et al., 2012, 2016).
With the finally updated time-depth functions t(I, z), In the case of rapidly varying structures in space, sin-
we tie all the velocity logs to the seismic image, as shown gle well-seismic ties may be more reliable than the lateral
in Figure 8a. A 3D image-guided nearest neighbor inter- correlations of synthetic seismograms or well logs. In such
polation (Figure 8a) of these tied velocity logs is laterally case, it may help to use well-seismic ties to improve well-
more continuous along seismic reflectors than the one in log correlation (or synthetic seismogram correlation), in-
Figure 3b. This suggests that the simultaneous multiple stead of using the latter to improve the former as proposed
well-seismic ties (Figure 8b) are more consistent than the in this paper. The general idea in this case would be to
single well-seismic ties (Figure 3a). first match well logs with the seismic image one-by-one us-
ing any single well-seismic tying method (e.g., Herrera and
DISCUSSION van der Baan, 2014; Herrera et al., 2014; Muñoz and Hale,
We have proposed to first laterally correlate the synthetic 2015; Cheverry et al., 2015). These well logs could then
and real seismograms to obtain flattened seismograms. be laterally correlated by the large scale structure trend
We then vertically matched the flattened synthetic seis- computed via seismic image flattening (e.g., Lomask et al.,
Simultaneous multiple well-seismic ties 10
2006; Fomel, 2010; Wu and Hale, 2015a,b). ACKNOWLEDGMENTS
In this scenario, a seismic image would only provide a In much of the research described in this paper, we ben-
low frequency structure trend for well-log correlation be- efited greatly from discussions with Dr. Dave Hale. We
cause of the uncertainties involved in seismic imaging pro- appreciate suggestions by J. Kim Welford, Roar Hegg-
cess itself, in computing well-seismic ties, and in seismic land, and two anonymous reviewers that led to significant
image flattening. Still, such trend can be used to reduce revision of this paper. This research is supported by the
ambiguities in automated stochastic well correlation prob- sponsor companies of the Consortium Project on Seismic
lems, as proposed by Julio et al. (2012). One might use Inverse Methods for Complex Structures at the Colorado
this low frequency trend to first compute an initial corre- School of Mines. The Teapot Dome seismic and well data
lation of the well logs, and then use a well-log correlation were provided by the Rocky Mountain Oilfield Testing
method (e.g., Wheeler and Hale, 2014) to further tune Center.
the correlation to reflect high-frequency information that
is measured in well logs but not in seismic data.
REFERENCES