2003 EAGE

75
tutorial
ell ties are a very important part of the interpreters
trade. They provide a means of 1. correctly identi-
fying horizons to pick, and 2. estimating the
wavelet for inverting seismic data to impedance.
Just as well ties are paramount in the calibration of a seismic
interpretation, so too they are the cornerstone of using seismic
amplitudes in impedance and AVO inversion, and ultimately
of inferences fed into the risking process. Figure 1 shows an
example where the geologically relevant amplitude informa-
tion is highly focussed in a particular seismic loop (the blue
loop in this case). The wavelet is ~30 degrees rotated from
symmetry and its main loop is 24 ms deeper in time than time
zero. This is nominally minimum phase data displayed with
European polarity, i.e. with negative values corresponding to
compressions and plotted as troughs. Given this convention,
together with the fact that the red loop is closest to the check-
shot time, some interpreters might expect that the red pick is
the one to make. If you are after the amplitudes that map out
the sedimentology, then it clearly isnt. Once you know the
wavelet, it is evident that the wavelet energy is concentrated in
the blue loop and this loop is delayed relative to the checkshot
time. Readers are referred to our previous tutorial article
(Simm and White 2002) for a discussion of phase, timing,
polarity and the interpreters wavelet and their impact on seis-
mic interpretation.
What is important in this example is that both the wavelet
shape and timing were estimated without making assumptions
about the wavelet (i.e. what the wavelet should look like) or
the timing (i.e. which loop represents the top of the reservoir).
The subject of this paper then is the process by which
wavelets are estimated through a well tie procedure that results
in quantitative measures of synthetic to seismic goodness-of-t
and likely wavelet accuracy. In the authors view this is a good
practice approach, which should form at the very least the ini-
tial stages of a well tie study. Activities such as stretch and
squeezing a synthetic to t a seismic trace are unscientic and
denitely not good practice! Regrettably, while stretch and
squeeze is a common feature of well tie software, q.c. features
are at best rudimentary.
The well tie process
Making a well tie (Figure 2) is deceptively simple. A synthet-
ic seismogram is matched to a real seismic trace and features
from the well are correlated to the seismic data. The prime
concept in constructing the synthetic is the convolutional
model. This represents a seismic reection signal as a
sequence of interfering reection pulses of different ampli-
tudes and polarity but all of the same shape. This pulse
shape is the seismic wavelet, formally the reection wave-
form returned by an isolated reector of unit strength at the
target depth. Because reecting boundaries are spaced much
more nely in depth than the length of the reected pulses,
the degree of interference is generally severe and only the
strongest reectors or reection complexes stand out in the
reection signal.
Well tie procedure
In outline the procedure for tying a well-log synthetic seismo-
gram to seismic data comprises the following steps (Figure 3).
1. Edit and calibrate the sonic and density logs.
2. Construct the synthetic seismogram from the calibrated
well-logs:
a. choose the appropriate reection series (usually primar-
ies only),
b. construct the reection series in two-way time.
3. Perform the match, comprising
a. determine the best match location,
b. estimate the wavelet and its accuracy.
rst break volume 21, October 2003
Tutorial: Good practice in well ties
Roy White
1
and Rob Simm
2
1.
School of Earth Sciences, Birkbeck, University of London ([email protected])
2.
Rock Physics Associates Ltd. ([email protected])
W
Figure 1 Making the wrong pick can make a difference to
geological interpretation: a) horizon picks and amplitude
maps, b) the wavelet
FB october v5 18-09-2003 17:20 Pagina 75
2003 EAGE
76
tutorial
In some situations there is also the issue of comparing
wavelets estimated from different wells in the survey area
and combining them if appropriate.
Log calibration
In the interests of brevity we shall omit any discussion of
editing and conditioning the logs. This is not to deny its
importance but it is too big a topic to be satisfactorily
reviewed here.
Well log calibration brings the timing of the sonic log into
agreement with seismic times from a checkshot survey or
Vertical Seismic Prole (VSP). Practical and theoretical
analysis of the factors that inuence the accuracy of well ties
(White 1997) shows that timing is paramount. Timing
errors in the synthetic seismogram are much more detrimen-
tal than amplitude errors. They are especially damaging to
estimating correctly the higher frequency components of the
seismic wavelet. Well-log calibration establishes the timing
of the synthetic seismogram. Without it, the tie can only be
fudged.
Calibration obviously starts with picking checkshot
times. Current practice with regard to picking checkshot
times seems to favour trough-to-trough picking of VSP times
in which times are picked from the rst trough of the signal
on the source hydrophone to the rst trough of the down-
hole signal. Some geophysicists have advocated picking
times from the onsets of these signals. A correction for dis-
persion-attenuation may then be necessary as the onsets
tend to travel at the speed of the highest frequency compo-
nents in the seismic waveform. The rst-order effect of dis-
persion-attenuation in the earth at common target depths is
a slight stretching of travel times and trough picking can be
regarded as a way of compensating for this effect (White
1998).
Nowadays the interpretation geophysicist is likely to
receive either calibrated sonic and density logs or times
picked and corrected by the VSP contractor. In the latter
case, a calibration (or drift) curve is tted to the difference
between the integrated sonic log times and the corrected VSP
times, as illustrated in Figure 4. In this case the calibration
uses straight line segments between knee points. Knee points
have to be chosen at major jumps in the sonic log so as to
avoid introducing articial or abnormal reection coef-
cients. A popular alternative is a smooth spline t which pro-
vides smoothly varying corrections to the log. Spline calibra-
tion requires a choice of spline knots and this can be auto-
mated. Nonetheless calibration points often show sudden
bends at unconformities and changes in character of the logs
and knee points allow direct control of the calibration. In
practice the density of VSP recordings usually offers very lit-
tle scope to vary the calibration curve signicantly and any
reasonable t to this many points generates a satisfactory
timing.
The drift curve is applied to the sonic log which is then
integrated to give a time-depth relationship that ts the
checkshots to within 1-2 ms. The picking accuracy of VSP
times is of the order of 1 ms and the calibration should con-
strain the timing of the synthetic seismogram to a similar
accuracy. Stewart et al (1984) discuss the causes of different
types of drift and their physical explanation.
Constructing the synthetic seismogram
Constructing the synthetic involves creating a reection
series in time and convolving this with a wavelet. There is
also the question as to whether to use a primaries-only
reectivity or whether to include internal multiples or even
to allow for other multiples. In most cases it is the correla-
rst break volume 21, October 2003
Figure 2 The components of a well tie. Time zero of the
wavelet is indicated by the horizontal line.
Figure 3 Well tie procedure, indicating the key parameters
required by the process.
FB october v5 18-09-2003 17:20 Pagina 76
2003 EAGE
77
tutorial
tion of the data with the primary reectivity that interpreta-
tion requires and this usually produces a satisfactory well tie.
a) Creating a reection series in time
There is no single approach to creating a broadband reection
series in time. Some modern software gives the option of either
using a velocity log with the effect of drift applied (i.e. the sonic
log values are effectively changed) or using the original sonic
values together with the time-depth relationship constructed
from applying the drift curves to the sonic log (i.e. the depths are
relabelled in time). The conversion to time obviously involves a
re-sampling step. In order to mitigate any aliasing effects the ini-
tial sampling gives a sample rate ner than the initial depth sam-
pling. These days there is no reason why the synthetic trace can-
not be computed from this nely sampled log and a wavelet
with the same sample rate. However the synthetic seismogram
generally benets from some blocking of the logs. In the past,
limited computing power may have dictated further sample
coarsening, commonly to match the seismic sample rate.
This coarsening of the sample rate is effectively a time aver-
aging process. Theoretically, in order to account for the effects
of velocity dispersion between sonic and seismic frequencies the
sonic logs need to be averaged using the Backus average rather
than the time-average approach (Marion et al 1994, Rio et
al.1996). In this approach velocities and densities are used to
derive the bulk and shear moduli, which are then averaged using
the harmonic average. Velocities are re-constituted from the
averaged moduli and the arithmetically averaged density log.
Experience has shown that in sand/shale environments there is
very little difference between well ties made using the time-aver-
age and the Backus average approach. It is usually safe to block
to a ne sample interval (0.5 ms say) and then resample the syn-
thetic to the required sample interval. If there are large contrasts
in velocity then the Backus average should be used and in this
case it may pay to optimise the blocking length.
For well ties on angle stacks, reection series need to be
computed for a range of angles of incidence or recording off-
sets. Converting the logs to elastic impedance (Connolly 1999)
offers a quick and convenient way of computing angular reec-
tivities using standard normal incidence code. This procedure
becomes seriously inaccurate on going from siliciclastic to car-
bonate lithologies or at any sudden change in the background
V
s
/V
p
ratio. Alternatively computing angle or offset-dependent
primary reection series is a straightforward task.
b) Estimating the wavelet to be convolved with the reection
series
The broadband reectivity has to be ltered by the seismic
wavelet in order to tie the well to the data. A problem in estimat-
ing the wavelet is that the transmission response of the earth has
to be included if the ltered synthetic seismogram is to match the
seismic data. A matching process is therefore required. There are
a number of approaches to estimating the wavelet:
rst break volume 21, October 2003
Figure 4 An example of well-log calibration from the central
North Sea. Left: uncalibrated sonic log; centre: integrated log
time minus check-shot time tted using knee points; right:
calibrated sonic log. The slopes of the calibration curve (cen-
tre panel) provide the correction factors subtracted from the
uncalibrated log.
Figure 5 The trade off between bias (distortion) and noise-
induced errors in matching a synthetic seismogram and seis-
mic trace. The top panel shows wavelets extracted from the
data shown in Figure 6 using various maximum correlation
lags; the dashed vertical lines mark zero time. The centre
panel plots the computed NMSE of these wavelets against
maximum lag; the solid curve is a theoretical t to these val-
ues using a N/S ratio of 0.15; the dashed curve is for a N/S
ratio of 0.5. The bottom panel plots the predictabilities (P)
tted by a smooth curve.
FB october v5 18-09-2003 17:20 Pagina 77
2003 EAGE
78
tutorial
1. Forward modelling approach
One approach is to build a model of the seismic wavelet
from the response of the acquisition and processing system,
adding perhaps a constant-Q lter to represent attenuation
in the earth. Constant-Q attenuation is at best a simple
approximation to the characteristic amplitude decay with
frequency from transmission through the earth and it may
or may not approximate the concomitant phase and timing
response. Ghosting, the effect of processing and an off-the-
shelf source signature inject further signicant uncertainties
into this approach. After stacking and migration the wavelet
will no longer be a causal or even a physical wavelet.
Without some further adjustment, the well tie is likely to be
crude and sub-optimal.
2. Phase rotation of zero phase synthetic
A more straightforward alternative is to construct a zero-
phase wavelet that matches the energy spectrum of the data
window and phase rotate and time shift it to optimise the
match. This can be a very effective way of making a charac-
ter match of synthetic to seismic. The reason it works is
that, after processing, the phase of the seismic wavelet is
usually fairly linear across the seismic bandwidth.
Deviations from linearity at the edges of the bandwidth are
likely to matter only if the wavelet is to be used for inver-
sion. Owing to differences in spectral character between
Ricker wavelets and real seismic data, Ricker wavelets
should not be used for the initial zero phase synthetic
(Hosken 1988).
3. Wavelet estimation through matching
The method that is recommended here and that is described
in the following sections is to estimate the wavelet through a
coherency matching technique (White 1980,Walden and
White 1998). This procedure gives a number of outputs that
effectively dene whether the tie is good or not.
a. the shape of the wavelet
b. the phase characteristics of the wavelet
c. quantitative measures related to the tie
i. an estimate of the goodness-of-t of the synthetic to
seismic
ii. an estimate of the likely phase error (or accuracy) of the
wavelet
Measuring the tie: goodness of t and accuracy
The distinction between goodness-of-t and accuracy is a
very important one in assessing the reliability of the tie. In
order to dene goodness-of-t and accuracy, we introduce
two terms:
1. the energy of a trace is the sum of squares of a segment of
a time series
2. the residuals are the difference between a seismic trace
and its matched or ltered synthetic seismogram.
A simple measure of goodness-of-fit is the proportion of
trace energy predicted by the synthetic seismogram, which
we can call the predictability P:
P = 1 (energy in the residuals/trace energy).
The correlation coefcient R is another goodness-of-t
measure but implicit in it is the assumption of an error-free
synthetic seismogram. If the assumption holds, then P = R
2
.
Many statisticians consider R
2
a less deceptive measure of
correlation than R.
rst break volume 21, October 2003
Figure 7 Contours of goodness-of-t from matching a 2.7-
3.2s segment of well-log synthetic seismogram with a cube of
data centred on the well. The goodness-of-t is measured by
P, the proportion of trace energy in the seismic trace predict-
ed by the match, which is approximately the square of the
cross-correlation coefcient between the synthetic seismo-
gram and data.
Figure 6 The well tie using the 124ms wavelet in Figure 5.
Two copies of the synthetic seismogram are spliced into the
seismic data at traces 9 and 10. The best matching trace is
repeated on each side of the synthetic seismogram at traces 7
and 12. The best matching trace was found by scanning a
cube of data around the well (Figure 7).
FB october v5 18-09-2003 17:20 Pagina 78
2003 EAGE
79
tutorial
To measure accuracy we start with the normalised mean
square error in the synthetic seismogram:
NMSE = (energy in the errors in the synthetic)/(energy in
the synthetic)
Since the true synthetic is an abstraction, we cannot
know what the errors are but estimation theory tells us
what the expected NMSE is, namely:
T is the duration of the matched data segment and b is
called the analysis bandwidth (White 1980). bT is dimen-
sionless. In frequency domain terms, b is the bandwidth of
a moving spectral window within which the frequency
response is analysed. In time domain terms, b is a constant
divided by the maximum correlation lag employed in the
analysis. The maximum lag is a proxy for wavelet length.
Thus the NMSE is proportional to the ratio of wavelet
length to data segment length. (1-P)/P approximates the
ratio of misfit energy to fitted energy; and it depends on the
S/N ratio of the seismic data and the accuracy of the syn-
thetic model built from the calibrated well logs.
Simulations show that this approximation measures the
NMSE in the synthetic seismogram rather well. If the reec-
tivity spectrum is fairly at, it also estimates the NMSE in
the wavelet. However it measures only the mist from noise
in the seismic data and errors in the synthetic. If the analy-
sis bandwidth b is too broad, or equivalently the maximum
lag is made too short to accommodate the wavelet, the esti-
mated wavelet will be distorted. On the other hand, if the
maximum lag is too long, the tie will be overtted, match-
ing seismic noise as well as reection signal.
The NMSE is a somewhat crude measure of wavelet
accuracy. Error bars on the amplitude and phase spectra of
the wavelet are more informative. These are illustrated later
in the section on well ties and seismic data quality. A rough
idea of the phase error within the seismic bandwidth can be
found from
This equation also approximates the relative standard error
in the amplitude spectrum.
A practical example
Figure 5 illustrates a range of wavelets which have been
extracted from the data shown in Figure 6. At a wavelet length
of 52 ms the wavelet is compressed and has limited side lobe
energy whereas at 152 ms the wavelet has well developed side
lobes with secondary oscillations. As wavelet length increases
so does the goodness-of-t (bottom panel, Figure 5). The accu-
racy of the wavelet as measured by the NMSE is shown by the
points in the middle panel of Figure 5. The solid curve in this
panel is a theoretical estimate of the NMSE when wavelet dis-
tortion is included, for a noise-to-signal power ratio of 0.15,
corresponding to a predictability P of 0.87 (R0.93). The min-
imum in the curve is very at and any choice of maximum lag
(or analysis bandwidth) that avoids distortion yields a satisfac-
tory wavelet estimate. The synthetic seismogram in Figure 6
used the wavelet from the 124 ms maximum lag.
NMSE increases for wavelets shorter than the optimum
maximum lag owing to the effects of distortion or bias. It
increases for wavelets longer than the optimum through the
inclusion of noise into the wavelet estimation. In practice since
distortion rises rapidly at low maximum lags, it pays to err
towards larger not smaller values. Lengthening the wavelet
may make the side lobes look noisier but bias distorts the main
lobe which is more serious since it carries most of the wavelet
energy. Unfortunately simply inspecting a wavelet estimate does
not necessarily reveal that it is seriously distorted, although
there may be indications from inspecting its spectrum. If
wavelet length is a problem, there is no substitute for simula-
tion. Automated criteria which take no account of spectral dis-
tortion, such as the Akaike Information Criterion, do not opti-
mise wavelet length (Walden and White 1984). The dashed
curve shows how the minimum in the total NMSE would move
to a smaller maximum lag if the N/S ratio increased to 0.5.
The mechanics of matching: nding the best match
location
Matching is usually performed on time migrated seismic
data. A volume of data around a well is scanned for the best
match location. For various reasons, the best match often
does not occur at the well location. Since velocity typically
increases with depth, time migration commonly moves the
best match location up-dip from the well. The effect of
imperfect migration velocities is less predictable and it is even
possible that the well location is suspect. During the scanning
rst break volume 21, October 2003
Figure 8 Contour plot of two-way time to a prominent reec-
tion at ~2.5 s from the cube of data scanned in Figure 7. The
up-dip direction is from NE to SW.
FB october v5 18-09-2003 17:21 Pagina 79
2003 EAGE
80
tutorial
procedure the choice of analysis parameters in estimating the
wavelet can be investigated.
The best match time location is found by scanning a
number of traces around the well for goodness-of-t using,
say, a 500 ms segment of synthetic seismogram. Over such a
segment it can be generally assumed that phase is not vary-
ing signicantly. Reliability of the procedure deteriorates
rapidly if shorter segments are used. When more than 500 ms
of calibrated sonic is available the best time gate in which to
establish the tie should also be investigated. Figure 7 shows
contours of goodness-of-t from matching a 2.7-3.2 s seg-
ment of well-log synthetic seismogram with a cube of data
centred on the well. The best match location is about 100 m
SW of the well. A contour plot of two-way time to a promi-
nent reection above the matched time gate (Figure 8) con-
rms that the overlying strata dip upwards from NE to SW.
Good well ties reveal this up-dip pattern. When tying wells
to poor seismic data the best match location may be pulled
around by variations in S/N ratio. The well tie at the best
match location of Figure 7 is shown in Figure 6.
Well ties and seismic data quality
The quality of the seismic data has an impact on our con-
dence in well ties and wavelet extractions. This is illustrated
with regard to a well which has sections of good and poor
ties. Starting with the zone of poor data, Figure 9 shows two
possible ties in a time section from 2.5-2.9 s. Visually there
is nothing to choose between them nor can any measurement
of goodness-of-t help. For example, the correlation coef-
cients between the nearest data trace and the synthetic seis-
mograms are 0.796 and 0.790. Yet seismic wavelet 1 has a
36 degree phase advance and wavelet 2 a 48 degree phase
rst break volume 21, October 2003
Figure 10 Left: Log likelihood function, measuring goodness-of-t, from estimating timing and phase for the well tie of
Figure 9. The axes are time and phase advance in ms and radians respectively; i.e. negative values are lags. The maximum
likelihood estimates are a time lag of 1.0+/-3.5 ms and a phase lag of -0.46+/-0.66 radians. The third contour from the cen-
tre marks the 90% condence zone. The correlation coefcient between timing and phase is 0.98.
Right: phase spectrum (solid curve) with standard errors (dashed curves) of the seismic wavelet estimated from the well tie
for the 2.5-2.9 s time window shown in Figure 9. The intercept and slope of the sloping straight line give the maximum like-
lihood best t of time and phase.
Figure 9 Which is the better well tie? Two equally good well
ties from a time interval showing poor seismic bandwidth.
The well ties were constructed from a calibrated impedance
log using the two wavelets shown.
FB october v5 18-09-2003 17:21 Pagina 80
2003 EAGE
81
tutorial
lag, a difference of almost 90 degrees. Their timing differs by
8 ms. In seismic inversion an error of 90 degrees in phase will
reveal a step in impedance as a thin bed and vice versa.
The problem stems from the 400 ms segment of data. Its
spectrum and that of the well-log reectivity segment both
show a spectral peak at 28 Hz. The data bandwidth is just
19 Hz. This bandwidth is inadequate for determining timing
and phase accurately. Figure 10 (left) illustrates the ambigu-
ity. It shows the log likelihood function, a weighted measure
of goodness of t, from estimating the phase and timing of
this data segment. Timing and phase estimates are negative-
ly correlated, and strongly so: the correlation coefcient
between them is always in the range 0.866 to 1.0. Timing
and phase combinations represented by the central red zone
will give a reasonable tie. The lower the data bandwidth the
more elongated is this zone.
Figure 10 (right) illustrates another way of looking at the
problem. Estimated phase spectra and their error bounds from
which the phase and timing are computed are shown. In the
frequency domain the slope of the phase spectrum determines
the timing and the intercept at zero frequency the phase. There
is clearly a number of different ts (slope and intercept) to the
data that can be made. This ambiguity of timing and phase is
similar to the anti-correlation of intercept and slope also seen
in AVO intercept-gradient crossplots (Simm, White & Uden
2001) and velocity-depth analysis (Al-Chalabi 1997).
The data with good bandwidth (2.9-3.4 s) are shown in
Figure 11. The tie is clearly a good one and the effect of the
rst break volume 21, October 2003
Figure 11 Left: log likelihood function from estimating timing and phase for the well tie over the 2.9-3.4 s time window
shown below. The axes are time and phase advance in ms and radians respectively; i.e. negative values are lags. The maxi-
mum likelihood estimates are a time lag of 1.6+/-1.0 ms and a phase lag of 0.31+/-0.18 radians. The third contour from the
centre marks the 90% condence zone. The correlation coefcient between timing and phase is -0.92.
Right: phase spectrum (solid curve) with standard errors (dashed curves) of the seismic wavelet estimated from the well tie
in the 2.9-3.4 s time window. The intercept and slope of the sloping straight line give the maximum likelihood best t of time
and phase.
FB october v5 18-09-2003 17:21 Pagina 81
2003 EAGE
82
tutorial
broader bandwidth in resolving the intercept and slope is evi-
dent.
This example illustrates that the quality of a well tie
depends primarily on the quality of the data seismic and log.
One cannot expect to make good well ties from poor data.
The key seismic indicators are rst of all bandwidth, and then
the S/N ratio and duration of the data available for making a
tie. The example also illustrates a distinction between good-
ness-of-t and accuracy: the ties of Figure 9 look reasonable
but their accuracy is poor. The eye can judge goodness-of-t
fairly well but accuracy needs to be measured.
Comparing wavelets
Is the seismic wavelet from well A in Figure 12 different from
that from well B? At a supercial level they may appear to be
but the question cannot be answered reliably without knowing
how accurate the wavelets are. One simple test would be to
compare the normalised mean square difference between the
two wavelets with their estimated NMSE. Comparing their
amplitude and phase spectra is more informative. The ratio of
their amplitude spectra should hover around 1.0 within expect-
ed error limits (and it is in this case). Similarly the difference of
their phase spectra should be zero, again within error bounds.
There are two qualications to this: rst, the phase difference
must be unwrapped to remove spurious 2 jumps and second,
it may show a linear trend due a time difference. The error
bounds for a ratio and a difference can be readily calculated
from the individual standard errors.
The lower panel in Figure 12 shows the phase difference
between wavelets A and B. The dashed lines show the standard
errors. Standard errors correspond to 68% condence bounds;
i.e. roughly two thirds of the estimates are expected to fall with-
in these bounds. The light line in this panel is the best t
straight line through the origin of the phase difference within
the 8-50 Hz seismic bandwidth. It is within the error bounds
for 65% of that bandwidth. A regression of the phase differ-
ence against frequency within the 8-50 Hz seismic bandwidth
has a phase intercept is 22 22, making it doubtful whether
the wavelets really differ in phase.
Overtting always makes wavelets appear different since
noise has been propagated into them. Measuring their accura-
cy puts the differences into perspective.
rst break volume 21, October 2003
Figure 12 Comparing the phase response of two seismic wavelets. The top panel shows (left) wavelet A extracted from the
data in Figure 6 and (right) wavelet B extracted from the data shown in Figure 3 from the same 3D survey. The phase spec-
trum of wavelet B is shown in Figure 11. The bottom panel plots the unwrapped phase difference between the wavelet phase
spectra: the dashed lines indicate the standard error bounds; the light solid line is a best t through the origin of the phase
difference within the 8-50 Hz seismic bandwidth.
FB october v5 18-09-2003 17:21 Pagina 82
2003 EAGE
83
tutorial
When comparison shows wavelet differences are within
error bounds, the estimates can be combined, possibly through
a weighted average, into a survey-wide wavelet. This is especial-
ly benecial where no one wavelet stands out as the most accu-
rate. It is very important that the combination does not blur the
wavelet through small misalignments during averaging and this
is most conveniently achieved in the frequency domain.
Comments on stretch and squeeze
One common bad practice is the stretching and squeezing of
the synthetic seismogram to beautify the tie without any con-
rmatory evidence for the time shifts applied. The time shifts
involved are often implausible on grounds of both geophysi-
cal reasoning and timing accuracy (White 1998). For exam-
ple, any geophysical explanation of a time shift of just 2 ms
would exhibit other measurable effects. The structural dip,
lateral velocity variation or attenuating layer that produced
this time shift would be unmissable in the seismic data.
Stretch and squeeze may indicate a problem but it is not a
valid remedy.
What do you do when all the ties are bad?
When a well tie is poor, it is difcult to offer any prescription
for improving it beyond the general one of going back to q.c.
the seismic data and calibrated logs. In general effort spent on
practical details of the logs and seismic data are more likely
to improve a well tie than niceties such as the time variance of
the seismic wavelet. Time variance can be built into the syn-
thetic model or, more dangerously for overtting, into the
matching procedure. Any improvement in the t can be meas-
ured and tested. Our experience is that many perceived con-
ceptual limitations of well ties, such as time variance of the
seismic wavelet, are generally overridden by practical issues.
A more productive line of action is to look for improvements
in the log conditioning and calibration and, if feasible, some
reprocessing of the seismic data. White, Simm and Xu (1998)
documented an example where log conditioning and repro-
cessing combined to improve a well tie signicantly. The out-
standing well tie reported by White and Hu (1998) came
about after a 2D line was reprocessed.
Conclusions
Well ties are a key part of the interpreters art. Owing to the
potential pitfalls in making assumptions about phase and
timing, good practice in well ties demands a quantitative
approach to the determination of wavelet shape and timing.
We have shown how an appreciation of wavelet accuracy is
required as well as the goodness-of-t of the tie (as described
by parameters such as the cross-correlation coefcient).
Approximate measures of accuracy can be computed using
the equations in this paper. A quantitative approach helps
build experience on what matters in making a good well tie
in a particular area. It relies on measurements from the data,
not preconceptions of what seismic wavelets should look
like.
There are good reasons for adopting a rigorous approach
to well ties. Making assumptions about phase and timing can
be misleading. A quantitative approach to estimating wavelet
shape, timing and likely accuracy is recommended.
Approximate measures of accuracy can be computed from the
equations in this paper, starting from the cross-correlation
coefcient between the ltered synthetic seismogram and the
matched data trace. Measuring accuracy as well as goodness-
of-t is important in the quality control of a well tie.
A quantitative approach helps build experience on what
matters in making a good well tie. It relies on measurements
from the data, not preconceptions of what seismic wavelets
should look like.
Acknowledgements
We thank Mike Bacon for practical suggestions that improved
the paper.
References
Al-Chalabi, M. [1997] Parameter nonuniqueness in velocity
versus depth functions. Geophysics 62, 970-979.
Connolly, P. [1999] Elastic impedance. The Leading Edge,
18(4), 438-452.
Hosken, J.W.J. [1988] Ricker wavelets in their various guises.
First Break 6(1), 24-33.
Marion, D., Mukerji, T. and Mavko, G. [1994] Scale effects
on velocity dispersion: from ray to effective medium theories
in stratied media. Geophysics 59, 1613-1619.
Rio, R., Mukerji, T., Mavko, G. and Marion, D. [1996]
Velocity dispersion and upscaling in a laboratory-simulated
VSP. Geophysics 61, 584-593.
Simm, R., White, R. and Uden, R. [2000] The anatomy of
AVO crossplots. The Leading Edge 19(2), 150-155.
Simm, R. and White, R.E. [2002] Phase, polarity and the
interpreters wavelet. First Break 20(5), 277-281.
Stewart, R.R., Huddleston, P.D. and Kan, T.K. [1984] Seismic
versus sonic velocities: A vertical seismic proling study,
Geophysics 49, 1153-1168.
Walden, A.T. and White, R.E. [1998] Seismic wavelet estima-
tion: a frequency domain solution to a geophysical noisy
input-output problem, IEEE Transactions on Geoscience and
Remote Sensing 36, 287-297.
White R.E. [1980] Partial coherence matching of synthetic seismo-
grams with seismic traces. Geophysical Prospecting 28, 333-358
White, R.E. [1997] The accuracy of well ties: practical proce-
dures and examples, Expanded Abstract RC1.5, 67th SEG
Meeting, Dallas.
White, R.E. [1998] Stretch and squeeze just keeping up
appearances? EAGE 60th Conference and Technical
Exhibition, Leipzig, Extended Abstract P138.
White, R.E. and Hu, T. [1998] How accurate can a well tie
be? The Leading Edge 18(8),1065-1071.
White, R.E., Simm, R. and Xu, S. [1998] Well tie, uid sub-
stitution and AVO modelling - a North Sea example.
Geophysical Prospecting 46, 323-346.
rst break volume 21, October 2003
FB october v5 18-09-2003 17:21 Pagina 83