on a combination
of rock physics modeling, seismic attribute generation and
pattern recognition via neural network analysis. The result
was a new lithologically calibrated attribute that showed
the producing wells to be inside the indicated oil sand area
and the non-producing wells to be outside this area.
Introduction
In many cases poststack seismic data is the only available
source of information on interwell stratigraphy and
lithology. In such a situation the amount of information
that can be extracted on reservoir properties such as
porosity or hydrocarbon content is usually quite limited.
This project had core, well log, and fully processed 3-D
poststack seismic data available. To extract the most
lithologic knowledge from this data, a method was
developed that combined rock physics, seismic modeling,
neural networks and seismic attributes in a unique way.
The geologic setting for this study was a North Sea tertiary
turbidite system. The seismic survey covering the area of
interest was about 325 square km. There were 5 wells
inside the survey area, two of which had encountered oil
saturated pay sands. These wells had a full suite of high
quality logs, and the two producing wells also had dipole
shear wave data.
Reservoir Classification Using Borehole Data
Basic log analysis
Five wells were available for this study. Wells 2, 3, and 4
were non-producers in the Heimdal Formation or had no
Heimdal sand present. Wells 5 and 6 had significant oil
content in the Heimdal. These two wells also had
measured dipole shear wave logs that were used to calibrate
the Vs prediction in the other three wells. Other log curves
available were gamma, neutron, density, sonic, ILD, caliper
and in some wells, MSFL.
Total porosity (Phi), shale volume (Vshale), and water
saturation (Sw) were computed first. The zones of interest
were primarily sand-shale sequences between 2100 and
2300 meters. The shale volume was determined mainly
from the gamma log, but in some cases we used gamma,
neutron, and density logs. The total porosity was based
mainly on the density log, which appeared to be of high
quality throughout the entire zone of interest. In
hydrocarbon-bearing intervals an average of neutron and
density logs was used to find total porosity. Caliper logs
showed good wellbore conditions in the zones of interest in
all wells.
Core derived lithologic boundaries and well log derived
volumetric curves were used to train a neural network to
automatically identify five important lithologies from logs.
Shear wave velocity prediction
Two main methods of Vs prediction were tested in the non-
hydrocarbon bearing wells. These methods are
implemented within the PetroTools software package. The
first was Greenberg-Castagna. In the producing wells this
relation showed good agreement with the dipole shear log,
except in the oil sands. This suggested that there might be
a problem with mud invasion or wellbore washouts. But
since the caliper log showed no indication of washout, the
former was deemed more likely. The good agreement
between Greenberg-Castagna and the dipole shear log in
non-pay zones was considered to be at least partial
confirmation of our Vshale and porosity values.
The second Vs prediction method was based on the
Castagna mudrock equation calibrated to the two wells
with known shear wave velocities. Only the water
saturated portions of the wells were used to establish the
calibration. The relationship derived was;
Vs = 0.73 Vp 767 (m/sec).
This simple relationship gave a slightly better correlation
than Greenberg-Castagna so it was used to compute Vs in
the water-saturated wells 2, 3, and 4.
Mud filtrate invasion correction
Deep and shallow resistivity logs indicated that water-
based mud filtrate had invaded the near wellbore region in
Well 5. Because density and sonic are both shallow
investigation logs, we performed a calculation to obtain
values that were more representative of the true reservoir.
The resistivity from the MSFL curve (Rmsfl) showed that
the near wellbore Sw was essentially 100%. It also
indicated an apparent water resistivity (Rwa) that was very
similar to that indicated by the deep induction log (Rt).
This means that the reservoir brine and the mud filtrate
have essentially the same salinity. Therefore to correct the
sonic and density, we assumed that they responded to 100%
Sw when logged and we performed a fluid substitution to
the Sw indicated by the Rt log. This resulted in lower
density and Vp in the pay zone than recorded by the
logging tool. There was no shallow resistivity log available
for Well 6 but comparisons of the dipole sonic to the
predicted Vs from the modified mudrock equation
suggested that this well also had significant mud filtrate
invasion. Therefore, it was corrected for mud invasion in
the same manner as Well 5. Reservoir characterization using rock physics, seismic attributes, and neural networks
Fluid substitution
Biot-Gassmanns relations were used to create additional
pseudo-wells in which the oil saturation was removed
from the actual oil sands and oil was added to one of the
wells where the Heimdal was water saturated. The
following properties of the solid and fluid components were
used in the fluid substitution. The fluid moduli and density
varied with pressure and temperature as described by
Batzle and Wang:
Solid
Moduli and Density
Quartz
K=36.6 Gpa, u=45 Gpa, rho=2.65 g/cm3
Shale
K=20.8 Gpa, u=6.9 Gpa, rho=2.58 g/cm3
Fluid
Description
Oil
32 API, GOR = 65 L/L
Brine
60,000 PPM NaCl (~0.05 Ohm-m @ 80C)
Training of neural network with well log data
The purpose of this step was to:
Define lithology classes to be identified at the wells
and ultimately across the 3-D volume.
Build a classification scheme using neural network
training to predict the lithology classes.
Lithology classes were defined as sedimentary units with
distinguishable characteristics such as clay content,
bedding configuration (massive or interbedded), grain size,
cementation, and rock mineral properties. Also pore fluids
of the specific identified reservoir units were included in
the classifications. These parameters were obtained from
the 75 meters of classified core from well 5. The following
reservoir classifications were used:
1. Pure shale; 2. Silty shale; 3. Interbedded sandstone-
shale; 4. Massive wet sand; 5. Unconsolidated wet sand; 6.
Planar laminated oil sand; 7. Unconsolidated oil sand;
8.Undefined.
In addition to the lithologic column output, multiple well
log curves were provided as inputs. These curves were
density, total porosity, Vp, Vs, clay volume, and water
saturation.
The neural network used this information and developed
weights and scalars by iterative correction to minimize the
discrepancies between the predicted lithology results and
the actual classes. The training was done on a limited range
of depths encompassing the reservoir interval. We selected
an interval from 2100m to 2300m for the reservoir
classification. An example of this classification for Well 5
is shown in Figure 1. Once the training operations were
completed and validated, the neural network weightings
and scalars were applied to the processed logs to obtain a
reservoir classification at each well. Where core data was
available in Well 5 the match between actual and predicted
lithology was 92%.
Reservoir Classification Using Seismic Data
Synthetic seismic generation
To tie the well log-derived attributes and the seismic-
derived attributes, synthetic seismograms were generated
and used to train the seismic properties to predict the
lithology classes.
The following tasks were performed:
1)

Wavelets were extracted at each well location based
on a 9X9 grid of seismic traces around the well.
2)

Logs input to the synthetic seismogram program were
Vp, Vs, and density, after correction for mud filtrate
invasion. The predicted classification curves at each
well were also input for later conversion to seismic
travel time.
3)

Synthetic seismograms were calculated based upon the
extracted wavelet and wireline logs. The synthetics
were calculated by building a synthetic offset model of
10 offsets spanning a range of 0-5000m. The offsets
then had normal moveout applied and were summed to
generate a stacked synthetic trace.
4)

The reservoir classification curves were then
resampled to time using the synthetic seismogram -
derived time-depth curve.
At this point the calculated synthetic seismograms and
time-converted lithology classes were available for attribute
generation and neural network training.
Generation of seismic attributes
We define all seismic-derived parameters as seismic
attributes. They can be velocity, amplitude or rate of
change of any of these with respect to time or space.
Furthermore, we are able to classify these attributes based
upon their computational characteristics, e.g., some of the
attributes that are computed from the complex trace such as
envelope, frequency and phase, etc., correspond to various
measurements of the propagating wavefield. For these
attributes we adopt the term 'Physical Attributes'. Other
attributes are computed from the reflection configuration
and continuity properties of the sub-surface and we group
these together as 'Geometric Attributes'.
The principle objective of computing seismic attributes is
to provide accurate and detailed information to the
interpreter on structural, stratigraphic and lithological
parameters of the seismic prospect (Taner et al.).
For this pr