Chapter (V)

Pre-stack seismic inversion versus neural-network analysis: A

case study in the scarab field offshore Nile Delta, Egypt
In clastic depositional systems such as those encountered in the Nile Delta Basin,
simultaneous pre-stack seismic-amplitude inversion is an effective method for detecting and
appraising gas-bearing sandstone reservoirs. However, the method has limitations concerning the
requirement of a reliable set of wavelets, suitable wireline logs, and a sufficiently dense initial
model. The neural-network analysis method is an alternative technique which sometimes can
provide similar or better results and does not require significant volumes of data. Simultaneous
pre-stack inversion was applied over the Scarab field, West Delta Deep Marine concession,
offshore Egypt. The field comprises submarine channel-based gas reservoirs that extend laterally
over 20 km². Six wells were analyzed in a rock-physics study prior to performing inversion.
Three angle gathers (near: 0–15°; mid: 15–30°; far: 30–45°) were inverted for P-wave impedance
( Z p ), S-wave impedance ( Z s), P-wave velocity (V p), S-wave velocity (V s ), V p /V s , and density
(ρ) using the pre-stack inversion method. Neural-network analysis was performed using full-
stack seismic data along with well logs in the training stage, followed by cross-validation of
results and rendering of V p, V s , V p /V s , and density volumes. The V p /V s volumes produced
from the two methods were used to infer water saturation ( Sw ). Direct comparisons were made
between neural-network and pre-stack inversion results at a blind-well location to assess the
relative quality of each method. Results suggest that application of the proposed neural-network
method leads to reliable inferences. Hence, using the neural-network method alone or along with
the pre-stack inversion method has a positive impact on reserves growth and increased
production.

Introduction
Ther Quaternary clastic reservoirs in the Scarab field and in the Nile Delta Basin at large:
(1) reducing the risk of finding productive sands and (2) delineating the boundaries of these gas-
bearing sandstone reservoirs. One strategy to solve these problems is to perform pre-stack
seismic-amplitude inversion and use its outputs for the estimation of reservoir properties. Pre-
stack inversion yields volumes such as VP /VS (P-wave velocity over S-wave velocity) which
are useful for lithology discrimination and water-saturation inference. Neural networks have
been used recently as efficient tools to solve similar problems. They can circumvent general
nonlinearity among seismic-log relationships and do not require a deterministic model algorithm.
Neural networks are adaptive systems that can map input seismic attributes directly into
corresponding output log properties. We applied the two approaches to the Scarab field 3D
seismic-amplitude data sets offshore Nile Delta, Egypt.
Results show that the neural-network method provides comparable results to those obtained with
pre-stack inversion.
The study area
The Nile Delta is one of the world’s classic deltas (Figure 1). Offshore exploration using
high-quality 3D seismic-amplitude data in the last decade has resulted in the discovery of
significant reserves. Using a geology-based assessment method in 2010, the U. S. Geological
Survey estimated an average of 1.8 billion barrels of recoverable oil, 223 trillion cubic feet of
recoverable gas, and 6 billion barrels of natural-gas liquids in the Nile Delta Basin Province
(Kirschbaum et al., 2010).
The West Delta Deep Marine (WDDM) concession covers 6150 km² and lies
approximately 50 to 100 km offshore the northwest margin of the Nile cone. It includes the
Pliocene Scarab gas field (Figure 6.1).

Figure 6.1. (a) Bathymetric map of the eastern Mediterranean region; (b) index map showing the
location of Scarab field in the West Delta Deep Marine concession offshore Nile Delta, Egypt.
After Commission for the Geological Map of the World of UNESCO.
 The Scarab field reservoir is one of several north-north-west–south-southeast-trending
submarine slope channel complexes in the Late Pliocene El Wastani Formation (Figure 2). On
seismic-amplitude data, the stratigraphy is represented by large-scale channel cuts that have
acted as conduits through which Pliocene sediments were transported from the paleo-Nile River
into the deep water of the Mediterranean Basin. Analogous deepwater slope channel systems
with associated gravity-driven sediment transport and depositional processes are active on the
present-day deepwater Nile Delta slope (Islam A. Mohamed et.al 2014) .
Based on subsurface correlations and conventional petrophysical analysis, the canyon fill
comprises multiple stacked channels that are as much as 170 m in gross thickness and that have
50 m of pay with an average non-shale porosity of 28% and average water saturation of 28%. An
extensional normal fault cuts the two channels (Figure 6.2) with a maximum displacement of 120
m in the center of the field, decreasing to less than 10 m on its eastern and western limits (Islam
A. Mohamed et.al 2014) .
The Scarab field is regarded as a combination trap which consists of two faulted main
incised channels named Channel 1 and Channel 2 (Figures 6.2 and 6.3). It was the first of the
WDDM fields to enter production, in March 2003, eventually reaching a peak production of 800
million standard cubic feet of gas per day to the domestic Egyptian market (Islam A. Mohamed
et.al 2014).

Method
Seven wells (two exploration and five development wells) in the Scarab field produce
hydrocarbons from channel-fill sandstone reservoirs. All the wells are vertical and have a full
suite of wireline logs over the reservoir interval. The Scarab-Dc well (Figure 6.2) has been used
as a quality control well but was not used in inversion or in neural-network analysis. (Figure 6.3)
shows an example of well logs and reservoir stratigraphy at the Scarab-De well (Islam A.
Mohamed et.al 2014).
The seismic-amplitude data made available to this study consist of full offset stack and
partial-angle stacks of near (0–15°), mid- (15–30°), and far (30–45°) angles with a recorded
duration of 3 s and sampled at 4 ms. The available rms stacking-velocity volume was useful for
cokriging of the initial model and can be used as an external attribute in neural-network
processes. Data conditioning was applied to seismic amplitudes, including time-misalignment
correction and amplitude-offset scaling for pre-stack inversion. Another conditioning workflow
was applied to well logs for neural-network analysis. The workflow includes resampling well
logs at the seismic scale and low-pass-filtering them to limit their spectral content to that of the
seismic-amplitude data (120–130 Hz) so that aliasing is avoided when down sampling log data to
seismic resolution. A smoothing operator with low value was then applied to avoid residual
spikes (Islam A. Mohamed et.al 2014).
To define the lithology model in the Scarab field, a variety of cross plots was constructed
to identify gas sands, brine sands, and shale lithologies. The cross plots combine rock physics
parameters with the third dimension, depicted as color coding of water saturation and shale
volume. The aim is to assess the feasibility of discriminating different reservoir facies from
seismic attributes. It is found that the cross plot of VP /VS versus P-wave impedance (Figure 6.4)
is the most reliable discriminator among litho-fluid facies (Islam A. Mohamed et.al 2014).
 Figure 6.2. (a) Seismic inline through the Scarab-2 well location with gamma-ray log, (b)
schematic section, and (c) paleogeographic map illustrating the Scarab channels and faults.
Channel 1 is younger and incises into Channel 2 (After Islam A. Mohamed et.al 2014).

Figure 6.3. Stratigraphy and representative well logs of El Wastani Formation at the Scarab-De
well location. Well logs show the corresponding thick- and thin-bedded gas sands and shale
over the reservoir interval. Well logs, from left to right, are: gamma ray, deep resistivity,
calculated water saturation ( Sw ), P-wave velocity (V p), S-wave velocity (V s ), and density (After
Islam A. Mohamed et.al 2014).
 Pre-stack inversion simultaneously estimates P-impedance, S-impedance, and density. In
addition, it calculates V p, V s , and V p /V s volumes using angle-limited stacks from seismic-
amplitude data. The inversion process uses Fatti’s equation, which is a reformulation of the
original Aki-Richards equation. Coefficients and reflectivity terms in Fatti’s equation contain
information about the P- and S-wave impedances and densities as well as the angles of incidence
(Islam A. Mohamed et.al 2014).

Pre-stack inversion is an iterative algorithm which starts with an initial-guess model (<
10 Hz) and proceeds to modify it in a series of steps. Each step is guaranteed to produce a match
with field data at least as good as that of the previous step. P-impedance, S-impedance, and
density models are updated iteratively while the data misfit is minimized until reaching an
acceptable value. The final updated models are the inversion results. An initial estimate of
reservoir properties is provided by the low-frequency trend of the well logs: V p, V s , and density.
Interpolation among well logs uses the inverse distance-squared algorithm, whereas
extrapolation over the entire model uses a cokriging method. The advantage of using cokriging is
the ability to use the processing rms velocity volume to guide extrapolation, whereby the model
ties the wells perfectly while following the trend of the external velocity volume away from the
wells (Islam A. Mohamed et.al 2014).
The initial model was defined by five seismic horizons obtained from structural and
stratigraphic interpretation, including seabed surface, top Channel 1, base Channel 1, top
Channel 2, and base Channel 2.
 Figure 6.4. Cross plot of P-impedance versus V p/V s using data from the six wells in this study.
Scarab reservoir lithologies are defined using (a) water-saturation and (b) shale-volume color
bars. The red ellipse identifies potential gas-sand points, the blue ellipse identifies potential
brine-sand points, and the black ellipse identifies potential shale points (After Islam A. Mohamed
et.al 2014).
 Three wavelets were estimated from the seismic-amplitude data statistically that represent
near-, mid-, and far-angle stacks, respectively. Next, three deterministic wavelets were estimated
by computing a matching operator between the seismic-amplitude data and the reflectivity series
in the frequency domain. The validity of the extracted wavelets was verified by convolving them
with the reflectivity series, followed by cross correlation between the new synthetic data with the
seismic traces surrounding the relevant well (Islam A. Mohamed et.al 2014).
Only four wells have sufficient logging length for deterministic wavelet extraction:
Scarab-1, Scarab-2, Scarab-Dd, and Scarab-De. Extracted wavelets were averaged from those
wells. The extraction window was 550 ms (from 1950 ms to 2500 ms). Different wavelet-
extraction parameters were tested, and three taper lengths were investigated: 40, 50 and 60 ms,
with a constant wavelet length of 100 ms. Of these taper lengths, the 40-ms taper gave the best
correlation. All wavelets were extracted using the full-waveform method, and their phases were
close to 180°. Synthetic traces were then matched to the seismic data at well locations using
pattern matching with the assistance of predefined well-log markers, as shown in Figure (6.5).
Inversion analysis suggests that it is better to use deterministic wavelets than statistical ones
because deterministic wavelets provide higher correlation values with the original seismic traces
(Islam A. Mohamed et.al 2014).

Figure 6.5. An example of well ties in Scarab-De well. The water-saturation log shows reservoir
zones. Blue traces represent the calculated synthetic seismograms using the deterministic
wavelets. Red traces represent the composed seismic amplitudes. Deterministic wavelets
represent near- (blue), mid- (green), and far-angle (red) stacks (After Islam A. Mohamed et.al
2014).
 The neural-network estimation attempts to infer log properties from seismic attributes.
We used the probabilistic neural-network (PNN) method to estimate V p, V s (then calculate V p /
V s ), and density volumes, following the workflow described by Hampson et al. (2001). A linear
multi-attribute analysis was performed first to choose the set of attributes that most effectively
estimated the target log. The optimized combination of the different attributes was obtained
using a stepwise regression analysis. Subsequently, those attributes were used to train the neural-
network analysis for the estimation of target well logs. After the validation step, we applied the
neural-network method to the entire data set.
Using a validation error plot, we can determine the best number of attributes and the
operator length for each log. The minimum validation error occurs when a nine-point operator is
used with five attributes. Any other combination results in a larger validation error. The
attributes are: 1/rms velocity, integrated amplitude, amplitude envelope, second-derivative
instantaneous amplitude, and quadrature trace, respectively. Using those attributes, the
probabilistic neural network was trained and validated. The PNN shows a good result in the
prediction of VP logs. Correlation with actual logs was 0.89(Islam A. Mohamed et.al 2014).
We then applied the PNN to the complete seismic volume to produce the V p volume. The
same workflow was applied to produce V s and density volumes. Using V p and V s volumes, we
calculated the V p /V s volume.

Results
(Figure 6.6) compares P-wave velocity, S-wave velocity, and density obtained from neural-
network analysis to those obtained from pre-stack seismic inversion at the Scarab-2 well
location. (Figure 6.7) compares the stacked seismic and V p /V s results obtained from inversion
and neural-network analysis in cross-section format, whereas Figure 8 makes the comparison
with maps. Average correlation values among actual / V p, /V s , density, and /V p //V s logs and
estimated ones at the six well locations were 81%, 79%, 78%, and 72%, respectively, for the pre-
stack inversion and 89%, 85%, 62%, and 84%, respectively, for the neural-network method.
Neural-network results are comparable to those obtained with pre-stack seismic-amplitude
inversion (Islam A. Mohamed et.al 2014).
Figure 6.6. Comparison of pre-stack inversion results (three right columns) and neural-network
results (three left columns) at the Scarab-2 well location: (a) P-wave velocity section, (b) S-wave
velocity section, and (c) density section. Note that the two methods yield results that are
comparable to the actual logs, but inversion results seem to resolve the upper sand better and
preserve fine vertical details. (Figure 6.2) shows the location of this line (After Islam A.
Mohamed et.al 2014).
Figure 6.7. (a) Seismic inline through the Scarab-2 well location with gamma-ray log across the
reservoir interval: comparison of V p / V s results from (b) pre-stack inversion against (c) neural-
network results overlaid by seismic traces. Figure 2 shows the location of this line (After Islam A.
Mohamed et.al 2014).

Blind-well testing
The Scarab-Dc well (the blind well) was not included in the inversion and neural-network
analysis, and it was kept aside for quality control of estimation products. In this case, we verified
the success of the two techniques used in the study with a good degree of confidence. Figure 9
compares inversion results with neural-network results at the Scarab-Dc well location. Figure 10
shows the same comparison beside the average correlation at the six wells (Islam A. Mohamed
et.al 2014).

Estimation of water saturation

As a final benchmark for the comparative study, we tested the reliability of the estimation
methods to quantify water saturation. The idea was to establish a relationship between / V p //V s
and water saturation. This estimation is relatively difficult for pre-stack inversion, for which it is
relatively simple to detect gas-saturated units. The differentiation among gas sand, brine sand,
and shale was performed with cross plots of / V p //V s versus P-impedance. Assuming water
saturation and volume of shale cutoffs equal to Sw < 50% and VSh < 40%, respectively, we
defined a /V p //V s cutoff value (≈ 2.6) that separates gas-bearing sands from other lithologies
(Figure 6.4). Subsequently, we implemented that cutoff on the / V p //V s maps obtained from the
inversion and neural-network procedures to detect gas-bearing sands. Figure 6.11 shows the
successful definition of the upper Pliocene gas sands on /V p //V s maps (Islam A. Mohamed et.al
2014).
 Figure 6.8. /V p//V s inferred over the Channel 1 reservoir: (a) pre-stack inversion result
and (b) neural-network result. The neural-network procedure yields /V p//V s maps similar to
those obtained with inversion (After Islam A. Mohamed et.al 2014).
 Figure 6.9. Inferred logs (in red) compared with actual logs (in black) at the Scarab-Dc well
location: (a) pre-stack inversion result; (b) neural-network result (After Islam A. Mohamed et.al
2014).
Figure 6.10. Correlation differences between the two methods for the inference of each target
log. Dark colors identify the average correlation at the six wells, and light colors identify
correlation at the blind-well location (After Islam A. Mohamed et.al 2014).
 To establish a relationship between /V p //V s and water saturation, we implemented a
linear equation obtained from the regression line between water-saturation / V p //V s values. The
average correlation between estimated and actual water saturation at the six well locations was
68% for pre-stack inversion and 78% for the neural-network method. Concerning the blind-well
location, the correlation is 43% for inversion and 66% for the neural-network method. The low
correlation yielded by inversion refers to / V p //V s itself, which was affected by the failure in the
VS-log estimation. That failure probably resulted from the quality of the initial model and the
seismic amplitude data (Islam A. Mohamed et.al 2014).
The neural-network method can be trained with calculated water-saturation logs at well
locations against the seismic response and / V p //V s (Figure 6.12). Hence, we can take advantage
of inferring nonlinear relationships among given well logs, seismic-amplitude data, and / V p //V s
volume. The average correlation at the six well locations is 91% and at the blind well location is
78%. The better prediction with the neural network method has been interpreted as resulting
from the ability of neural networks to capture the nonlinear relation- ship between / V p //V s and
water saturation (Islam A. Mohamed et.al 2014).

Figure 6.11. 3D view of the Channel 1 gas sand (red), which exhibits /V p//V s values
lower than 2.6 (dimensionless): (a) pre-stack inversion result; (b) neural-network result (After
Islam A. Mohamed et.al 2014).
 Figure 6.12. Water saturation inferred over the Channel 1 reservoir. The inset panel shows the
inferred water-saturation (Sw) log (in red) compared with the actual log (in black) at the Scarab
Dc well location (After Islam A. Mohamed et.al 2014).

Conclusions
Neural-network analysis provided comparable or even better estimation results than those
obtained with pre-stack seismic-amplitude inversion for the data considered in this article.
Results indicate that using the neural-network procedure instead of pre-stack inversion mitigates
some of the limitations of the inversion process. It requires neither horizon picking nor wavelet
extraction; conversely, one can predict any log property, not only impedance logs. The neural-
network approach was used successfully to predict water-saturation values and captured the
nonlinear relationship between water saturation and /V p //V s (Islam A. Mohamed et.al 2014).
However, pre-stack inversion still has the advantage of preserving lateral continuity and
resolving fine vertical details, which makes inversion more reliable geologically in the 3D
context. In addition, the neural-network method has its own limitations regarding quality of
seismic-amplitude data and well logs, length of the logged target interval, seismic attributes, and
neural-network parameters selected for the estimation. The probabilistic neural-network method
is expected to work favorably in the same time window used for training and whenever the
reservoir maintains the same seismic signature. However, away from well control or the training
window, PNN results cannot be trusted (Islam A. Mohamed et.al 2014).
Further investigation about the validity of the neural-network method is needed. The most
limiting factor of the PNN method is the need for at least three wells for the training stage. Data
are readily available in the development and production phases of a reservoir. The more well
data available, the more reliable and accurate the neural-network results. Thus, using the neural-
network procedure alone or along with pre-stack inversion results improved the definition of late
Pliocene gas sands in the Scarab field, which should have an ultimate impact on increasing
production and economic efficiency (Islam A. Mohamed et.al 2014).