A New Approach To Improve Reservoir Modeling Via Machine Learning

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

A new approach to improve reservoir modeling

via machine learning


Islam A. Mohamed1, Adel Othman2, and Mohamed Fathy2
Downloaded 08/01/20 to 125.70.254.71. Redistribution subject to SEG license or copyright; see Terms of Use at https://library.seg.org/page/policies/terms

https://doi.org/10.1190/tle39030170.1

Abstract the reservoir. The new approach represents a shortcut for the
In highly heterogeneous basins with complex subsurface seismic-to-simulation process, providing a reliable and fast way
geology, such as the Nile Delta Basin, accurate prediction of of constructing a reservoir model and making the seismic-to-
reservoir modeling has been a challenge. Reservoir modeling is simulation process easier.
a continuous process that begins with field discovery and ends
with the last phases of production and abandonment. Currently, Introduction
the stochastic reservoir modeling method is widely used instead Reservoir simulation is simply the prediction of fluid flow
of the traditional deterministic modeling method to consider behavior in a particular reservoir over time. This process is essential
spatial statistics and uncertainties. However, the modeling work- for estimating reserves, predicting future production, and making
flow is demanding and slow, typically requiring months from the decisions regarding field development. This requires having accu-
initial model concept to flow simulation. In addition, errors from rate static and dynamic 3D models of the reservoirs that reflect
early model stages become cumulative and are difficult to change the original well logs, seismic data, and production history.
retroactively. To overcome these limitations, a new workflow is In the static case, the model represents the physical space of
proposed that implements probabilistic neural network inversion the reservoir by an array of discrete cells, delineated by a 3D grid.
to predict reservoir properties. First, well-log data were conditioned Each cell has specific values of the reservoir properties, i.e.,
properly to match the seismic data scale. Then, the networks were porosity, permeability, and water saturation (Sw), outlining the
trained and validated, using the conditioned well-log data and 3D distribution of the reservoir (Tyson, 2007). The conventional
DOI:10.1190/tle39030170.1

seismic internal/external attributes, to predict water saturation reservoir modeling workflow, however, is a complicated and
and effective porosity 3D volumes. The resulting volumes were demanding process. After gathering and conditioning of the input
sampled in simulation 3D grids and tested using a blind well test. data set, we must build a structural framework and design a 3D
Subsequently, the permeability was calculated from a porosity- grid. Then, we have to distribute the facies and petrophysical
permeability relationship inside the reservoir. Finally, a dynamic properties in a geologically reasonable manner, run different
simulation project of the field was performed in which the historical scenarios, and perform uncertainty analysis. After that, we design
field production and pressures were compared to the predicted a simulation grid to simulate the flow of fluids within the reservoir
values. One of the Pliocene deepwater turbidite reservoirs in the over its production lifetime. The final step is the model optimiza-
offshore Nile Delta was used to demonstrate the proposed tion and production history matching (Figure 1).
approach. The results proved the accuracy of the model in predict- Although this is a widely used and standard workflow for
ing the reservoir properties and honoring the heterogeneity of most oil companies, it suffers from three main problems. First

Figure 1. Steps of the conventional seismic-to-simulation workflow.

1
Rashid Petroleum Company (Rashpetco), Cairo, Egypt. E-mail: islam.ali@rashpetco.com.
2
Al-Azhar University, Cairo, Egypt. E-mail: a3othman2@yahoo.com; green_geophysics@yahoo.com.

170 The Leading Edge March 2020 Special Section: Reservoir characterization Part II
is the demanding nature of the workflow. Building a robust present-day shoreline (Figure 3). The field was discovered in 2000
reservoir model requires many inputs. As an example, we need based on seismic amplitude anomalies (bright/flat spots) and
accurately interpreted horizons and faults. Also, we need many encountered gas-bearing sands in a slope canyon setting.
Downloaded 08/01/20 to 125.70.254.71. Redistribution subject to SEG license or copyright; see Terms of Use at https://library.seg.org/page/policies/terms

equally distributed wells to represent the reservoir heterogeneity Subsequently, the first exploration well was followed by another
via the geostatistics. A low number of inputs likely will produce three appraisal wells (2000–2002). The field was sanctioned in
a poor model. The second problem is the cumulative error prob- 2008, and six development wells were drilled as a result, three in
lem. This means that early-stage errors become cumulative and each sector (i.e., north and south) and tied back to existing facilities
are difficult to change retroactively. It is even more difficult to (2008–2009). Production started in August 2009, and the cumula-
define where exactly the error is coming from when the final tive gas production reached approximately 662 billion cubic feet.
model is not consistent with the production data. The third The Sequoia Field is a submarine slope canyon system that
problem is slowness. This workflow requires months from the consists of multistacked channel systems associated with gravity-
initial conceptual model to the flow simulation, and in addition driven sediment transport (turbiditic) deposits (Cross et al., 2009).
there is the extra time spent fixing the model to better tie the This channel complex runs from south to north and starts narrow
production profile. (5 km wide) but broadens to more than 20 km wide in the north
To overcome these problems, we have designed a new workflow with many branches (Figure 4). The field gets its name due to a
that implements the probabilistic neural network (PNN) to predict similarity in shape to a type of giant redwood tree that grows in
reservoir properties. These properties are mainly S w and effective California. The total length is more than 30 km.
porosity (ϕ). The resulting volumes were sampled in simulation The reservoir inside the canyon consists of a succession of
3D grids and tested using a blind well test. Subsequently, the sandstones and mudstones formed by many stacked channels,
permeability is calculated from the porosity-permeability relation- organized into a composite upward-fining profile. Those channels
ship inside the reservoir. After the initial history match process, are different in shape and behavior and include laterally amalgam-
PNN parameters were adjusted to improve the match. The final ated channels, sinuous channels, channels with frontal splays,
model represents the best match to original field measurements and leveed channels. Following the work of Cross et al. (2009)
and production data (Figure 2). and the application of seismic variance and spectral decomposition
DOI:10.1190/tle39030170.1

Compared with the conventional workflow, this workflow has on 3D seismic data presented by Mohamed et al. (2017), we are
four main advantages: (1) it does not require many inputs; (2) it able to map and understand the geometry of the incised channels.
is short and direct; (3) it mitigates the cumulative error problem
and decreases the statistical uncertainties; and (4) it is fast, reducing
reservoir modeling time by 80% (taking two weeks instead of 10).

Geologic setting
The Sequoia Field is one of the Pliocene gas fields in the
offshore Nile Delta, Egypt. It is located on the western side of
the West Delta Deep Marine (WDDM) and Rosetta concessions
to the north and south, respectively, and about 50 km from the

Figure 3. (a) Bathymetric map of the eastern Mediterranean region. (b) Index
Figure 2. The proposed new approach that implements the PNN to predict the map showing location of the Sequoia Field in the WDDM concession offshore Nile
reservoir properties. The new workflow represents a reliable shortcut that mitigates Delta, Egypt. Upper Pliocene gas fields are in gray, and Sequoia Field is in orange.
the cumulative error problem and reduces the modeling time by 80%–90%. Modified from Mohamed et al. (2019).

Special Section: Reservoir characterization Part II March 2020 The Leading Edge 171
The Sequoia story started with the initial erosion of the channel the trained network to the 3D seismic and inversion volumes to
(stage I), followed by the deposition of several laterally amalgam- predict the target reservoir property.
ated channels (stage II), then another deposition of sinuous chan- Practically, we have used the following data as an input for this
Downloaded 08/01/20 to 125.70.254.71. Redistribution subject to SEG license or copyright; see Terms of Use at https://library.seg.org/page/policies/terms

nels with aggradational stacking patterns (stage III). The next study: a seismic full-stack prestack depth migration volume acquired
cycle was characterized by channels with frontal splays and leveed in 2006 and reprocessed in 2014 with a record duration of 6 s and
overbank (stage IV). Finally, the channel system is abandoned sampled at 4 ms. We also have used the available in-house prestack
during mudstone deposition (stage V) and buried by prograding seismic inverted volumes — P-impedance, S-impedance, P-wave
sand sheets (stage VI). All of these channels are stacked to form velocity, S-wave velocity, VP /VS ratio, lambda-rho, and mu-rho
a thick reservoir that is approximately 200 m in gross thickness volumes. The interpretation data include structural maps (channel
and 77 m of pay gas sand. The average nonshale porosity is 24%, top, gas base, and channel base), fault sticks, and fault polygons.
and the average Sw is 34%. Eight wells were used in the study. These wells are sorted
into exploration (Sapphire-2 and Rosetta-10) and development
Methodology wells (Sequoia-D1, -D2, -D3, -D4, -D5, and -D6). One of these
Many machine learning algorithms have proved their validity
and show significant accuracy in predicting various reservoir
properties. The PNN is one of these algorithms. The PNN can
be used either for classification or for reservoir properties predic-
tion. In the prediction mode, the algorithm trains a neural network,
utilizing the information from multiple wells and seismic data,
to predict the reservoir properties away from the well control. The
procedure involves three main steps — finding the best set of
attributes via multilinear regression technique, network training,
and validation.
In the first step, we use the technique of multilinear regression
DOI:10.1190/tle39030170.1

to find the set of attributes that most effectively estimates the target
reservoir property, i.e., the set of attributes that will predict a given
reservoir parameter with the lowest prediction error. For this step,
we use the prestack inversion volumes as external attributes besides
the seismic-derived (internal) attributes. The seismic-derived attri-
butes are grouped into instantaneous, windowed frequency, filter
slice, derivative, integrated, and time attributes. Moreover, all of
the attributes are increased by applying nonlinear transforms — Figure 4. Root-mean-square amplitude map over Sequoia Field overlaying
natural log, exponential, square, inverse, and square root (Mohamed variance slices extracted at the base of channels. Seven wells were used in
et al., 2017). We try all of these attributes and then determine this study, and one was kept as a quality control blind well.
statistically the best order and num-
ber of attributes that give the lowest
prediction error and save them for
the subsequent network training.
In the second step, the PNN
algorithm trains the network. It uses
Gaussian weighting functions to fit
the seismic attributes to the training
reservoir property samples. Each
mean of these functions is centered
at each training sample, and the
algorithm optimizes the standard
deviation or the smoothness param-
eter, sigma, for each attribute. Those
sigma values are determined auto-
matically by the cross-validation
method. In the validation step,
performed simultaneously with the
training, we remove part of the well
data (one-third of the wells) from Figure 5. Cross sections through the blind well location (Sequoia-D5) show results of the PNN method. S with actual
w
the training stage and try to predict Sw log is shown at top, and ϕ with the actual ϕ log across the reservoir interval is shown at bottom. Figure 4 shows the
these wells blindly. Finally, we apply location of this line.

172 The Leading Edge March 2020 Special Section: Reservoir characterization Part II
Downloaded 08/01/20 to 125.70.254.71. Redistribution subject to SEG license or copyright; see Terms of Use at https://library.seg.org/page/policies/terms

Figure 6. Cross sections show the resampling of (a) ϕ and (b) Sw in the simulation 3D grid for the dynamic simulation process. Some details were lost during this
process, and to make up for that, we had to use multipliers at the final stage.

VP /VS ratio, inverted VS , square root of inverted VP, frequency filter


DOI:10.1190/tle39030170.1

5/10–15/20, amplitude weighted phase, and cosine instantaneous


phase, respectively. The input attributes for ϕ prediction are
1/inverted VP /VS ratio, the square of lambda-rho, the logarithm of
mu-rho, two-way time, and frequency filter 5/10–15/20.
After training and validation of the networks, we applied the
trained neural networks to the 3D seismic data volume to generate
3D volumes. The PNN shows outstanding results in the prediction
of S w and ϕ, the average correlations with actual logs were 0.90
and 0.94, respectively. Figure 5 shows the resulting Sw and ϕ
sections through the blind well location. The normalized correla-
tions with the actual logs for the blind well were 0.86 for the S w
and 0.81 for the ϕ.
It is easy to note the high level of detail in the resulting
volumes and the lateral continuity inside the reservoir. This degree
of resolution is required in estimating flow continuity and barriers
in the dynamic simulation model. The simulation 3D grid was
designed in conjunction with the reservoir engineer. The grid
Figure 7. (a) The porosity-permeability relationship. Permeability values from the
linear trend are overestimated at the low porosity values. (b) Permeability values was designed to cover the main channel complex and also include
were calculated from the porosity using a third-order polynomial equation. the faults. For practicality, it should be a coarse grid that is
100 × 100 m instead of the original seismic cell dimensions
wells, Sequoia-D5, was kept as a blind quality control well (25 × 25 m). The average cell thickness was 1.5 m, and the layers
(Figure 4). The S w and ϕ logs for these wells were provided by are proportional inside the reservoir. Both Sw and ϕ were upscaled
an in-house petrophysicist. using the arithmetic averaging method in the simulation grid.
We have applied the same PNN procedure that was proposed Figure 6 shows the results before and after the upscaling
by Mohamed et al. (2017), using Hampson-Russell software, to process. Some details were lost during this process. To make up
predict Sw and ϕ in two different separate runs using different sets for that, we had to use pore volume and permeability multipliers
of attributes. First, we filtered, resampled, and smoothed the logs at the final stage, but before that, we created the permeability
to match the seismic data vertical scale (4 ms). Second, we trained volume using the porosity-permeability relationship that is
the network using the conditioned well-log data and seismic internal/ coming from the special core analysis and the well test data. As
external attributes. Simultaneously, we validated the network by shown in Figure 7a, the relationship between porosity and
hiding some of the wells and predicted them using the trained permeability could be expressed either by a linear or polynomial
network. The input attributes for Sw prediction were: 1/inverted trend. It has been observed that the permeability values from

Special Section: Reservoir characterization Part II March 2020 The Leading Edge 173
Downloaded 08/01/20 to 125.70.254.71. Redistribution subject to SEG license or copyright; see Terms of Use at https://library.seg.org/page/policies/terms
DOI:10.1190/tle39030170.1

Figure 8. (a) Saturation at one of the training wells location and the corresponding production history versus the model’s prediction for (b) the conventional model versus
(c) the PNN model. The observed gas production is represented by red circles, while the predicted gas production from the model is the red line. Bottom hole pressure
is represented by black circles, while the pressure prediction is the black line. The actual water production (blue circles) has no equivalent prediction from the model
because this run was dedicated to gas production only.

the linear trend are overestimated at the low porosity values. Both simulation models show superb history matching and
Hence, we chose to apply the polynomial equation. The result tie the production data very well. However, the PNN model is
is shown in Figure 7b. much more efficient and requires much less effort and time. The
high predictability of the PNN model can be attributed to its
Results ability to honor the heterogeneity of the reservoir and the accurate
To measure the quality of the model, we have kept one of prediction of the reservoir properties. Keeping in mind that the
the wells (Sequoia-D5) aside from the study to be used as a blind accuracy of the PNN prediction is dependent on many factors
test. In the static case, the predicted reservoir properties (satura- such as the well-log and seismic-amplitude data quality, any
tion and porosity) were very close to the actual logs as shown imaging artifacts or poor illumination zones probably will cause
in Figure 5. However, the model should reflect the original well incorrect reservoir property predictions.
logs, seismic data, and production history to be considered as a The final PNN model represents the best deterministic
“good model.” A dynamic simulation project of the field was match to original field measurements and production data.
performed in which we compared the historical field production Consequently, this model is a good base-case model. For the
and pressures to the predicted values in a history matching low/high-case model, we may use the error percentiles from
process. The dynamic simulation results from the conventional the neural-network prediction processes. Otherwise, using the
model and the proposed PNN model at one of the training wells geostatistical inversion probabilities as inputs for the PNN
and at the blind well locations are shown in Figures 8 and 9, instead of the deterministic inversion could give different sce-
respectively. The observed gas production (red circles) from the narios to address the uncertainties.
well matches perfectly the predicted gas production from both
models (red lines). Also, the pressure forecasting (black lines) Conclusions
generally matches the bottom hole pressure trend (black circles) In real life, where the reservoir can be developed and produced
in both cases. However, the pressure predictions from the PNN only once and mistakes can be tragic and irrefragable, we need
model look more accurate. The actual water production (blue an accurate reservoir that mimics the well logs, seismic data, and
circles) has no equivalent prediction from the models because production history. The conventional way of modeling is demand-
these runs were dedicated to the gas production only. ing and slow, typically requiring months from the initial model

174 The Leading Edge March 2020 Special Section: Reservoir characterization Part II
Downloaded 08/01/20 to 125.70.254.71. Redistribution subject to SEG license or copyright; see Terms of Use at https://library.seg.org/page/policies/terms
DOI:10.1190/tle39030170.1

Figure 9. (a) Saturation at the blind well location and the corresponding production history versus the model’s prediction for (b) the conventional model versus
(c) the PNN model. The observed gas production is represented by red circles, while the predicted gas production from the model is the red line. Bottom hole pressure
is represented by black circles, while the pressure prediction is the black line. The actual water production (blue circles) has no equivalent prediction from the model
because this run was dedicated to gas production only.

concept to flow simulation. In addition, errors from the early Data and materials availability
model stages become cumulative and are difficult to change Data associated with this research are confidential and cannot
retroactively. In this study, we propose a new approach that be released.
implements the PNN to predict reservoir properties. The results
proved the accuracy of the model in predicting the reservoir Corresponding author: islam.ali@rashpetco.com
properties and honoring the heterogeneity of the reservoir. The
proposed workflow represents a shortcut for the seismic-to- References
simulation process, providing a reliable and fast way of constructing Cross, N. E., A. Cunningham, R. J. Cook, A. Taha, E. Esmaie, and N.
a reservoir model, which will in turn effectively enhance the El Swidan, 2009, Three-dimensional seismic geomorphology of a
deep-water slope-channel system: The Sequoia Field, offshore west
reserves estimation and production planning.
Nile Delta, Egypt: AAPG Bulletin, 93, no. 8, 1063–1086, https://
doi.org/10.1306/05040908101.
Acknowledgments Mohamed, I., M. Hemdan, A. Hosny, and M. Rashidy, 2019, High-
We would like to thank the Offshore Mediterranean Conference resolution water-saturation prediction using geostatistical inversion
2019 organizers and the Society of Petroleum Engineers Italian and neural network methods: Interpretation, 7, no. 2, T455–T465,
Section for awarding this work first prize in the student paper contest, https://doi.org/10.1190/INT-2018-0153.1.
PhD category. We thank Egyptian General Petroleum Corporation Mohamed, I., O. Shenkar, and H. Mahmoud, 2017, Understanding
reservoir heterogeneity through water saturation prediction via
and Rashid Petroleum Company (Rashpetco) for permission to
neural network — A case study from offshore Nile Delta: The
publish this study and for providing the data. We also thank Rashpetco Leading Edge, 36 , no. 4, 298–303, https://doi.org/10.1190/
staff, especially Basem Khalaf and Ayman Shamsia, for their valuable tle36040298.1.
contribution to this study. Finally, we thank the anonymous reviewers Tyson, S., 2007, An introduction to reservoir modeling: Pipers’ Ash Ltd.
who helped improve the quality of this work.

Special Section: Reservoir characterization Part II March 2020 The Leading Edge 175

You might also like