SPE-198656-MS

Downloaded from http://onepetro.org/SPEGOTS/proceedings-pdf/19GOTS/2-19GOTS/D021S007R001/1181910/spe-198656-ms.pdf by Universidade Federal De Pelotas user on 25 September 2024

Well Placement Optimization with an Artificial Intelligence Method Applied to
Brugge Field

Jérémie Bruyelle and Dominique Guérillot, Texas A&M University at Qatar

Copyright 2019, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Gas & Oil Technology Showcase and Conference held in Dubai, UAE, 21 - 23 October 2019.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
Planning the position of a new well is an essential task to increase hydrocarbon production. However, trying
to capture the uncertainty related to the geological properties, an ensemble of geological realizations of the
models must be considered. The objective can be to identify the locations of the new wells that maximize the
average oil recovery or Net Present Value (NPV), for an ensemble of geological models. Also, considering
an ensemble of geological models, an objective can be to minimize the standard deviation of the NPV (or
cumulative oil recovery). Reducing the uncertainty on the NPV leads to reduce the risk for the placement
of the new wells. These two objectives can be defined with a "mean-standard deviation" formulation of
the objective function. Following the weight given to each objective, the placement of the wells is not the
same. To know all the well-placement possibilities and associated risk, many optimization problems must
be done. Such as procedures, require the use of optimization method and reservoirs simulations on each
reservoir model. This optimization problem leads to a high number of reservoir simulation to determine
the appropriate location and can be expensive in computation time. To reduce the computation time, a
method consists in to substitute the reservoir simulator by a meta-model. In this paper, the use of an artificial
neural network as a proxy model is considered. The objective is to propose a method to speed up the well-
placement optimization process using a proxy model based on an artificial intelligence technique to replace
the reservoir simulator. The proposed methodology considers an ensemble of realization and aims to provide
an overview of solutions in an economic point of view inspired by portfolio optimization problems. An
application of the methodology is presented for the Brugge field.

Introduction
The localization of new wells is a challenging task in the oil field development plan. Many authors are
proposing to maximize the oil recovery or the Net Present Value (NPV) by controlling the location of the
new wells. Some of them consider a deterministic reservoir model while others consider an ensemble of
reservoir models to take into account the geological uncertainty.
This problem has been addressed considering the reservoir model as deterministic. Gradient-based
methods have been applied to maximize NPV (Sarma and Chen, 2008; Zandvliet et al., 2008). From a
mathematical point of view, well-placement optimization is a challenging task due to the complexity of
2 SPE-198656-MS

the objective function usually composed of several local minima. To avoid the convergence to these local
minima, several authors proposed to use derivative-free algorithms. Emerick et al. (2009) had developed
a computer-aided optimization tool based on a genetic algorithm to optimize well-placement considering
arbitrary well trajectories. Bouzarkouna et al. (2013) proposed to couple the Covariance Matrix Adaptation
Evolution Strategy (CMA-ES) (Hansen and Ostermeier, 2001) optimizer with surface response models

Downloaded from http://onepetro.org/SPEGOTS/proceedings-pdf/19GOTS/2-19GOTS/D021S007R001/1181910/spe-198656-ms.pdf by Universidade Federal De Pelotas user on 25 September 2024

of the NPV for each well. This approach leads to improve the convergence of the optimization method
knows to be costly in term of direct evaluation of the objective function, i.e., the number of reservoir
simulations. Isebor et al. (2014) proposed a method to optimize simultaneously the number of wells, their
locations, and the control parameters using a derivative-free algorithm. Awotunde and Sibaweihi (2014)
used the CMA-ES and Differential Evolution (DE) algorithms and considered a multiobjective approach
composed of the NPV and the Voidage-Replacement Ration (VRR). Forouzanfar et al. (2016) used the
CMA-ES to optimize the well-placement and the control parameters sequentially or simultaneously for a
deterministic geological model. Alrashdi and Sayyafzadeh (2019) compared (μ + λ) Evolution Strategy,
Genetic Algorithm (GA), Particle Swarm Optimisation (PSO) and CMA-ES for well control, placement,
and joint optimization problems.
To compensate for insufficient information on the structure of the reservoirs at distances less than the
spacing of the soundings, an approach proposed by geologists consisted in reasoning by analogy (Montadert
(1963), Groult et al. (1966), Tomutsa and al. (1986), Ravenne et al. (1989, 1991)): to look out for geological
structures comparable to those of the oil formations that interest us. The fundamental hypothesis that is
made in this step can be expressed as follows: The way the sediments have been deposited can be explained,
and these depositing mechanisms have been repeated for different deposits: there are families of reservoirs
that have been built "roughly" in the same way. Thus, if we can reconstruct a well-known deposit, if we can
characterize it with only a few parameters, then we can reconstruct other deposits of the same type of deposit,
using available information (electrical logs, well tests, seismic surveys, etc.). The probabilistic character of
these "images" of reservoirs is the ransom of the uncertainty that exists in an immense region where little
is known about the information available in the wells. For heterogeneous deposits, one possible approach
is to assume different geometries for a deposit and to make each one of them reservoir studies. Of course,
the number of these assumptions can only be limited in number, given the time it would take to interpret
the calculations, which themselves are very heavy. But the remaining deposits to be exploited are becoming
more and more difficult and fewer and fewer, and the costs of calculations continue to fall considerably.
Computers are getting faster and faster and allow an increasing amount of data to be processed. Thus, three-
dimensional representations of the reservoirs are now simulated by geostatistical methods integrating data
from various sources outcrop, wells, seismic, dynamical data.
In the well-placement optimization context, Güyagüler and Horne (2004) used the utility framework with
Genetic Algorithm (GA) to maximize the NPV. Wang et al. (2012) proposed to maximize the NPV using
a Retrospective Optimization (RO) procedure consisting in increasing the number of geological realization
sequentially at each optimization iteration.
Within the well-placement optimization process, most the works are focused on the maximization of
the NPV and consider the risk only as an input variable due to geological uncertainties without thinking
of the risk in an economic point of view. In this paper, we take place in the concept introduced by the
portfolio optimization problems in finance that consists of analyzing a ‘mean-standard deviation’ model
of an objective function (Markowitz, 1952). Couet et al. (2004) apply this methodology to optimize the
waterflooding strategy. Capolei et al. (2015) propose a mean-variance objective function and evaluate the
‘risk-expected return’ through the Sharpe index (Sharpe, 1994).
Whatever the formulation of the objective function, the well-placement optimization is a process
involving a large number of reservoir simulation, moreover if geological uncertainties are taken into account
through many realizations of the geological model. In this context, a classical approach consists in using
of response surface model to bypass the reservoir simulator. Proxy models (e.g., kriging, polynomial,
SPE-198656-MS 3

and artificial neural network) have been widely used for reservoir engineering problems such as history-
matching, production forecasting, and production optimization. With the continued increase in computer
power, artificial intelligence techniques have rapidly met interest in the reservoir engineering community.
In 1988, Guérillot proposed an Expert System (ES) based on the concept of fuzzy logic to identify the most
suitable Enhanced Oil Recovery (EOR) process to apply in a given reservoir. Artificial neural networks

Downloaded from http://onepetro.org/SPEGOTS/proceedings-pdf/19GOTS/2-19GOTS/D021S007R001/1181910/spe-198656-ms.pdf by Universidade Federal De Pelotas user on 25 September 2024

(ANN) to replace the reservoirs simulator has been applied to optimization processes (Cullick et al. 2006;
Silva et al. 2007; Costa et al. 2010; Foroud et al. 2014; Bruyelle and Guérillot 2014, 2016, 2017). Guérillot
(2014) proposed a workflow to replace the calculation of thermodynamic equilibrium in reservoir simulation
by an ANN. Guérillot and Bruyelle (2018) proposed to speed-up reservoir simulations with an ANN as a
proxy model of the geochemical equilibrium solver in a reactive transport simulator.
In this work, the methodology proposed by Bruyelle and Guérillot (2019) for waterflooding optimization
is applied for well-placement optimization. This approach is in line with the concept introduced by the
portfolio optimization problems in finance that consists of analyzing a ‘mean-standard deviation’ model of
the NPV. This approach requires a considerable number of reservoir simulation to optimize the production
strategy for different formulations of the objective function. An artificial neural network has been used to
provide the average and standard deviation of the NPV and also to bypass the use of the reservoir simulator.
The proposed approach allows evaluating of all the formulation of the objective function to give the optimal
production strategies with the "risk-expected return" index associated. The risk is considered by the standard
deviation of the NPV for an ensemble of geological realizations.
In this paper, first, the "mean-standard deviation" objective function formulation for the well-placement
optimization problem and the Sharpe ratio are recalled. Then, the methodology that consists in to replace
the NPV calculations (mean and standard deviation) by an artificial neural network is presented. The
methodology to design the artificial neural network predicting the average and standard deviation of the
NPV is described. Finally, an application of this method for well-placement optimization to the Brugge
field is proposed.

Mean-Standard Deviation Objective Function

A classical formulation of the Net Present Value (NPV) in a well-placement optimization context is defined
as follows:

(1)

where X is a vector of optimization variables (i.e., well location, well production/injection parameters), T
is the number of time steps, is a time step, QO is the oil production rate, QW is the water production rate,
QO is the water injection rate, rO is the oil price, rw is the water production cost, rI is the water injection cost,
d is an annual discount rate, ti is the cumulative time, N is the number of new well and Cj is the drilling
cost applied to new wells.
Uncertainty is considered through the use of a set of reservoir models representing the variations of the
geological properties. For all these models, we suppose that the calibration of the history of production data
has been done. The objective is to provide an optimal localization of new wells in term of return/risk ratio.
The multi-criterion problem to solve consist of maximizing the average of the NPV and minimizing the
standard deviation of NPV. The aim is to find such as:
(2)
4 SPE-198656-MS

where X is a vector defining the optimization variables (i.e., well location, well production/injection
parameters), U is the search space, α is a coefficient between [0,1], Ω is a set of geological realizations, EΩ
is the average operator and σΩ is the standard deviation operator.
Following the value of α, the objective function advantage the maximization of the NPV or the
minimization of the standard deviation. A value of α equals to one correspond to the classical formulation

Downloaded from http://onepetro.org/SPEGOTS/proceedings-pdf/19GOTS/2-19GOTS/D021S007R001/1181910/spe-198656-ms.pdf by Universidade Federal De Pelotas user on 25 September 2024

of the objective function, i.e., maximization of NPV. Increase the value of α allows reducing the standard
deviation of NPV and decreasing the risk.

Sharpe Ratio
Sharpe (1994) defined the performance of one solution respect to its risk with an indicator measuring the
ratio of the gain expected and the standard deviation of the gain.

(3)

where rf is a benchmark return of a risk-free configuration.

Here, the benchmark return of a risk-free configuration is assumed to be the one obtained for the
objective function with, i.e., the configuration that minimizes the standard deviation. The Sharpe ratio will
be computed only for values of α greater than zero.
This ratio is an indicator of the performance of a given solution and can be interpreted as follow. A
negative value designates a solution with poorer performance than the least risky configuration, i.e., the
configuration that minimizes the standard deviation and corresponds to the solution obtains for the objective
function with α=0. A ratio value between 0 and 1 designates a solution with a risk too high compared to the
benefit expected. A ratio upper than 1 designates a solution with a high expected benefit for a low taken risk.
In an economic context, the best solution should be the one that maximizes the Sharpe ratio. Thus, the
definition of α does not matter. However, in a real industrial application, the operators want to know a
maximum of information on the gain and risk taken. The Sharpe ratio must be calculated for different value
of α, i.e., the well-placement optimization must be done for each value of α.

Methodology
The proposed methodology aims to provide an overview of the possible solutions with the associated "risk-
expected return" associated. This workflow implies to perform an optimization process for each formulation
of the objective function, i.e., for different value of α. Given the geological uncertainty resulting from a set
of geological realizations, each optimization process is all the longer as the number of reservoir models is
significant. To reduce the computation time, an artificial neural network as a proxy model is used. The use
of a proxy implies to generate a database representing the problem to reproduce. The ANN is then trained
using the learning database. An approximation of the objective function is designed with the outputs of the
ANN. Global optimizations are done for each value of α. To finish, an overview of the solutions with the
associated Sharpe index is done. The steps of the methodology are described below.
1. Artificial neural network. A multilayer perceptron is used as a proxy model providing the average
NPV and standard deviation of the NPV of an ensemble of geological realization for a given
configuration, i.e., the well location parameters (cf. Fig. 1). The ANN is composed of one hidden layer.
The activation function of the hidden neurons is sigmoidal. The activation function of the neurons of
the output layer is linear. Cybenko (1989) and Hornik (1993) have shown the universal approximation
capacity of such ANN;
2. Experimental design. To efficiently generate a learning database, a classical approach based on
experimental designs (ED) consists of select the data to cover the space of parameters to obtain the
SPE-198656-MS 5

most information with a minimum of data (Amudo et al. 2009). Following the type of proxy model,
some experimental designs are more appropriate to fit a model, e.g., Plackett-Burman experimental
designs suit to design first-order models, Box-Behnken experimental designs suit to design second-
order models. According to the complexity of the model, experimental designs consider or not the
interactions between parameters. Here, the Box-Behnken experimental design is used. This type of

Downloaded from http://onepetro.org/SPEGOTS/proceedings-pdf/19GOTS/2-19GOTS/D021S007R001/1181910/spe-198656-ms.pdf by Universidade Federal De Pelotas user on 25 September 2024

ED suits to fit quadratic models, considering the interactions between two factors, the linear and
squared terms. However, no assumptions about the model to reproduce have been made a priori, and
random samples are added to the learning database. The predictive quality of an ANN is dependent
on the quality and number of data in the learning base, i.e., the higher the number of learning data,
the better the prediction quality of the ANN. The simulations requested to constitute the learning
database are independent one to the other and can be quickly done in a parallel manner to reduce the
total computing time;
3. Learning of the ANN. The learning phase aims in to adjust the weights of the ANN in order to
the simulated outputs fit the learning outputs. One of the most method used to train an ANN is the
backpropagation algorithm (Rumelhart et al. 1995). This algorithm consists in to compute the total
error between the outputs provided by the ANN and the actual outputs given by the database. The
derivative of the error, combined with a learning rate and a momentum term, is used to modify the
weights of the ANN. This iterative process is repeated until a minimal error is reached. To avoid the
problem of over-learning, i.e., the inability of the network to predict outputs on non-learned data, a
validation dataset is used and the learning phase is stopped when the error on the validation dataset
increases. As long as the generalization capacity of the ANN is not satisfactory, the learning dataset
needs to be improved by adding new samples, e.g., with new random samples, and the learning phase
is repeated;
4. Response surface model of the objective function. The objective function is defined with the outputs
of the ANN, i.e., the average and standard deviation of the NPV. The problem to solve is to identify
X^*#U such as:
(4)
where is the first output of the ANN corresponding to the average of the NPV for a given
input X and is the second output of the ANN corresponding to the standard deviation of
the NPV for a given input X;
5. Optimization. The CMA-ES is used as an optimizer to maximize and avoid local minima of
the objective function (4). This evolutionary strategy based on the principle of natural selection
mechanism, evaluate the objective function for an ensemble of configuration, i.e., the well location,
and update the probability sampling distribution with the best individual, i.e., the well locations
that maximize the objective function. This iterative procedure is done until the solution reaches a
convergence criterion. This approach is high consuming in term of the number of evaluation of the
objective function. The coupling of the CMA-ES with the response surface model allows performing
optimizations for several values of α in a short time compared to classical approaches that use a
reservoir simulator. After convergence of the optimizer, the solution given by the optimizer using the
ANN is cross-checked with the reservoir simulator results. If the average and standard deviation of
the NPV are too different compared to those given by the simulator, the learning base is amended by
adding new random samples, and the process is reiterated (learning and optimization) until that the
precision of ANN is satisfactory;
6. Solution analysis. For each value of α, a well locations solution is given with the associated average
and standard deviation of the NPV. The ensemble of the couples "average - standard deviation" of the
6 SPE-198656-MS

NPV form the efficient frontier, i.e., the location of the wells that maximizes the average NPV for a
given level of risk. For each solution, the Sharpe ratio is given.

Downloaded from http://onepetro.org/SPEGOTS/proceedings-pdf/19GOTS/2-19GOTS/D021S007R001/1181910/spe-198656-ms.pdf by Universidade Federal De Pelotas user on 25 September 2024

Figure 1—The artificial neural network structure where the inputs correspond to the location of the wells, and well
production/injection parameter (pressure or rate), and the outputs are the average and the standard deviation of the NPV.

Figure 2—Flowchart of the well placement methodology: (1) Learning phase to design
an artificial neural network; (2) Optimization of well-placement for different value of α
using the ANN; (3) Solution analysis to identify the best "risk-expected return" solution.

Brugge Field
The presented methodology is applied to optimize the well-placement to the Brugge field. This
synthetic case study has been provided by TNO to benchmark methods addressing history-matching and
waterflooding optimization problems. The geological description of the case is given by Peters et al. (2010).
SPE-198656-MS 7

The Brugge field properties are similar to the Brent field located in the North Sea. The reservoir is defined
with four zones coming from multiple deposit environments.
Thirty vertical wells (20 producers and 10 injectors) has been drilled. First, the 20 producers are drilled
from 1998 to 1999, then each month an injector well is drilled. The first ten years of production data are
available. To evaluate the methodology, the existing injection wells have been removed from the Brugge

Downloaded from http://onepetro.org/SPEGOTS/proceedings-pdf/19GOTS/2-19GOTS/D021S007R001/1181910/spe-198656-ms.pdf by Universidade Federal De Pelotas user on 25 September 2024

model. The methodology is applied to identify the optimal location of these injector wells. The injection
rate and the date of drilling of these wells are assumed to the same that the initial model. A comparison
of results is made between (1) the production without optimization, i.e., the reference case, and (2) the
production obtained with the proposed method. The proposed exercise aims to highlight that the mean-
standard deviation formulation, coupled with an accurate proxy model, can provide an overview of the risk-
expected return.
The economical parameters of the NPV function used are the same as those given for the production
optimization problem (Table 1).

Table 1—Economic parameters of NPV function

NPV parameters Value

Oil price 80 $/bbl

Water production cost 5 $/bbl
Water injection cost 5 $/bbl
Annual discount rate 10%

To generate the neural network learning database, a Box-Behnken experiment plan has been used (761
configurations) and randomly generated configurations in the parameter space (250 configurations). Fifty
randomly generated configurations were used to build the validation database. For each configuration of
the placement of wells, a flow simulation is performed for the 50 reservoir models assumed to represent the
uncertainty associated with the geological parameters. The output data are the average of the NPV and the
standard deviation of the NPV of the 50 simulations.
Figure 3 shows the ability of the ANN to reproduce the learning data and to predict the validation data
for the average and the standard deviation of the NPV for an ensemble of geological models.

Figure 3—Cross-validation plots showing the ability of the ANN to reproduce the
learning data (•) and the validation data (˚) for average EΩ[NPV] (left) and σΩ[NPV] (right).
8 SPE-198656-MS

Optimization of the placement of the wells is done, for ten values of α, using the outputs of the ANN
to compute the value of the objective function. Figure 4 shows the efficient frontier for the mean-standard
deviation optimizations and associated Sharpe ratio for ten different value of alpha α. The optimized
solutions for α=0.1 et α=0.2 have a Sharpe ratio between 0 and 1, i.e., a risk too high compared to the
expected return. For α ≥ 0.3, the optimized solutions have a Sharpe ratio higher than one, i.e., a solution

Downloaded from http://onepetro.org/SPEGOTS/proceedings-pdf/19GOTS/2-19GOTS/D021S007R001/1181910/spe-198656-ms.pdf by Universidade Federal De Pelotas user on 25 September 2024

with a high expected benefit for a low taken risk. The higher Sharpe ratio is for α equals 0.3. The NPV of
the optimized solution has been increased by 2.88% compared to the base case.

Figure 4—Efficient frontier for the mean-standard deviation optimizations

(left) and associated Sharpe ratio (right) for different value of α.

Conclusions
Plan the position of a new well is an essential task in order to increase hydrocarbon production. This
optimization problem requires a high number of reservoir simulation to determine the appropriate location.
In an economic context, it is essential to evaluate the risk to reach an expected NPV. The mean-standard
deviation analysis, well known for portfolio optimization problems in finance, allows evaluating this
risk. For the well placement optimization where the uncertainty is taken into account considering a set a
geological realization, the mean-standard deviation analysis is costly in term of the number of reservoir
simulations.
In this paper, a methodology to address the problem of well placement optimization using the mean-
standard deviation formulation, coupled with an artificial neural network as a substitute to the reservoir
simulator is presented. The ANN provides the average NPV and standard deviation of the NPV of an
ensemble of geological realization for a given wells configuration. The CMA-ES optimizer is coupled with
the ANN to speed-up the optimization processes for many formulations of the objective function (i.e.,
different values of α). This approach allows providing a complete overview of the information on the gain
expected and risk is taken through the efficient frontier and the Sharpe ratios.
The workflow has been applied for optimization of the injector wells location of the Brugge field.

Acknowledgments
This project was supported by Texas A&M University at Qatar. We acknowledge Schlumberger for their
donation of Petrel software and Eclipse 100 simulator.
SPE-198656-MS 9

Nomenclature
Cj = drilling cost applied to new well j [$]
d = annual discount rate [-]
J = objective function
= approximation of the objective function

Downloaded from http://onepetro.org/SPEGOTS/proceedings-pdf/19GOTS/2-19GOTS/D021S007R001/1181910/spe-198656-ms.pdf by Universidade Federal De Pelotas user on 25 September 2024

N = number of new wells
NPV = Net Present Value [$]
QO = oil production rate [bbl.day^(-1)]
QW = water production rate [bbl.day^(-1)]
QO = water injection rate [bbl.day^(-1)]
rf = benchmark return for a risk-free configuration [$]
rI = water injection cost [$.bbl^(-1)]
rO = oil price [$.bbl-1]
rw = water production cost [$.bbl-1]
Sh = Sharpe ratio
ti = cumulative time [day]
U = search space
X = vector defining the control parameters
α = coefficient between [0,1]
= time step [day]
σΩ = standard deviation operator
Ω = set of realizations of the reservoir properties
EΩ = average operator

References
Alrashdi, Z. and Sayyafzadeh, M. (2019). (μ+ λ) Evolution strategy algorithm in well placement, trajectory, control and
joint optimization. Journal of Petroleum Science and Engineering, 177, 1042-1058.
Amudo, C., Graf, T., Dandekar, R. R. and Randle, J. M. 2009. The pains and gains of experimental design and response
surface applications in reservoir simulation studies. In SPE Reservoir Simulation Symposium. Society of Petroleum
Engineers
Awotunde, A. A. and Sibaweihi, N. (2014). Consideration of voidage-replacement ratio in well-placement optimization.
SPE Economics & Management, 6(01), 40-54. doi:10.2118/163354-PA
Bouzarkouna, Z., Ding, D. Y. and Auger, A. 2013. Partially Separated Metamodels With Evolution Strategies for Well-
Placement Optimization. Society of Petroleum Engineers. doi:10.2118/143292-PA
Bruyelle, J. and Guérillot, D. 2014. Neural networks and their derivatives for history matching and reservoir optimization
problems. Computational Geosciences, 18(3-4), 549-561
Bruyelle, J. and Guérillot, D. 2019. Optimization of Waterflooding Strategy Using Artificial Neural Networks. SPE
Reservoir Characterisation and Simulation Conference and Exhibition (RCSC 2019)
Bruyelle, J. and Guérillot, D. 2019. Proxy Model Based On Artificial Intelligence Technique For History Matching -
Application To Brugge Field. Gas & Oil Technology Showcase and Conference (GOTECH 2019)
Capolei, A., Suwartadi, E., Foss, B. and Jørgensen, J. B. 2015. A mean–variance objective for robust production
optimization in uncertain geological scenarios. Journal of Petroleum Science and Engineering, 125, 23-37
Costa, L. A. N., Maschio, C. and Schiozer, D. J. 2010. Study of the influence of training data set in artificial neural network
applied to the history matching process. In Rio Oil & Gas Expo and Conference
Couet, B., Burridge, R. and Wilkinson, D. 2004. U.S. Patent No. 6,775,578. Washington, DC: U.S. Patent and Trademark
Office.
Cybenko, G. 1989. Approximation by superpositions of sigmoidal functions.Mathematics of Control, Signals, and Systems
Emerick, A. A., Silva, E., Messer, B., Almeida, L. F., Szwarcman, D., Pacheco, M. A. C. and Vellasco, M. M. B. R.
2009. Well placement optimization using a genetic algorithm with nonlinear constraints. In SPE reservoir simulation
symposium. Society of Petroleum Engineers.
10 SPE-198656-MS

Foroud, T., Seifi, A., and AminShahidi, B. 2014. Assisted history matching using artificial neural network based global
optimization method–Applications to Brugge field and a fractured Iranian reservoir. Journal of Petroleum Science
and Engineering, 123, 46-61
Forouzanfar, F., Poquioma, W. E. and Reynolds, A. C. 2016. Simultaneous and sequential estimation of optimal
placement and controls of wells with a covariance matrix adaptation algorithm. SPE Journal, 21(02), 501-521.
doi:10.2118/173256-PA

Downloaded from http://onepetro.org/SPEGOTS/proceedings-pdf/19GOTS/2-19GOTS/D021S007R001/1181910/spe-198656-ms.pdf by Universidade Federal De Pelotas user on 25 September 2024

Groult, J., Reiss, L. H. and Montadert, L. 1966. Reservoir inhomogeneities deduced from outcrop observations and
production logging. Journal of Petroleum Technology, 18(07), 883-891.
Guérillot, D. R. 1988. EOR Screening With an Expert System. Society of Petroleum Engineers. doi:10.2118/17791-MS
Guérillot, D. 2014. Method and system for dynamically modeling a multiphase fluid flow – European patent
EP2791712B1 - US Patent App. 14/365,053
Guérillot, D. and Bruyelle, J. 2018. Geochemical Equilibrium Determination Using an Artificial Neural Network in
Compositional Reservoir Flow Simulation. In ECMOR XVI-16th European Conference on the Mathematics of Oil
Recovery. doi: 10.3997/2214-4609.201802232, to appear by Springer10.1007/s10596-019-09861-4
Guérillot, D. R. and Bruyelle, J. 2017. Uncertainty Assessment in Production Forecast with an Optimal Artificial Neural
Network. Society of Petroleum Engineers. doi:10.2118/183921-MS
Guérillot, D. R. and Bruyelle, J. 2016. History matching methodology using an optimal neural network
proxy and a global optimization method. In Third EAGE Integrated Reservoir Modelling Conference. doi:
10.3997/2214-4609.201602403
Güyagüler, B. and Horne, R. N. 2004. Uncertainty Assessment of Well-Placement Optimization. Society of Petroleum
Engineers. doi:10.2118/87663-PA
Hansen, N. and Ostermeier, A. 2001. Completely derandomized self-adaptation in evolution strategies. Evolutionary
computation, 9(2), 159-195
Hornik, K. 1993. Some new results on neural network approximation. Neural networks, 6(8), 1069-1072.
Isebor, O. J., Durlofsky, L. J. and Ciaurri, D. E. 2014. A derivative-free methodology with local and global search for the
constrained joint optimization of well locations and controls. Computational Geosciences, 18(3-4), 463-482.
Markowitz, H. 1952. Portfolio selection. The journal of finance, 7(1), 77-91.
Montadert, L. 1963. La sédimentologie et l'étude détaillée des hétérogénéités d'un réservoir: application au gisement
d'Hassi-Messaoud. Rev. Inst. Franc. Pétrol., Paris, 17, 241-257.
Peters, L., Arts, R.J., Brouwer, G.K., Geel, C.R., Cullick, S.,Lorentzen, R.J., Chen, Y., Dunlop, K.N.B., Vossepoel,
F.C.,Xu, R., Sarma, P., Alhutali, A.H. and Reynolds, A.C. 2010. Results of the Brugge benchmark study for flooding
optimization and history matching. SPE Reserv. Evalu. Eng.13(3), 391–405
Ravenne, C., Eschard, R., Galli, A., Mathieu, Y., Montadert, L. and Rudkiewicz, J. L. 1989. Heterogeneities and geometry
of sedimentary bodies in a fluvio-deltaic reservoir. SPE Formation Evaluation, 4(02), 239-246
Ravenne, C., Galli, A., Beucher, H., Eschard, R. and Guérillot, D. 1991. Outcrop studies and geostatistical modelling
of a middle Jurassic Brent analogue. In Proceedings of the European Oil and Gas Conference, A Multidisciplinary
Approach in Exploration and Production R&D: London Graham and Trotman (pp. 497-520)
Rumelhart, D. E., Durbin, R., Golden, R. and Chauvin, Y. 1995. Backpropagation: The basic theory. Backpropagation:
Theory, architectures and applications, 1-34.
Sarma, P. and Chen, W. H. 2008. Efficient Well Placement Optimization with Gradient-based Algorithms and Adjoint
Models. Society of Petroleum Engineers. doi:10.2118/112257-MS
Sharpe, W. F. 1994. The Sharpe ratio. Journal of portfolio management, 21(1), 49-58
Silva, P. C., Maschio, C. and Schiozer, D. J. 2007. Use of neuro-simulation techniques as proxies to reservoir simulator:
application in production history matching. Journal of Petroleum Science and Engineering, 57(3-4), 273-280
Tomutsa, L., Jackson, S. R., Szpakiewicz, M. and Palmer, T. 1986. Geostatistical characterization and comparison of
outcrop and subsurface facies: Shannon shelf sand ridges. In SPE California Regional Meeting. Society of Petroleum
Engineers.
Wang, H., Echeverría-Ciaurri, D., Durlofsky, L. and Cominelli, A. 2012. Optimal well placement under uncertainty using
a retrospective optimization framework. SPE Journal, 17(01), 112-121.
Zandvliet, M., Handels, M., van Essen, G., Brouwer, R. and Jansen, J. D. 2008. Adjoint-based well-placement optimization
under production constraints. SPE Journal, 13(04), 392-399. doi:10.2118/105797-PA