OTC-28002-MS

Multi-Scale Ensemble-Based Data Assimilation for Reservoir

Characterization and Production Forecast: Application to a Real Field

Alexandre de Lima, Theophile Gentilhomme, David Riffault, and Fabien Allo, CGG; Alessandra S. Anyzewski,
Alexandre A. Emerick, and Marcos S. dos Santos, Petrobras

Copyright 2017, Offshore Technology Conference

This paper was prepared for presentation at the Offshore Technology Conference Brasil held in Rio de Janeiro, Brazil, 24–26 October 2017.

This paper was selected for presentation by an OTC 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 Offshore Technology Conference and are subject to correction by the author(s). The material does not necessarily reflect any
position of the Offshore Technology Conference, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Offshore Technology Conference 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 OTC copyright.

Abstract
The paper presents the application of multi-scale Levenberg-Marquardt ensemblerandomized maximum
likelihood(MS-LM-EnRML) method to an offshore field in Campos Basis. The objective of this method
is to update an ensemble of spatial properties and flow parameters using observed productiondata while
preserving the prior information already carried by the prior models. The final ensemble provides an
estimate of the uncertainties and can then be used for productionforecast andinfill drilling projects. Instead of
working directly with the gridblock values of the reservoir, the MS-LM-EnRML method performs a wavelet
transform of the spatial properties. Based on this parameterization, a multi-scale approach is followed:
different subsets of wavelet coefficients are successively updated with the data using the Levenberg-
Marquardt ensemble-based data assimilationmethod. Thanks to the compression property of the wavelets,
the field properties are well characterized with a small number of wavelet coefficients. The algorithm
starts by optimizing an initial sub-set of large-scale coefficients. The parameterization is then progressively
refined, and a new optimization is run on the new subset of coefficients. Because the data mismatch is
highly reduced by the initial optimizations at large scales, the perturbation of the fine scale coefficients is
minimized resulting in a better preservation of the prior models and a reduced level of noise in the final
models.
For comparisons, a second history matching was conducted using the ensemble smoother with multiple
data assimilation (ES-MDA) method without the multi-scale parameterization. The quality of the data
match, the preservation of the prior models and the ensemble variance are used to compare the two methods.
The results showed that although a better data matcheswereobtained with ES-MDA, MS-LM-EnRML
preserved better the prior models, the geological consistency and the variability of the ensemble. These
results indicate that the multi-scale approach is an attractive option especially when a large number of data
has to be assimilated (e.g., long production history and 4D seismic data) or when geological features and
seismic information of the initial models need to be preserved.
2 OTC-28002-MS

Introduction
Ensemble-based data assimilation (Evensen, 2003; Aanonsen et al., 2009; Oliver and Chen, 2011) have
become a popular method for history-matching thanks to its simplicity, flexibility and ability to use different
source of data associated with different simulation processes (e.g. reservoir flow simulator, geophysical
processes, etc.). These methods start with an initial ensemble of realizations, also called the prior ensemble,
reflecting the prior uncertainties and correlations of spatial properties (e.g. porosity, permeability, net-gross,
etc.) and flow parameters controlling input of flow simulator (e.g. relative permeability, fluid contacts, rock
compressibility, etc.). The prior ensemble integrates the geological knowledge about the reservoir as well
static data fromwell logs and seismic. In the ideal case, the prior ensemble should be able to represent the
uncertainty before the assimilation of the dynamic data (e.g., production and 4D seismic data) in a process
known in the petroleum literature as history matching.
The information carried by the prior ensemble is valuable and should be preserved as much as possible
during the history-matching process, in order to characterize the field maintaining the geological consistency
represented by the initial realizations meanwhile all the different data are assimilated. Consistency is an
assurance of reliability of the model parameters and predictions. However, ensemble methods can be quite
destructive when the number of realizations is limited (Gentilhomme et al., 2015). In fact, these methods use
an ensemble to statistically estimate correlations between the forward model outputs (e.g. simulated well
production or seismic responses) and the parameters of the models. Typical history-matching problems use a
limited number of realizations as a fluid flow simulation has to be run for each ensemble realization at every
update of the parameters, which is computationally expensive. As a consequence, the size of the ensemble
(usually hundreds of realizations) is much smaller than the number of reservoir parameters (millions of
variables) resulting in noisy updates of the model parameters due to spurious correlations, which damage the
consistency of the ensemble with the prior data and significantly reduce the its variability and its capacity
to predict the uncertainties. In extreme cases, a collapse of the ensemble may be observed, in which case
it is no longer possible to assimilate further data.
MS-LM-EnRMLwas proposed by Gentilhomme et al. (2015) to address this problem. A wavelet
transform is used to decompose the spatial properties into coefficients that are both localized in space and
frequency. The method starts by updating the large scale coefficients, less subject to spurious correlations
as they highly impact the flow response (Chen and Oliver, 2012), and progressively includes finer scale
coefficients. The objective of the method is to reduce as much as possible the mismatch by updating the large
scales so that after refinement, the perturbations of the fine scales, more subject to the spurious correlations,
are minimized. Although the re-parameterization helps identifying sensitive parameters, it does not remove
the spurious correlations. As a good match is generally not obtained by updating the large scales, distance-
based localization (Chen and Oliver, 2010) is applied to reduce the effects of the spurious correlations on the
parameters. Distance-based localization assumes that the correlations between the data and the parameters
are a function of their separating distance. Since the wavelet coefficients are localized in space, distance-
based localization can be applied but has to be adapted to the multi-scale approach.
In this paper, the prior ensemble is derived from geological models updated with 3D acoustic data
characterizing the properties of a large turbidite reservoir in Campos Basin, offshore Brazil. The objective
of this study is to update the prior ensemble using the production data from 36 wells with the MS-LM-
EnRML method. A second history matching is performed using the method known as ensemble smoother
with multiple data assimilation (ES-MDA) (Emerick, 2013). The quality of the data match, the preservation
of the prior models and the posterior variance are used to compare the results.
The remaining of the paperis organizedas follows. The next section presents a brief overview of the
methods MS-LM-EnRMLand ES-MDA followed by a discussion on the localization process for each
technique. Then, the field case is introduced. After that, the results of the data assimilation with both methods
are discussed. The last section of the paper summarizes the conclusions.
OTC-28002-MS 3

Ensemble Data Assimilation

MS-LM-EnRML and ES-MDA are methods that iteratively update the model parameters, here denoted by
a vector m, using cross-covariances statistically estimated from the ensemble. Instead of assimilating the
temporal data sequentially, as in the standard ensemble Kalman filter (Evensen, 2003), MS-LM-EnRMLand
ES-MDA assimilate all the data at once through an iterative process. Both methods are based on the standard
ensemble smoother (ES) analysis equation

(1)

where CMD is the (np × nd) cross-covariance matrix between model parameters and the simulated responses,
with np and corresponding to the number of model parameters and the number of data points, respectively.
CDD is the (nd × nd) auto-covariance matrix of the simulated responses. Both matrices are estimated based
on the ensemble of dimension ne CD. is the (nd × nd) observed data-error covariance matrix; dobs,j is a (nd
× 1) vector of perturbed observations. This vector is computed as a sample from N (dobs, CD) for MS-LM-
EnRML and from N(dobs, αCD) for ES-MDA, where N(.,.) denotes a Gaussian distribution and dobs is the (nd
× 1) observation vector. mj and mju represent (np × 1) vectors oparameters for the jth realization before and
after data assimilation, respectively. g(mj) is the forward response (vector of simulated production data). K
is the so-called Kalman gain matrix.
SinceMS-LM-EnRML and ES-MDA are iterative processes, they better handle nonlinearity than the
standard ES. For each iteration, the update of the parameters is controlled by modifying the inverse term of
the Kalman gain. The ES-MDA uses an inflation α coefficient so that:
(2)
whereas the MS-LM-EnRML uses a Levenberg-Marquardt control parameter λj so that
(3)
The control parameters α and λ have very similar impact on the update in the two methods. They control
the amplitude and orientation of the parameter updates by finding the best update in a limited region around
the current state, which helps handle nonlinearity. Impact and control of these parameters are discussed in
more details in (Emerick and Reynolds, 2013; Gentilhomme et al., 2015; Oliver et al., 2008). One difference,
however, is that ES-MDA requires the sum of the inverse of to be equals to one.
One of the main problems with ensemble-based methods are spurious correlations present in the matrices
CMD and CDD. As a reservoir simulation is required for each realization of the ensemble, the number
of evaluations of the forward model (few hundreds) is very limited compared to the number of model
parameters. This leads to correlations between the parameters and the flow responses that do not exist,
but artificially appear in the ensemble because of its limited size. These correlations cause an abnormal
reduction of variability in the ensemble after each update of the parameters, and since they are random, they
introduce noise in the properties that is shared by all the realizations (and therefore visible in the ensemble
mean) as the same regression is used to update the entire ensemble. In order to help limiting the impact of
these spurious correlations, re-parameterization and localizationare used.

Scale decomposition and scale-by-scale optimization

Eqs. 2 and 3 show that the updating equations of MS-LM-EnRML and ES-MDA are very similar. The main
difference is in the parameterization of the problem. While ES-MDA updates the realizationsdirectly on the
gridblock values, MS-LM-EnRML decomposes the spatial properties using wavelet transforms (Sweldens,
1998). Moreover, in the ES-MDA all parameters are updated during all iteration. For MS-LM-EnRML, on
4 OTC-28002-MS

the other hand, the update equation(Eq. 1) is applied on a subset of selected coefficients γopt of size nopt ≤
np, such that

(4)

where W is the (np × np) wavelet transform matrix and γopt and are the (nopt × 1) vectors of wavelet
coefficients updated at the assimilation step and its complement vector, respectively. For sake of simplicity,
we assume here that m only contains spatial properties. Thanks to sparse representation property of the
wavelet transform, the spatial properties can be well represented only using a subset of wavelet coefficients.
Thus, by perturbing a limited number of coefficients, it is possible to highly impact the flow response of
the models. Although, the coefficients are not inverted, they are used to reconstruct the properties used
by the reservoirsimulator. Their effects on the property reconstructions are however weakened so that no
bias is introduced when computing the correlations between the optimized parameters γopt and the simulated
responses (Gentilhomme et al., 2015).
The multi-scale approach starts by selecting a subset of large scale parameters (Fig. 1). As previously
discussed in this article and argued by Chen and Oliver (2012), large scale coefficients are highly sensitive
parameters and their corresponding correlations with the flow responses are well estimated even with a small
ensemble. Therefore, it is possible to improve the match without corrupting the ensemble (limited impact
of the spurious correlations). However, anacceptable data match may not be obtained by only updating the
large scales. Fine scale features close to the wells or defining structures (e.g. barrier or drain) impacting the
fluid flow can have an important impact on the response. Therefore, the parameterization is progressively
refined by finer scale coefficients (Fig. 1). After each refinement, an optimization is performed on the new
subset of coefficients and the mismatch is reduced. Thus, the perturbations of the finest scale coefficients
that are highly subject to spurious correlations, is limited as the mismatch has been reduced (Eq. 1). More
details about the MS-LM-EnRML can be found in (Gentilomme et al. 2015).

Figure 1—Scale-by-scale optimization. Optimization is performed on a subset of large scale

wavelet coefficients leaving the others scales unmodified. This subset is progressively refined
by included finer scale wavelet coefficients. Optimization is performed after each refinement.

One drawback of the scale-by-scale approach is that all the sensitive parameters are not directly included
at the beginning of the optimization (they are unknown). As a consequence, the process may have to wait
for refinement in order to update sensitive parameters, which leads to a slower convergence than with the
standard approach where all the parameters (sensitive and insensitive) are included from the beginning
of the iterations. Although the scale-by-scale approach helps minimize the perturbation of the fine scale
coefficients, it does not remove the spurious correlations. Therefore, localizations process need to be applied
in both MS-LM-EnRML and ES-MDA.
OTC-28002-MS 5

Localization
Localization aims to regularize the Kalman gain by removing the components affected by the spurious
correlations inherent to the limited size of the ensemble used to construct it. Localization can be applied
to the individual covariance matrices of the Kalman gain or directly to the Kalman gain itself (Chen and
Oliver, 2010). The second approach is used in this work with a tempering matrix, A, containing elements
in [0, 1] applied to K such that
(5)
where is the regularized Kalman gain and ∘ is the element-wise Schur product. In distance-based
localization, the elements of A are related to the separating distance of the data location (e.g. well
locations) and spatial parameters (e.g. gridblocks). Each column of K corresponds to the perturbations of
the parameters weighted by the data mismatch (Eq. 1) for a given update step. The goal of the localization
is to remove the perturbations induced by the spurious correlations.
Standard distance-based localization method is applied to ES-MDA where only one localization function
(i.e. column of A) is defined for each datum. A different approach is used in MS-EnOpt. The columns of
the matrix A are not fixed, but they depend on the current scale to be optimized. As mentioned before, the
estimated correlations between the flow response and the large scale coefficients are usually reliable. In
this case, no localizations need to be applied and A only contains values of one. However, after refinement,
the Kalman gain becomes more and more subject to spurious correlations, and the localization has to be
more restrictive. The range (i.e., the distance from which the impact of data becomes negligible) is reduced.
Note that before applying this localization, the Kalman gain K must be reconstructed as is defined in the
gridblock space (see Gentilhomme et al., 2015).
A second localization is directly applied in the wavelet space for the MS-LM-EnRML method. It uses
sensitivity analysis of the wavelet coefficients and spatial coverage. For example, the localization will
discard the correlations of the coefficients with a spatial coverage that does not overlap a given data location
(e.g. well position). Thus, the correlation corresponding to large scale coefficients (with large coverage) are
more likely to be preserved whereas the correlations corresponding to fine scale coefficients (with small
coverage) are only maintained close to the well locations. The gridblock localization only limits the update
of the parameters inside pre-defined areas without distinguishing between frequencies, whereas the wavelet
localization only restricts the updates of some frequencies at some locations, but allows global modifications
of the property fields. More information can be found in (Gentilhomme et al., 2015).

Field Case
The field is a large heavy-oil turbidite reservoir located in Campos Basin. Reservoir rocks are mainly
composed of poorly consolidated sandstones with very high porosity and permeability and low clay content.
Currently there are 36 wells operating in this field. The field production startedwith a pilot well in 1994.
Most of the wells only started operation six years later. After the first six year of depletion, injector wells
were placed in the field considering water flooding as the main recovery mechanism. More details about
this field can be found in (Bruhn et al, 2001; Emerick, 2016). This is the same reservoir model history
matched in (Emerick, 2016). However, there are some modifications compared to the previous work. First,
the production history has been updated with five additional years. Second, the transmissibility across
several faults not present in the previous study havebeen included as history matching parameters. Finally,
the prior ensemble of petrophysical properties has been updated by the incorporation data from a 3D seismic
in a petrophysical inversion process illustrated in Fig. 2.
6 OTC-28002-MS

Figure 2—Generation of the prior ensemble used for history-matching. (a) Realizations
generated by SGS conditioned to the well log. (b) Ip realizations generated by stochastic seismic
inversion. (c) Petrophysical inversion based on a petro-elastic model to update the realizations.

The petrophysical inversion started with an initial ensemble of 200 realizations generated by sequential
Gaussian simulation constrained by the well log data (Fig. 2a). Porosity, permeability and net-to-gross are
generated by this process. In order to sample the uncertainties related to the field geology, the realizations
were created using different variograms; four subsets of realizations were generated using short, medium,
large and average rangevariograms, respectively. Moreover, the permeability fields were generated using
facies proportion as a trend map for soft conditioning. The vertical permeability was derived from the
horizontal permeability using an anisotropy ratio.
Porosity and net-to-gross of this ensemble were conditioned to acoustic seismic derived attributes (P-
impedance, Ip) through apetrophysical inversion process. The inversion used the results of a stochastic
seismic inversion (Buland and Orme, 2003), which produces an ensemble of fine scale Ip realizations
constrained by the seismic amplitudes (Fig. 2b). During the petrophysical inversion, each petrophysical
geostatistical realization was constrained by one Ip realization. This process is called randomized maximum
likelihood (RML, Oliver et al., 2008). The petrophysical inversion uses a petro-elastic model (PEM) as
forward model. For a given configuration of petrophysical values, the PEM provides an elastic response.
In this work, the PEM was based on an unconsolidated sandstone model calibrated with well logs and lab-
measurements.
The petrophysical inversion uses a novel sampling based on optimization process to update the
petrophysical realizations (Gentilhomme, 2017). This method allows the sampling of the uncertainties
associated with the petrophysical inversion associated with the RML approach in order to produce a seismic
constrained ensemble with large variability. Sinceno prior correlation was observed between the seismic
characterized petrophysical parameters (porosity and net-to-gross) and the permeability, the permeability
fields were not updated at this step. Table 1 presents the ranges determined for the petrophysical properties
in the data assimilation.

Table 1—Petrophysical parameters and its respective ranges

OTC-28002-MS 7

Gentilhomme et al. (2017) presents a detailed discussion on the petrophysical inversion used in the study.
Besides the petrophysical realizations, the history matching also includes transmissibility multipliers
across 100 faults and the water-oil relative permeability curve, which was modeled using a Corey-type
function, i.e.,

(6)

where, krw is the relative permeability of the water phase, krw, max is the maximum water relative permeability,
Sw is the water saturation, Swi is the initial water saturation, Sor is the residual oil saturation and ew is the
exponent of the Coreyfunction. Table 2 presents the prior distribution used to generate the initial ensemble
of fault transmissibility and water relative permeability.

Table 2—Scalarparameters and their respective distribution

Observed data corresponds to measured oil and water rate, gas-oil ratio and bottom-hole pressure at the
oil producing wells and water rate and bottom-hole pressure at the injection wells. Table 3 summarizes the
observed data used with their assumed respective errors, which were used to construct the elements of the
matrix CD (Eq. 1).

Table 3—Observed data and their respective measurement errors

History Matching Settings

The main difference between MS-LM-EnRML and ES-MDA comes from the parameterization of the
spatial properties. ES-MDA works directly on the property values inside each gridblock ofthe model and
all iterations are applied in the finniest scale, which englobes all frequencies. On the other hand, MS-LM-
EnRMLuses wavelet transformations and decomposes the property in several scales, thus the methoditerates
through different scales, starting with a reduced parameterization.
Both methods were run with 4 iterations. In the MS-LM-EnRMLcase, the total number of iterations was
distributedthrough three different scales, one iteration was used foreach intermediate/coarse scale and two
iterations for the finniest scale. For ES-MDA we used a constant coefficient α = 4 for all iterations.
8 OTC-28002-MS

Another important difference between MS-LM-EnRML and ES-MDA is the localization. ES-MDA
usesthesame localization for all iterations, whileMS-LM-EnRML uses ascale-adaptive localization, which
becomes more restrictive from the coarsest scale to the finniest scale. Fig. 3 shows an example of localization
applied to one well of the field. For ES-MDA, we tested two localization functions. The first one corresponds
to the localization used in the intermediate scale of MS-LM-EnRML (Scale 2). The second localization
was computed using a more restrictive correlation range. Fig. 4 presents the average porosity after history
matching with ES-MDA using both localization functions. Both cases resulted in acceptable data matches.
However, localization 1 resulted in several artifacts while localization 2 resulted in a smoother posterior
mean. For this reason, localization 2 is used in the remaining results presented in this paper.

Figure 3—Localization images for a particular well. (a) Scale adaptive localization used by the MS-LM-EnRML.(b) Two
alternative localizations tested for ES-MDA. Localization 1 was defined as the intermediate localization of MS-LM-
EnRML (Scale 2) and Localization 2 was defined using a more restrictive correlation function with range of 2 kilometers.

Figure 4—Average porosity of the prior and posterior ensembles. (a) ES-MDA with localization 1 and (b) localization 2.

Results
Fig. 5 shows the histograms of data mismatch objective function obtainedby the prior and the posterior
ensembles generated by MS-LM-EnRML and ES-MDA. Both methods reduced the objective function.
However ES-MDA resulted in a better data match. The average objective function from ES-MDA is 2.2
while MS-LM-EnRML obtained a distribution of objective function with mode around 6.7. Fig. 6 shows the
water-cut for a selected well of the field illustrating the better data match obtained by ES-MDA. Fig. 7 shows
the gas-oil ratio (GOR) for another well of the field indicating that both methods improved the data match
compared to the prior, but ES-MDA resulted in a predicted GOR in closer agreement with the observations.
OTC-28002-MS 9

Figure 5—Histograms of the data mismatch objective function.

Figure 6—Water-cut (Well 2).

Figs. 8 and 9 show the ensemblemean and standard deviation of porosity before and after data assimilation
with MS-LM-EnRML and ES-MDA. These figures show that MS-LM-EnRML resulted in a smoother
posterior mean which is closer to the prior mean. Moreover, MS-LM-EnRML resulted in larger posterior
standard deviation evidencing a larger variability in the final ensemble. Note that the ensemble variability
may be interpreted as a simplified measure of uncertainty. Therefore, it is desirable that the variability is not
10 OTC-28002-MS

severely reduced after data assimilation. Otherwise, we could underestimate the uncertainty in production
forecasts. Moreover, keeping the ensemble variability is important to allow the assimilation of further data.
In fact, we are currently working on assimilating data from two 4D seismic acquisitions in this model and
our previous experience with the same field shows that keeping the ensemble variability with large amounts
of data is a critical issue; see, e.g., Emerick (2016) where it is shown that the assimilation of production and
seismic data resulted in an almost collapse of the ensemble.

Figure 7—Gas-oil ratio (Well 3).

Figure 8—Ensemble mean of porosity

OTC-28002-MS 11

Figure 9—Ensemble standard deviation of porosity

Fig. 10 shows the average permeability and Fig. 11 shows the permeability for a selected realization of
the ensemble. These figures show that MS-LM-EnRML preserved the prior permeabilities better than ES-
MDA. An important difference between the permeability fields obtained MS-LM-EnRML and ES-MDA is
around the region of the Well 1 (Fig. 10). Note that ES-MDA introduced a severe reduction of permeability
around this well while MS-LM-EnRML resulted in very little change compared to the prior model. Fig. 12
shows the observed and predicted bottom-hole pressure (BHP) at this well. According to this figure, the
prior ensemble predicted biased BHP compared to the observations. This is clearly an anomalous situation
which requires a revision of the prior parameterization (e.g., investigation on existing of local barriers or
even the presence of a severe skin in the well) before proceeding with data assimilation. Because MS-LM-
EnRML searches to preserve long-range features of the model, it did not introduce the severe reduction in
permeability observed in the model generated by ES-MDA. As a result, MS-LM-EnRML did not improve
the data match of BHP at this well.

Figure 10—Ensemble mean of horizontal permeability.

12 OTC-28002-MS

Figure 11—Horizontal permeability of realization 51.

Figure 12—Bottom-hole pressure of Well 1.

OTC-28002-MS 13

Conclusion
This paper presented the application of the multi-scale Levenberg-Marquardt ensemblerandomized
maximum likelihood method to a real field case. For comparison, we also applied data assimilation without
the multi-scale parameterization using the method known as ES-MDA. The results warranted the following
conclusions:

• The multi-scale parameterization preserves more features of the prior models, resulting in smoother
realizations with fewer localized changes in the petrophysical properties.
• The multi-scale method also resulted in a large variance in the posterior ensemble, which is
important for uncertainty quantification and assimilation of large amounts of data.
• The better preservation of the prior and the ensemble variance of the multi-scale approach came
with the cost of inferior production data matching. In the present study, we limited the number of
iterations of MS-LM-EnRML so that the computational cost was the same of ES-MDA. However,
it is unclear if we could achieve the same data match quality of ES-MDA with more iterations of
MS-LM-EnRML and still preserve the prior and the ensemble variance.

Acknowledgment
The authors would like to thank CGG and Petrobras for the permission to publish this paper.

References
Aanonsen, S.I., Nævdal, G., Oliver, D.S., Reynolds, A.C., Vallès, B.: The ensemble Kalman filter in reservoir engineering–
a review, SPE J 14(3), 393–412 (2009)
Bruhn, C.H.L.; Gomes, J.A.T.; Luchesse, C.D. & Johann, P.R.S. Campos Basin: Reservoir Characterization and
Management – Historical Overview and Future Challenges, Proceedings of the Offshore Technology Conference,
Houston, Texas (2003)
Buland, A. and Omre, H.: Bayesian linearized AVO inversion. Geophysics, 68(1):185–198 (2003)
Chen, Y. and Oliver, D.S.: Multiscale parameterization with adaptive regularization for improved assimilation of nonlocal
observation, Water resources research, Wiley Online Library, 48 (2012)
Chen, Y. and Oliver, D.S.: Cross-covariances and localization for EnKF in multiphase flow data assimilation. Comput
Geosci 14, 579–601 (2010)
Emerick, A.A. and Reynolds, A.C.: History matching time-lapse seismic data using the ensemble Kalman filter with
multiple data assimilations, Comput Geosci, 16, 639–659 (2012)
Emerick, A.A. and Reynolds, A.C.: Ensemble smoother with multiple data assimilation, Computers & Geosciences, 55,
3–15 (2013)
Emerick, A.A.: Analysis of the Performance of Ensemble-based Assimilation of Production and Seismic Data, Journal
of Petroleum Science and Engineering, 139, 219–239 (2016)
Evensen, G.: The ensemble Kalman filter: theoretical formulation and practical implementation. Ocean Dyn 53(4),
343–367 (2003)
Gentilhomme, T.; Oliver, D. S.; Mannseth, T.; Caumon, G.; Moyen, R. and Doyen, P.: Ensemble-based multi-scale history-
matching using second-generation wavelet transform, Comput Geosci 19, 999–1025 (2015)
Gentilhomme, T.: Adaptive ensemble-based method for sampling nonlinear non-convex problems. Submitted to Comput
Geosci (2017)
Gentilhomme, T.; Lima, A.; Piero, M.; Riffault, D.; Emerick, A.A.. Sebastiao, M and Anyzewski, A.: Real field
petrophysical inversion using adaptive ensemble-based optimization method. Submitted to Comput Geosci (2017)
Oliver, D.S. and Chen, Y.: Recent progress on reservoir history-matching: a review, Comput Geosci 15(1), 185–221 (2011)
Oliver, D.S.: Reynolds, A.C. and Liu, N.: Inverse theory for petroleum reservoir characterization and history matching,
Cambridge University Press (2008)