SPE-187803-MS

Automated Geologically-Consistent History Matching of Facies Distribution

and Reservoir Properties in Inter-Well Space by Adjoint Methods

E. S. Zakirov, O. V. Lyubimova, I. M. Indrupsky, D. P. Anikeev, and E. Yu. Arkhipova, OGRI RAS

Copyright 2017, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Russian Petroleum Technology Conference held in Moscow, Russia, 16-18 October 2017.

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
Comprehensive geological analysis at the stage of 3D geological modeling is necessary to obtain a high-
quality flow model. The next important step is history matching to dynamic well data without distortion of
the initial geological basis of the model. It is necessary to preserve static data on facies indexes, porosity and
other parameters in the well gridcells, adjusting their distributions in the inter-well space, while maintaining
the geological concepts of the object. The purpose of the study presented in this paper was the development
of automated computationally efficient algorithms for solving this problem.
Earlier, based on the computationally efficient adjoint methods, we developed algorithms and their
implementation in the in-house SimMatch® simulator for geologically consistent history matching with
identification of parameters of the variograms and porosity-to-permeability relations for a given distribution
of facies in the inter-well space. In this study, we make transition from the discrete representation of facies
at wells to the continuous values in the inter-well space reflecting the fractional contribution of the facies in
the formation of model cell properties. Considering implicit parametric dependencies on static data at wells
and variogram parameters for properties and facies, a computationally efficient algorithm was developed
for consistent adjustment of the distributions of facies and reservoir properties during history matching of
the 3D model.
The algorithms developed are implemented within the frameworks of the forward and inverse problems.
In the forward problem, a distribution of a continuous facies parameter is constructed, taking into
account well data and the variogram for the facies. In current implementation, the continuous "facie"
value is interpreted as being transitional (weighted) between the adjacent integer values, which is typical
for geological environments with sequential change of facies usually modeled by methods such as the
truncated Gaussian simulation. Further, for each facie, distribution of the reservoir parameters (porosity,
permeability) is independently constructed, taking into account the variogram for porosity and the porosity-
to-permeability relation for this facie. The resulting property value in each cell is determined by weighing
by the portions of the "pure" facies.
Within the framework of the inverse problem, the parameters of anisotropic variograms for the facies
and for the reservoir properties within each facie, as well as the coefficients in the porosity-to-permeability
relation for each facie, serve as the control parameters. For efficient implementation of the automated
2 SPE-187803-MS

gradient procedure for adjustment control parameters, the adjoint problem is solved at each iteration, and the
object function gradient with respect to the control parameters is calculated taking into account the implicit
dependencies of the reservoir properties in the model cells on the variogram parameters for the facies and
reservoir properties. The results of approbation of the approach on a realistic example of the 3D reservoir
model and on a 3D model of a real deposit section are presented.
During the study, effecient algorithms for consistent adjustment of the facies and reservoir properties
distributions in a 3D model were for the first time constructed on the basis of adjoint methods and
implemented in the in-house simulator. The advantage of the approach is the significant reduction in
computational costs (number of simulator runs) for solution of the inverse problem in comparison with
alternative methods of automated history matching, while preserving the consistency of the facies and
reservoir properties distributions with the original principles of the geological model construction.

Introduction
History matching problems, being inverse ones, are most often applied in flow simulation to quantify
unknown or uncertain reservoir properties, such as porosity and permeability [1, 2]. Actual measurements
obtained during field development are employed as known input data. They consist of gas, condensate,
oil, water, liquid well flow rates, measured reservoir and bottomhole pressures, etc. Proper characterization
of reservoir heterogeneity serves as the principal basis for correct dynamic forecasting of reservoir
development using the 3D flow model.
Solution of the history matching problem is often realized in practice by means of "manual" adjustment
of parameter values. This procedure takes vast majority of reservoir engineer's working time within the
project. And its result is ambiguous and strongly depends on qualification, experience and intuition of the
engineer. The problem is aggravated by the fact that flow models are becoming increasingly complex, non-
stationary multiparameter three-dimensional block structures. Therefore, efforts and time spent for data
assimilation within a model, as well as the degree of resulting solution's subjectivity, are not to be considered
more acceptable.
Partially or fully automated history matching of reservoir models has become recently more widespread
[3]. Most often, it is performed using commercial software packages (MEPO®, Enable®, RFD
tNavigator®, etc.), based on the methods of function approximation, experimental design, etc. The purpose
of utilizing such mathematical approaches consists in reducing the number of dynamic simulator runs
needed for analysis of the effect of a limited set of parameters specified by an engineer with certain ranges of
values on the quality of history matching. In this case, the number of required forward problem simulations
is measured by many tens, more often by hundreds and even by thousands of simulator runs.
Automated history matching procedures based on directional search for a minimum of the given
performance criterion are computationally more effecient [2]. In this case, values of reservoir parameters
(porosity and permeability) for many model zones or even for each grid block can serve as control (adjusted)
parameters. But with this approach, important geological information about spatial distribution of reservoir
properties inhrent in the initial geological model is lost. Therefore, petroleum engineers are increasingly
interested in the possibility of adjusting parameters of the 3D model during history matching without
violation of the underlying original geological principles [4, 5]. The appropriate problem statement for
automated history matching and the solution algorithm based on up-to-date methods of optimal control
theory are presented in [6, 7] and in the appendix of this paper, applying the basic geostatistical concept
– the (semi)variogram. In the latter case, history matched 3D models are supposed to be as complex as
possible: multifacial, with anisotropic semivariograms.
Further development of the geologically-consistent approach was aimed at improving the description
and history matching of the spatial facies distribution in the inter-well space. A unified approach is applied
within the framework of both forward and inverse problems.
SPE-187803-MS 3

Within the forward problem, an important issue is related to the improved description of the influence of
facies environment. Under the various facial conditions we mean the genetic features of reservoir formation,
expressed in the specifics of the reservoir properties, such as the mean value and the range of variation for
porosity, the type of functional dependences between porosity and permeability, etc. Usually the facies type
is associated with some class of rocks characterized by specific properties, and is expressed by an integer
index.
At the heart of the approach an idea is postulated that in real sedimentation conditions there is no clear
boundary between two discrete (integer) facies. It is especially so in terms of grid cells with areal sizes in
real 3D models equal in most cases to 50×50 m or 100×100 m, and for large fields – up to 200×200 m
and even more. In addition, within a single model cell, due to its significant dimensions, a change in the
facies types can happen. As a result, averaged properties of the cell can be described by a "transition zone,"
characterized by the change of one facies to another. From mathematical point of view, this features can
be represented by transition from the discrete facies description in cells to the continuous one. That is, a
separate cell of three-dimensional grid model can "average" itself several facies.
Formation of a geological body in various facial conditions predetermines the corresponding reservoir
properties. And to calculate the properties associated with a grid cell, averaging of the property values is
required throughout all the facies affecting this cell.
Thus, our continuous facies parameter can be interpreted as the weighting fraction of particular facies in
a given cell. And it represents a transitional reservoir type between pure facies characterized by integers.
Our approach to continuous facies distribution is conceptually similar to the method of truncated Gaussian
simulation (TGS) [8]. The key idea of the TGS consists in creating realizations of a continuous Gaussian
spatial variable and then to cutting it off with a set of threshold values to simulate integer facies realizations.
An important inherent feature of the truncated Gaussian model is associated with sequential change of facies
in resulting facies models.
We should also note growing interest of the petroleum industry to adjustment of the facies distribution
during history matching, which is manifested, for example, in publications [3–5] issued independently of
the authors’ papers.
The last remark. The developed approach is based on the kriging procedure. Therefore, the proposed
method can be used fruitfully in all approaches based on the two-point geostatistics.

Extension of the geologically-consistent history matching algorithms to

adjustment of parameters in the inter-well space based on production data
Geologically-consistent history matching
The semivariogram is the quantitative measure of uncertainty of spatially distributed data. It reflects the
variance of the difference in parameter values at two distinct points as a function of directional distance
between them [8, 9]. Semivariogram values increase from the minimum one at zero lag (distance), caused
only by the measurement error variance, to the maximum one at a distance at which mutual correlation of
parameter values at two points is no longer observed. Constructed from known static parameter values at
wells (by core and log data), the experimental semivariogram is to be approximated by means of a model
one.
Availability of a sufficient number of reliable theoretical models, as well as taking into account
the variogram anisotropy and facies differentiation, usually allows correct description of the complex
heterogeneous reservoir structure in accordance with the geological concepts. However, values of
semivariogram parameters often cannot be set with sufficient certainty due to the lack of information.
In addition, errors in well data, especially in the case of permeability, can be significant. Therefore,
permeability is often specified as a function of porosity.
4 SPE-187803-MS

All methods of two-point geostatistics are based on the variogram concept for calculating properties in
the inter-well space. These methods are widely used in construction of 3D geological reservoir models.
The inverse problem of geologically-consistent history matching of a 3D model to actual development
data can be set in the following way. Parameters of the anisotropic semivariogram for each facies, as well
as of the porosity-to-permeability relation or the permeability multupliers at wells [6, 7], are to be adjusted.
The corresponding computational procedures are presented in the appendix.
This approach allows us to adjust porosity and permeability distributions in general, in a consistent
manner with the principles of construction of the initial geological model. However, known (measured)
static data at wells are often unevenly distributed over the reservoir volume in a 3D model. Therefore, the
question arises of how to clarify peculiarities of the property distributions far from wells. At the same time,
it is well-known that on the basis of well operation data, e.g., the well water-cut dynamics, some local
features of reservoir properties distribution in the inter-well space can be determined and evaluated. In the
framework of the authors’ approach, this problem can also be solved in a geologically-consistent manner,
for example, using the pilot point method [10, 11]. In this case, the set of the adjusted control parameters
for the inverse problem is extended with the values of reservoir properties at the selected pilot points. And
during generation of the property distribution as a whole with the geostatistical methods all those values
should be taken into account analogously to the original static data at wells.
Applicability of the described approach was confirmed in [12] by studying the effect of changing reservoir
property values in areas distant from wells on the property distribution as a whole. Both deterministic and
stochastic approaches to reservoir properties simulation were considered [8]. The inverse problem solution
algorithm is based on the methods of optimal control theory (adjoint method) [2].

Algorithm for solving the inverse problem

The inverse problem solution in an optimization statement consists in minimizing of the so-called
performance or objective function (criterion). It reflects the degree of actual data assimilation in the model,
that is, how close the simulation results with the model correspond to the measured values at wells [2]:

(1)

where j is the serial number of a certain measurement time point, N is the total number of measurement
time points; and correspond to the vectors of measured and calculated values of dynamic data (bottom
hole pressures, reservoir pressures, component rates, rate ratios –water cut (WCT), gas-oil ratio (GOR),
current fluid contact levels, logged phase saturations, and other available dynamic data at different wells/
intervals/grid blocks) taken at jth time point; superscript T is the transposition sign. Symmetric (often
diagonal) weighting matrix Ω serves for normalization, units matching and confidence control of various
measurements.
The vector of adjusted (control) model parameters includes anisotropic semivariogram parameters
(azimuth of the major direction, ranges along the three axes, the nugget effect and the plateau) for porosity,
as well as parameters of porosity-to-permeability functional relation for each facies, or the semivariogram
parameters and well multipliers for permeability in the case of its independent distribution [6, 7]. In the case
of parameter adjustment in the inter-well space, porosity and permeability values at pilot points are also
included in the vector. For the choice of pilot points, it is expedient to take into account the performance
criterion sensitivity, that is, the map of the objective function (1) derivatives with respect to porosity and
permeability in the grid cells. An alternative approach assumes that cell centers of the first-level grid are
used as pilot points in the case of constructing geostatistical parameter distributions using the nested grid
method [12].
SPE-187803-MS 5

To minimize the objective function, iterative gradient methods are employed providing a superlinear
convergence rate. Namely, the conjugate gradient method and quasi-Newton methods [2] were
implemented. An improved estimate of the control vector on (v + 1) -th iteration is calculated from the
previous one as follows:
(2)
where is the search direction towards the objective function minimum, and β(v) is the step size along the
search direction. Only objective function gradient ∇J with respect to is required for to be computed.
At each iteration of the inverse problem solution, only one forward problem solution (simulator run) is
performed. Then the value of J is calculated. Objective function gradient ∇J and the step size are computed
using two auxiliary problems (the adjoint problem and the problem for state variable variations) according
to the methodology of optimal control (adjoint) theory. The total computational cost per iteration does not
exceed that of two forward problems runs, irrespective of the number of control parameters [2].
As soon as the control parameter values on the new iteration are calculated, porosity and permeability
distributions are updated with known deterministic or stochastic two-point geostatistical methods based on
the kriging procedure.

Calculation of parameter distributions and their derivatives

Kriging provides the most probable value (and its variance) of the property at a given point as the best
unbiased linear estimate from known data at other points [8]:

(3)

Here λi are the kriging weights, x0 is the point where the property value is estimated, and z(x1), z(x2),
…, z(xN) are the known values at other points. Weights are chosen subject to the following conditions: the
estimate should be unbiased, and its variance should be minimal. These conditions lead to the well-known
set of N + 1 linear equations, namely, to the so-called ordinary kriging system:

(4)

and the variance of the obtained estimate (3) can be expressed in terms of the semivariogram:

(5)

where γij denotes the semivariogram value between i and j points of the measured data, γi0 corresponds to
the semivariogram value between ith data point and the point at which the property is being estimated, γ00
is the measure of the nugget effect [8].
Thus, kriging makes it possible to obtain a single most probable realization of the property distribution,
as well as the variance distribution reflecting the uncertainty of the obtained estimate at each point.
To assess the uncertainty of reservoir property distributions and its impact on the results of reservoir
development simulation, methods of stochastic modeling are usually used. They make it possible to get
many realizations of the property distribution for a given data set and chosen model parameters. One of the
most common methods is the sequential Gaussian simulation (SGS). The principle of sequential simulation
6 SPE-187803-MS

consists in the following. At each sequential point, the value of the property is determined by drawing a
Gaussian random variable with mathematical expectation and variance obtained from the kriging results,
taking into account the values already estimated at the preceding points. To determine the ordering of points
(grid cells) at which values are to be estimated, a random sequence is constructed for each realization. In the
limit, averaged SGS property distribution over many realizations coincides with the kriging distribution.
To apply a gradient method (conjugate gradient method or any quasi-Newton method) for minimizing the
function (1), it is necessary to be able to calculate partial derivatives of the objective function with respect to
control parameters. The derivatives to porosity and permeability values in each model cell are determined
by optimal control theory methods, as described in [2] and in the appendix of this paper. For conversion
from the property value in a cell to the semivariogram parameters (which was the basis for estimation of
the property value), we apply the following relation

(6)

derived from (3). Here is the control sub-vector corresponding to the semivariogram parameters.
Computing the vector from (3) does not cause any difficulties. For calculation of , the set of
equations (4) is differentiated as an implicit function of multiple variables:
(7)

Expanding derivatives in (7), we find that to determine each ith column of the Jacobi matrix , we need
to solve the following set of linear equations [6]:

(8)

Thus, having solved the system (8) for each unknown value of the considered parameter (porosity or
permeability), it is possible to obtain the distributions of derivatives with respect to the semivariogram
parameters.
In a similar way, derivative distributions with respect to the coefficients of the functional relation between
porosity and permeability are obtained, as well as to the permeability multipliers at wells [6, 7]. Exactly the
same procedure is employed for property values at pilot points.
The described algorithms for calculating derivatives are implemented as a plug-in for the Schlumberger
Petrel® software package using the "Ocean" framework. In addition to calculating the derivative
distributions, the module provides calculation of the property distributions by kriging and sequential
Gaussian simulation. The module is incorporated in the software environment, like any other standard
Petrel® tool, and has the same set of adjusted parameters. The obtained property distributions can be
visualized and exported to standard formats for further use in common simulators for solving forward
problems, and also in the SimMatch® software of OGRI RAS for solution of inverse problems on the basis
of optimal control theory methods [2]. Such integration extends the capabilities of reservoir engineers and
provides additional options for analyzing data within the framework of the familiar software tool.
SPE-187803-MS 7

The new approach to facies modeling

Let's suppose that there are F facies. For each facie, within its extents, semivariogram for porosity (and
also, if necessary, for permeability) is determined according to the actual data, and also (if the variogram
for permeability is not specified) the porosity-to-permeability functional relation should be specified.
To estimate the porosity value in rth grid cell, it is necessary to solve F sets of linear equations
independently of each other with different sets of semivariogram parameters (in the general case, each facies
possesses its own semivariogram):

(9)

where is the semivariogram value as a function of distance

between the points with original data calculated for the set of semivariogram parameters . The latter
characterizes distribution of porosity of the ith facies in the reservoir. denote indices of actual
property measurements for a given facie; denotes the solution of the system of equations and represents
a vector of weighting coefficients for estimating the property at the point r using the original data set for the
ith facie. Nsd stands for the quantity of input data for the ith facie
i
, and the following equality holds:

where Nsd is the total number of measurements for all facies.

As a result, there are F porosity values in each grid block r, and each of them represents a weighted
average of the static data for a certain facie:

A well-known kriging feature is that kriging weights (the solution of the system (9) for each facie) do not
depend on the original set of static data values. Instead, they depend only on spatial locations of the static
data. In the case of only one unified set of data for all the facies including the ith facie, to calculate kriging
estimate of porosity at any point in space it is enough to invert the matrix Γi in the left-hand side of linear
equations (9) just once for any grid point, and then multiply it by the vector composed of semivariogram
values computed for the distances between the current point (center of the grid block r) and data points:

The presented argumentation was applied during numerical implementation of the method for models
with several facies. The novelty of the proposed approach consists in the introduction of additional system
of equations responsible for kriging of facies. The facies index values along wellbores interpreted from the
log data are treated as measured values. And for the specified set of measurements, a semivariogram for the
distribution of facies with a set of parameters is determined.
To estimate the effect of original static initial data for facies indexes along wellbores on the value of the
continuous facie "index," a set of linear kriging equations is solved for a certain grid point r:
8 SPE-187803-MS

(10)

Here is the semivariogram value computed for two points

of the static data depending on the distance between them and the set of semivariogram parameters
responsible for facies distribution, is the vector of weights for estimation of continuous facie value for
the grid block r. The continuous facie value is then calculated as:

The resulting facie value fr may be interpreted as a weighted combination of the two facies with
neighbouring serial numbers:

where wsi and wsi + 1 represent the weights and are uniquely determined from the obtained value of fr:
(11)
The combined ‘pure’ facies in the grid block are taken as:

Here ⌊·⌋ is the integer part of the number (the largest integer less than or equal to the number).
This approach assumes a certain sequence for changing facies, i.e. transition from the facie i to other
facies is possible only through the facies i+1 and i-1. This assumption also underlies the truncated Gaussian
simulation (TGS) method and it is usually valid, in particular, for shelf sedimentation conditions. To describe
other conditions for changing facies, alternative interpretations of the continuous facial variable in the model
cells are possible.
As the number of facies by well data is F, then for each block of the model F + 1 sets of linear equations
are solved – F sets for the porosity weights within each facie and one set (equations (10)) for the general
distribution of facies in the inter-well space.
The final value of porosity in the grid block r is the weighted average of the two values of porosity:
(12)
In the presented problem statement considering improved facies distribution, the resulting porosity
distribution during assimilation of production data (dynamic well data) can be determined iteratively. At
each iteration of the inverse problem solution, all parameters of all semivariograms are updated by gradient
methods. Namely, variogram parameters for porosity distribution in the inter-well space within each ith
facie, and the variogram parameters for distribution of continuous-type facies, eventually affecting the
weighting factors wsi and wsi + 1 for the two facies in each grid block. All of them are included in the vector
of control parameters , as stated in formula (1).
The calculation of the corresponding derivatives is carried out through the functional relationship:
SPE-187803-MS 9

The partial derivative with respect to lth parameter of the ith facie semivariogram can therefore be
computed using the chain rule for differentiation as

(13)

From formula (11), the validity of the following transformations directly follows:

Since and can be obtained regardless of the facies distribution, these values are treated as
constants with respect to the variogram parameters:

Then in the chain of derivatives of the right hand side of the formula (13) the is to be computed.
From the formula (11), we obtain:

The derivative does not make sense if facie has an integer value in the grid block. The algorithm regulates
and eliminates the appearance of such values, making minor changes in the weights.
Calculation of the partial derivatives is similar to the calculation of derivatives with respect to
the semivariogram parameters for porosity (see the appendix). That is, to adjust the parameters of the
semivariogram used for facies distribution, it is necessary to solve a set of linear equations with the matrix
Γi from (10) for each parameter τl:

Fig. 1 shows an example of continuous facies distribution by kriging. Here the functional dependence of
facies on spatial coordinate belongs to the class of continuous functions. The standard Gulfaks field model
is taken as the geological basis, usually used for training in the Petrel® software package.
10 SPE-187803-MS

Figure 1—Facies distribution obtained by kriging

In the example considered, 4 fluvial facies (background floodplain, channel sand, levee sand, crevasse
splay sand) have been determined by geologists from well data. Original porosity data for each facie
were taken from well grid blocks associated with the facie, and porosity values were distributed in the
inter-well space for the whole model grid. Thus, distinctive features of porosity distribution were revealed
characterizing reservoir properties of each facies (see Figs. 2–5). Then the principle of facies transition was
used, which is typical for sedimentary environments with a successive change of facies. The real number
for the facie value in a grid block was computed through the kriging procedure. This value is interpreted as
a superposition of the two pure facies with the closest integer numbers, and their weights are equal to the
distance from the real number to the integers. The resulting porosity of the grid block is also the weighted
average of porosity values obtained in the independent calculations for the two facies with the same weights.
In probabilistic terms, this conjecture can also be interpreted as the mean porosity for the given probabilities
(weights) of the facies in the grid block. The resulting distribution of porosity is shown in Fig. 6.

Figure 2—Porosity distribution for the fluvial facie 1 (background floodplain)

SPE-187803-MS 11

Figure 3—Porosity distribution for the fluvial facie 2 (channel sand)

Figure 4—Porosity distribution for the fluvial facie 3 (levee sand)

Figure 5—Porosity distribution for the fluvial facie 4 (crevasse splay sand)
12 SPE-187803-MS

Figure 6—Resulting porosity distribution, obtained by combining porosity values of different facies in each cell

For a gradient method of objective function minimization, it is necessary to calculate the distributions
of porosity derivatives with respect to variogram parameters used for facies distribution in the inter-
well space. In Fig. 7, as an example, the corresponding distribution is presented for a single parameter
(the semivariogram range in the major direction) overestimated if compared to the "true" value. Relative
influence of the obtained derivatives (sensitivity coefficients) during the inverse problem solution is
characterized by the distribution of the relative derivatives (normalized by the porosity values in grid cells),
as shown in Fig. 8.

Figure 7—Map of porosity derivatives with respect to the major range of

the semivariogram used for facies distribution (range value overestimated)
SPE-187803-MS 13

Figure 8—Map of relative porosity derivatives with respect to the major range
of the semivariogram used for facies distribution (range value overestimated)

Mathware and software developed in 2014-2016 by various members of our team is now being integrated
into a single software package. This package is now subjected to a comprehensive testing procedure on
synthetic tests. In 2017, a full-field 3D model testing on a part of a unique gas field was initiated.

Application of the proposed approach to simulation of facies and property

distributions for a Cenomanian gas reservoir model
The example of the Gullfaks model shows the implementation specifics of the proposed approach and
the developed algorithms in accounting for and correcting the distribution of facies within the solution
of the inverse problem. Nevertheless, due to some artificiality, this example does not provide thorough
demonstration of the approach advantages for proper consideration of the geological features of the object.
Therefore, the approbation of the developed algorithms is also carried out on the model of a section (a
separate dome) of the Cenomanian massive gas deposit with bottom water at a unique field of Western
Siberia.
The reservoir is composed of sandy-silty rocks with shaly inclusions, often lenticular. Reservoir
development has long been at the stage of production decline, and about 80% of the initial gas-in-
place have already been produced. Though the wells productivity and mean reservoir properties are
quite high, their areal and vertical distributions are very heterogeneous. The existing 3D flow model
keeps low level of detailization and poorly reflects the actual specifics of reservoir property distributions.
Construction of the detailed 3D model and its geologically-consistent history matching are of great interest
for localization of remaining gas volumes and justification of efficient technological solutions for further
reservoir development.
The initial information for model generation included available structural surfaces and a 3D grid
constructed on their basis, as well as well log interpretation data (WLID) obtained with the TAVS method
[13] ("Technology of automated reconstruction of rock properties in a terrigenous column by well log
data – ESKS-TAVS "). Along with traditional parameters of porosity and gas/water saturation, estimates
of volumetric content of sand, silt and clay fractions, as well as the bound water content, are evaluated at
each logged interval in all wells, and on their basis permeability and prevailing rock lithotype (sandstone–
siltstone–clay–tight rock) are also determined.
For approbation of the developed approach, the lithofacial parameter LITO_cont was considered as the
continuous facies parameter, constructed on the basis of lithotypes identified by WLID. Integer LITO_cont
14 SPE-187803-MS

values at wells correspond to the WLID data (1 – sandstone, 2 – siltstone, 3 – shales, including clay and tight
rocks intervals joined due to small amount of data). On the basis of available data, an anisotropic variogram
model was identified for the distribution of LITO_cont as a continuous parameter in the inter-well space
with values in the range from 1 to 3 (Fig. 9).

Figure 9—Distribution of the continuous lithofacies parameter in the model cross-section

Then, three anisotropic variogram models for porosity were separately constructed. Each of them
corresponded to the data of those well intervals where LITO_cont took values of 1, 2, or 3, respectively.
Using the corresponding sets of static data and variogram models, three distributions of porosity were
generated over all model cells (Figs. 10–12) responding to the "pure" lithotypes of sandstones, siltstones
and clay/tight rocks, respectively.

Figure 10—Distribution of porosity for the "sandstone" lithotype in the model cross-section

Figure 11—Distribution of porosity for the "siltstone" lithotype in the model cross-section

Figure 12—Distribution of porosity for the "clay/tight rock" lithotype in the model cross-section

At the last stage, using the above-described "weighing" rule, the resulting porosity values in the model
cells were obtained (Fig. 13). Thus, the resulting porosity field takes into account the contribution of the
lithofacies (lithotypes) in the average reservoir properties of each model cell.
SPE-187803-MS 15

Figure 13—Resulting porosity distribution in the model cross-section

Similar steps were implemented to construct the permeability distribution. Fig. 14–16 show distributions
of permeability logarithm within one of the model cross-sections for "pure" lithotypes of sandstones,
siltstones and clay/tight rocks, respectively. The resulting permeability distribution is shown in Fig. 17 (with
logarithmic scale).

Figure 14—Distribution of permeability logarithm for the "sandstone" lithotype in the model cross-section

Figure 15—Distribution of permeability logarithm for the "siltstone" lithotype in the model cross-section

Figure 16—Distribution of permeability logarithm for the "clay/tight rocks" lithotype in the model cross-section

Figure 17—Resulting permeability distribution (logarithmic scale)

16 SPE-187803-MS

The presented figures reflect the advantages of the proposed approach to the construction of facies and
reservoir/flow parameters distributions. Figs. 10–12 and 14–16 demonstrate the "smearing" (averaging) of
the heterogeneities typical for the kriging method in the case of property distribution according to a single
variogram for a particular facie. However, the very distribution of lithofacies, due to discrete values at wells,
shows evident local heterogeneities, including the ones of lenticular type (Fig. 9). Through the realization of
the "weighing," they are inherited by the final distributions of porosity and permeability, while preserving
the continuity of the spatial distribution of the property values and the absence of jumps at the boundaries
of facies changes (Figs. 14, 17).
The generated variogram models for distributions of the continuous lithofacies parameter and the
reservoir properties serve as initial ones. They are further adjusted during the inverse problem solution
when history matching the model to the production data, with the appropriate recalculation of porosity and
permeability distributions subject to the described algorithms.

Conclusions
Previous studies [6, 7] confirmed the reality of automated geologically-consistent history matching of
reservoir models to production data. For synthetic examples, true values of geological model parameters
were restored with high degree of accuracy. In the subsequent studies, the authors have:

• confirmed the possibility of local features adjustment for reservoir property distributions in the
inter-well space based on pilot points by estimation of influence of the property value at such point
on the property distribution in the surrounding area;
• developed an approach to improve the description of facial specifics of oil and gas fields in 3D
models. For this, a transition is made from a discrete to a continuous facies parameter. The resulting
reservoir rock property distributions take into account fractional contribution of the pure facies to
calculation of reservoir properties in a cell.
Thus, the proposed approach provides a generalization over the previously presented method for inverse
problem solution to the case of adjustment of not only rock properties, but also the spatial distribution of
facies in the inter-well space. And it preserves both well data and general principles of 3D geological model
generation with geostatistical methods. At the same time, the distributions of not only reservoir properties,
but also facies are adjusted according to actual production data.
Thus, the developed approach provides a more realistic account for the facial features of sedimentation
environments. And it also makes possible implementation of realistic history matching of multifacial
deposits.

Acknowledgments
The authors appreciate the support for their studies provided over the last years as a part of the Program of
the Presidium of the RAS "Fundamental Problems of Mathematical Modeling". We are also thankful to the
«Gazprom Dobycha Nadym» LLC for the given permission to use the Cenomanian reservoir section model.

Reeferences
1. Oliver D.S., Reynolds A.C., Liu N. Inverse Theory for Petroleum Reservoir Characterization and
History Matching. – New York: Cambridge University Press, 2008.
2. Zakirov E.S. 3D multiphase problems of forecasting, history matching and optimization within oil
and gas fields development. – Moscow, 2001, 302 p. (in Russian)
SPE-187803-MS 17

3. Tsepelev V.P., Pislegin M.N., Timchuk A.S. Automatic history matching of reservoir simulation
models with using of permeability fields multiplier tuning. // Oil Industry. 2011. #10. Pp. 97–99
(in Russian)
4. Bakhitov R.R., Zimin S.V. Some aspects of creation and history matching of geological model
of Nkh3-4 formation, Vankor field // Proceedings of the III conference in the name of N.N.
Lisovsky "Current state and further development of basic principles of oilfield reservoir
engineering," Almetyevsk, September 25-28, 2012. – Tatneft, 2013. Pp. 191–207 (in Russian).
5. Potekhin D.V., Putilov I.S., Galkin V.I. Improve the reliability of geological models of oil fields
on basis of optimized technological settings multivariate 3D modeling. // Oil Industry. 2014. #6.
Pp. 16–19 (in Russian).
6. Zakirov E.S., Indrupskiy I.M., Lubimova O.V., Shiriaev I.M. Geostatistically-Consistent History
Matching // Proc. Of the 14th European Conference on the Mathematics of Oil Recovery.
Catania, Sicily, Italy, 8-11 September 2014.
7. Zakirov ES, Zakirov SN, Indrupsky IM, Anikeev DP, Lyubimova OV, Shiryaev IM, Baganova
M.N. Optimal control of field development in a closed loop // SPE Russian Petroleum
Technology Conference, 26-28 October 2015, Moscow, Russia. SPE-176642-EN.
8. Deutsch C.V. Geostatistical reservoir modeling. – New York: Oxford University Press, 2002, 376
p.
9. Dubrul O. Geostatistics in petroleum geology. – AAPG, 1998, 210 p.
10. de Marsily G., Lavedan G., Boucher M., Fasanino G. Interpretation of interference tests in a well-
field using geostatistical techniques. // In: ReidelD. (ed.) Geostatistics for Natural Resources
Characterization, Part 2, pp. 831–849.
11. Bissell R. Calculating optimal parameters for history matching // Proc. Of the 4th European
Conference on the Mathematics of Oil Recovery, Roros, Norway, 7-10 June 1994.
12. Zakirov E.S., Shiryaev I.M., Indrupsky I.M., Lyubimova O.V. About possibility of geologically-
consistent adjustment of hydrodynamic model parameters in the inter-well space from well
production data // Geology, geophysics and development of oil and gas fields. 2016 #12. Pp.
33–40 (in Russian)
14. Afanasyev S.V. Technology of complex reinterpretation of well logging data while creating a
three-dimensional geological model of a mature field. // Oil Industry. 2005. #2. Pp. 12–17.
15. Abasov M.T., Zakirov I.S., Palatnik B.M. Identification of relative phase permeability functions
during two-phase flow // Doklady Earth Sciences, Reports of the Academy of Sciences of the
USSR, Vol. 312. #4. 1990.
16. Abasov M.T., Zakirov S.N., Palatnik B.M. History matching of geological-mathematical model
of gas reservoir under water drive // Doklady Earth Sciences, Reports of the Academy of Sciences
of the USSR. Vol. 38. #2. 1989.
17. Zakirov S.N., Vasilyev V.I., Gutnikov A.I., Korshunova L.G., Kolbikov S.V.: Forecasting and
regulation of gas fields development. – Moscow: Nedra, 1987, 295 p. (in Russian)
18. Zakirov S.N., Kolbikov S.V., Palatnik B.M. Complex adaptable geological and facility models
for gas fields development // Proc. of Gubkin oil and gas institute. Vol. 214. – Moscow, 1989 (in
Russian)
19. Marchuk G.I., Agoshkov V.I., Shutyayev V.P.: Conjugate equations and perturbation methods in
nonlinear problems of mathematical physics. – Moscow: Nauka, 1993 (in Russian)
20. Palatnik B.M., Aanonsen S. I., Zakirov I.S., Zakirov E.S. New technique to improve the
efficiency of history matching of full-field models // Proc. 4th European Conference on the
Mathematics of Oil Recovery. Roros, Norway, 7-10 June, 1994.
18 SPE-187803-MS

21. Palatnik B.M., Zakirov I.S., Haugen S.A., Van Roosmalen J.J. New approach to multiphase
history matching // Proc. 7th European IOR Conference, Moscow, Oct. 27-29, 1993.
22. Gill F., Murray W., Wright M. Practical optimization. – Emerald Group Publ. Ltd., 1982. – 418
p.
23. Yang P.H., Watson A.T. Automatic history matching with variable-metric methods // SPERE,
August 1988, pp. 995–1001. Paper SPE 16977.
SPE-187803-MS 19

Appendix
Inverse problems of reservoir parameters identification
In the inverse problems, it is assumed that during several years of field development, actual data on dynamics
of well flow rates for oil, gas, water, bottomhole and reservoir pressures at wells, and phase saturations
along well boreholes (completed formation intervals) have been obtained. Obviously, those data itself bring
information about parameters of the deposit being developed. Therefore, a natural intention is, on the basis
of this information, to adjust reservoir and flow properties within the entire volume of the reservoir under
consideration.
The problems of this type are called the inverse (coefficient) problems (also called the history matching
problems), and their aim is to determine the properties of the reservoir.
Inverse problems are formulated in an optimization statement. As a performance functional, a weighted
sum of square differences between calculated and actual values of all measured parameters for all
production, injection and observation wells for the considered period of field development history is usually
utilized. In more advanced statements, objective function may also engage regularizing terms, including
those based on statistical principles (for example, the Bayes method).
Algorithms for solving various history matching problems with identification of different reservoir
parameters are presented, for example, in [14–21]. The application of a gradient procedure is common to
all those algorithms. The most efficient approach to calculation of the required derivatives of the objective
function with respect to control (adjusted) parameters are the methods of optimal control theory, also called
the adjoint methods. At each iteration of the inverse problem solution, the following problems are solved:

• forward 3D multiphase boundary problem with given actual well operation modes, for example,
oil production rates;
• adjoint 3D "multiphase" boundary problem for auxiliary (adjoint) functions.

In addition, a special procedure involving solution of an additional boundary value problem is used to
compute the step size along the search direction. All this improvements significantly accelerate the directed
search for a minimum of the objective function, regardless of the number of parameters to be determined,
number of wells, volume of actual information about wells exploitation, and duration of field development
history.
Let's give a fairly general description of the algorithm for automated model history matching using
effective methods of optimal control theory to calculate objective function gradient.
We write down the objective function in the following form [2]:

(A.1)

Here j is the serial number of a certain measurement time point, N is the total number of measurement time
points; and correspond to the vectors of measured and calculated values of dynamic data (bottomhole
pressures, reservoir pressures, component rates, rate ratios – water cut (WCT), gas-oil ratio (GOR), etc.,
logged phase saturations, and other available dynamic data at different wells/intervals/grid blocks) taken
at j-th time point; superscript T is the transposition sign. Symmetric (often diagonal) weighting matrix Ω
serves for normalization and confidence control of various measurements. For Bayesian formulation, Ω
is the inverse covariance matrix, and an additional a-priori term is added to the objective function (A.1),
which is trivial to handle throughout the algorithm. Vector stands for the values of state variables of the
forward problem at jth time point in each grid block. For example, for a 3-phase saturated grid block of a
typical black oil model the kth subvector of consists of oil pressure, water saturation and gas saturation
20 SPE-187803-MS

in the kth grid block of the model. The bottomhole pressures to be calculated for wells controlled by rates
are also included in [2].
In the formula (A.1), is the vector of control parameters to be determined through the solution of
the inverse problem and typically consists of porosities and permeabilities of individual grid blocks of the
model, or porosity and permeability multipliers for given reservoir zones. Any other control parameters,
like relative permeability exponents, local reservoir geometry or aquifer properties (both numerical and
analytical, such as Fetkovich) can also be included in .
Forward problem consists in solving a system of partial differential equations describing simultaneous oil,
gas and water flow in porous medium, according to, for example, the black oil model, with corresponding
initial and boundary conditions [2]. After discretization, the system of nonlinear equations of the forward
problem at the jth timestep can be written in the form of
(A.2)
To solve the system (A.2), the Newton method is used, and at each Newtonian iteration ω the following
set of linear equations is to be solved
(A.3)

where stands for the Jacobian matrix of with respect to , and is the increment of state variables
at the current iteration ω.
Note that in general the distance between measurement time points may include several time steps, but
we use the same time index j in (A.1) and (A.2) just for clarity.
The inverse problem is a problem of minimization of the objective function (1) on an admissible region
. That is performed by means of an iterative gradient-based optimization algorithm. Quasi-Newton
procedures, like BFGS or SSVM [22, 23], proved to be the best choice considering convergence speed and
reliability. At each inverse problem iteration v, the vector of control parameters is updated in the following
manner:
(A.4)
where is the quasi-Newton search direction, and β(v) is the step size. Only objective function gradient ∇J
with respect to is required for to be computed. Effective calculation procedure is constructed using
the discrete form of the Pontryagin maximum principle. The gradient ∇J is obtained on the basis of adjoint
problem solution (A.5) according to the expression (A.6):

(A.5)

(A.6)

where Δtj is the length of the jth time step, and is the adjoint vector. The adjoint "initial" condition
holds. The adjoint system has the same dimension as the set of equations of the forward problem, but it
is solved in the opposite direction in time. The adjoint problem (5) is linear, and its computational cost
is several times less than that of the forward problem, whatever the number of control parameters is. As
soon as (A.5) is solved and ∇J is computed, the search direction is readily obtained from quasi-Newton
update formulas [22, 23].
The usual way to calculate optimal step size β(v) requires evaluation of the objective function J at a
number of points, i.e. a number of simulator runs is required, which is very expensive computationally. A
SPE-187803-MS 21

special approximate procedure proved to be highly efficient, which includes the solution of the problem for
variations of the state variables
(A.7)
with the initial conditions . After solving the problem (A.7), the approximate optimal step size is
computed based on the linearization of the dependence and the minimum point of the quadratic
function (A.1) as follows:

(A.8)

taking into account the dependency of variation on . Again, the computational cost of the problem
(7) is several times less than that of the single forward problem run.
The algorithm just described is implemented in the in-house black oil simulator software SimMatch®
with extra options of automated history matching, optimal well rates allocation and upscaling [2].

The geologically consistent inverse problem formulation and solution

algorithm
Let's get down to the problem of geologically consistent model history matching. We modify the previously
described inverse problem for porosities and permeabilities of individual grid blocks by rewriting the
objective function (A.1) in the following form:

(A.9)

where are the anisotropic variogram parameters (major range, minor range, vertical range, nugget and
azimuth angle) and are the parameters of porosity-to-permeability relations (usually zero- and first-order
term coefficients of the semilog linear dependence) for each of the facies.
In contrast to the formulation (A.1), the new problem statement is treated as follows. The distributions
of porosity and permeability in model cells forming vector are not completely independent. Those
distributions are constructed using specified values of porosity at wells (according to well log interpretation
data – WLID) and geostatistical, in particular, kriging procedures for their distribution to the inter-well
space [8, 9]. Permeability of each grid cell is calculated as a function of porosity. Thus, the inverse problem
for and is to be solved.
The modifications to the algorithm described in the previous section are straightforward. The only
difficulty to overcome is the calculation of derivatives with respect to and instead of in (A.6) and (A.7).
First we note that differentiation of any vector function according to the chain rule gives
(A.10)
after which can be readily computed, since the porosity-to-permeability relation is expressed by a given
explicit function. Hence, the calculation of the Jacobian matrix is of the main interest. Moreover, only
matrix rows corresponding to porosities of grid blocks obtained by kriging are to be discussed. Calculation
of the other rows is trivial, as permeabilities are directly related to porosities through . Thus, to simplify
formulas, let us further treat as a vector of grid block porosities within certain facie and adopt the common
assumption that the probability distribution for is Gaussian (otherwise it first has to be normalized).
22 SPE-187803-MS

Following the well-known kriging procedure, the porosity of a certain grid block r is the
number of grid blocks in the model) is computed as

(A.11)

where i is the sequence number of the static porosity value usi obtained from logs, Nsd is the total number
of static well porosity data; λrl are the kriging weights. Since kriging is an exact method, there's no need to
exclude fixed porosity values on wells from . Moreover, the condition of the best linear unbiased estimate
leads to the following set of linear equations for λrl:

(A.12)

Here is the semivariogram value computed for two points of the static
data with sequence numbers i1 and is the semivariogram
value computed for the ith static data point and the center of the grid block r; μ is the Lagrange

multiplier for the non-bias condition .

As usi in (A.11) are fixed, the derivatives of ur with respect to are linearly related through (11) to the
Jacobian matrix . In its turn, the implicit relation between and is given by (12) and the variogram model
used. Hence, the implicit derivatives with respect to each variogram control parameter vl are computed from
the following set of linear equations:

(A.13)

where is the functional expression of the ith line of (A.12) (linear in and nonlinear in ).
The set of linear equations (A.13) is to be solved in every grid block for every variogram control
parameter. An efficient implementation is following that of a usual kriging procedure. Since the size of the
static data set is moderate, the matrix on the left side of (A.12) is inverted by a direct method once per each
inverse problem iteration. After that (A.12) and (A.13) are solved as many times as required by means of
only matrix-vector multiplications.

Alternative formulation for permeabilities

The described derivation (A.9)–(A.13) is applicable when porosity distribution is obtained by kriging and
permeability is computed by means of porosity-to-permeability relation. Another option to be considered
SPE-187803-MS 23

is when static permeability values at wells are taken from logs, so that permeability is also kriged. In that
case, the overall procedure of inverse problem solution is quite similar to what has been described, except
for several points to be mentioned.
First of all, the probability distribution for permeability is usually close to log normal, which means that
should contain log values for permeabilities. Also variograms for porosity and log permeability can be
different, so includes the parameters for both the semivariograms and for every facie.
Further on, one should take into account that log data for permeability are usually far from actual reservoir
values of effective permeability. A possible solution is to introduce well permeability multipliers, so that
the multiplier is applied to every static permeability value at a certain well. In that case, these multipliers
are to be included in the set of control parameters of the inverse problem. Thus, the formulation (A.9) still
holds, with designating the vector of well permeability multipliers.
Now it is left to describe the procedure to compute derivatives of grid block permeabilities with respect
to well permeability multipliers. The value of permeability in the grid block r is computed by the kriging
procedure (A.11) (with u now being log permeability). Permeability values at wells are not fixed but
calculated as the products of their initial values and the values of well permeability multipliers at current
inverse problem iteration, so that:
(A.14)
where superscripts denote inverse problem iteration numbers, is the permeability multiplier for well w
at current iteration v, and permeability value ksi corresponds to the well w. After substitution of (A.14) into
(A.9), one ends up with the derivative of kr with respect to aw:

(A.15)

with βiw = 1 if permeability value ksi corresponds to the well w, and 0 otherwise.