SPE 84464

History matching Using Time-lapse Seismic (HUTS)

O. Gosselin, SPE, TOTAL E&P UK plc, S. I. Aanonsen1, SPE, I. Aavatsmarka, SPE, NORSK HYDRO, A. Cominelli, ENI,
R. Gonard, M. Kolasinski, TOTAL E&P UK plc, F. Ferdinandi, L. Kovacic, ENI, K. Neylon, SCHLUMBERGER

Copyright 2003, Society of Petroleum Engineers Inc.

Introduction
This paper was prepared for presentation at the SPE Annual Technical Conference and The use of time-lapse, or 4D seismic data in reservoir
Exhibition held in Denver, Colorado, U.S.A., 5 – 8 October 2003.
management, characterisation and monitoring is steadily
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, as
increasing as the interpretations get more reliable. Inverted
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to seismic data has proven to be very valuable for locating
correction by the author(s). The material, as presented, does not necessarily reflect any
position of the Society of Petroleum Engineers, its officers, or members. Papers presented at remaining oil and plan infill drilling.1,2 Better and more
SPE meetings are subject to publication review by Editorial Committees of the Society of
Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper
reliable 4D data also triggers the need for a tool, which can
for commercial purposes without the written consent of the Society of Petroleum Engineers is help the petroleum engineers to condition the reservoir models
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 to this data.
acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O.
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.
Several authors have considered the problem of using 4D data
in the process of history matching reservoir simulation
models, but very few have described a complete software
Abstract system which can handle the integration of both production
This paper describes the results of a two-year EC- data and seismic data in a computer aided history matching
sponsored project which uses new information provided by loop. Also most papers either use synthetic data3-8 or include
repeated seismic acquisitions (4D seismic data) jointly with the data in a qualitative way9-14. Landa and Horne15 presented
production data in an extended, efficient and consistent history a sensitivity study looking at the relative influence of the
matching process. This process involves a simultaneous various types of data in the reservoir characterisation process.
minimisation of the mismatch between all types of measured Huang et al.16-17 have presented a quantitative approach where
and simulated data. A gradient-based technique has been the misfit between 4D real and synthetic amplitude from
developed and tested both in a prototype and in commercial reservoir simulations is minimised using a stochastic search
computer-aided history matching software. We show results procedure. The misfit function also includes production data,
on real cases, located in the North Sea and the Adriatic Sea, but there is a lack of documentation about the exact definitions
and discuss key issues of such seismic history matching. and algorithms used. The procedure was applied to a Gulf of
Most applications of time-lapse seismic to date have been Mexico dry gas field and improved the reliability of model
qualitative or semi-quantitative. We propose a quantitative predictions. Waggoner et al.18 have used a similar approach to
workflow. The seismic contribution in the objective function a simple gas condensate reservoir. Here the similarity between
is defined in terms of elastic parameter variations within the acoustic impedance variations from 4D data and impedance
reservoir and the data have been properly scaled using an calculated by a numerical simulator was maximised using a
estimate of seismic uncertainty (covariance matrix). The greedy global optimisation algorithm. These approaches,
“observed” values are obtained by inversion of the seismic however, apply a global optimisation routine, which requires
signal. For the “modelled” values, the flow simulator is hundreds and even thousands of simulation runs to obtain
coupled with a petro-elastic model to convert simulated fluid a match.
and static rock properties into simulated elastic properties. In this paper we present the development of a software, based
The techniques described in this paper allow us to reconcile on a gradient method, which can include production data as
production history matched models with 4D information, and well as inverted seismic data in a consistent way. The
to reduce the uncertainty in reservoir properties, which haven't development is the result of a two-year project sponsored by
a real impact on the well history, but which significantly drive the European Commission, and is implemented in a
future behaviour of the field. This is a further step towards the commercial reservoir simulator coupled with a history
necessary integration of available data for better predictive matching software. The seismic contribution in the objective
simulations. Focusing on quantitative combined with function is defined in terms of elastic parameter variations
qualitative use of data enhances the within the reservoir, mapped from the seismic grid to the
multidisciplinary approach. reservoir simulation grid. The mapping involves upscaling in
the horizontal direction and both upscaling and downscaling in
the vertical direction. The data have been properly weighted
using an estimate of seismic uncertainty (covariance matrix).
1
Now with Center for Integrated Petroleum Research, University of Bergen. It is shown that in some cases this matrix will be non-diagonal
2 SPE 84464

(i.e., correlated data), and that the non-diagonal terms also

improve the performance of the optimiser. In fact, including Geological Model
Reservoir Characteristics (lithology, mineralogy)
correlations, may stabilise the optimisation routine even if the HM Parameters α

presence of correlations is not warranted by the data..Non- Fluid Flow Simulator

diagonal covariance matrices in the objective function mean Pressure/Saturation Pressure/Saturation

that efficient techniques for inversion of large matrices are Changes Not always
consistent
Changes

required. Parts of the development have been reported PetroElastic Modelling Iterative Interpretation
previously in four papers produced during the project3,6,19,20. In
Elastic Parameter Our choice Elastic Parameter
this paper, an overview of the methodology is presented Changes Changes
together with some new results from the application to real
Seismic Forward Modelling Seismic Inversion
field data from a North Sea oil reservoir and a dry gas
reservoir in the Adriatic Sea. Seismic Response Too costly Seismic Response
Changes Changes

Scope
What 4D seismic – repeated seismic acquisition after field Fig 1: Possible Levels of Matching
production – can bring to reservoir monitoring is now well
known and proven. Reservoir monitoring might be taken to The main reasons for this choice were:
mean simply observation of the reservoir during production, 1. It avoids to including into the iterative loop the cpu-time
but we believe that it should also involve updating the consuming forward seismic modelling; on the contrary each
properties of the reservoir as these observations become survey, separately or together, are inverted once;
available. The project then aimes to use the new information 2. It takes into account both saturation and pressure effect (through
on fluid movements provided by repeat seismic acquisitions PEM) and avoids a difficult and possibly not consistent
jointly with that offered by production data in an extended “inversion” in terms of fluid changes. These interpreted
history matching process to maintain consistency. So the main pressure/saturation changes can’t be considered as observations
purposes were defined as follows: independent of the reservoir model.
1.Focussing on reservoir engineering side, targeting the 4D
Our approach is then a quantitative use of 4D seismic data,
seismic impact to reservoir updates;
involving a simultaneous minimisation of the mismatch
2.Ensuring consistently matched models explaining both time- (objective function) between all types of measured and simulated
lapse seismic and well production data; data. - “observed” values obtained by inversion of the seismic
3.Using well proven Computer-Aided History matching (CAHM) signal, and “modelled” values, obtained by the fluid flow
techniques particularly gradient-based methods and sensitivity simulator coupled with a petro-elastic model. The seismic
analysis; contribution is defined in terms of elastic parameter variations
4.Developing and implementing an approach demonstrated on within the reservoir. The main ideas of CAHM are reviewed in
real cases, through internal prototypes; Appendix A.
5.Leading to a commercial tool available immediately after the In that framework, the main points of our “4D Seismic History
end of the project, closely following specifications coming from matching” are:
prototype tests. Global loop: gradient-based technique to reduce both production
HUTS Approach and 4D mismatch, each term weighted by prior observation
When we started the project most of the previous works uncertainties; coded first in prototype and then in commercial
referred to interpretations of 4D data in terms of saturation software (Fig. 2).
changes or to confrontation between synthetic and real Time-lapse seismic inversion: impedances (or whatever relevant
seismic. However we firmly opted for a matching loop at the parameters) mapped into the reservoir grid with associated
elastic domain level, which emphasised the importance of the uncertainty, such as posterior covariance matrix after inversion, to
petro-elastic modelling. PEM, or rock physics module weight the seismic term.
converts elastic properties of saturated porous medium from Petro-elastic Model: coupled with a fluid flow simulator, with
simulated fluid and static rock properties, and makes the gradient computations, to be used by the history matching
bridge between fluid flow and wave propagation domains (Cf. software; the formulation is general - for future cases - and
Fig. 1, and Ref. 3,6). specific - for the cases under study in HUTS.
This approach was designed to produce one useful tool to assist
the engineers in the global process, certainly not excluding
qualitative and manual procedures. The cases studied have
stressed the need for combining both approaches.
The up and down-scaling issues have been partially “solved”
using only the reservoir grid scale, that means if possible using
seismic inversion which provides better vertical resolution (like
geostatistical inversion constrained by well data), horizontal
upscaling of inverted impedances and including a correlation
matrix in the objective function.
SPE 84464 3

Project work packages Reservoir Seismic

The work load was divided into 6 main work packages, with Upscaling Geological Model Specific rules
specified leaders: (1) Research development – TOTAL, (2) Mineralogy, lithology...
lithology...
Vsh, etc.
Optimisation issues – NORSK HYDRO, (3) Seismic

Base Survey
Reservoir
Interpretation – NORSK HYDRO, (4) North Sea case – TOTAL, Parameters α Rocks
(5) Adriatic Sea cases – ENI, (6) Commercial development – Fluid Flow Petroelastic

UPDATING
SCHLUMBERGER. The main components and steps were: Modelling Fluids Model
Prototype coding (WP1): Hm4D prototype drives the

Survey
Well

BaseSurvey
Pressures
minimisation and calls the commercial fluid flow simulator27, performances Impedances
Saturations + gradients wrt α
+ gradients wrt α
+ gradients wrt α
matching pressure/saturation (version 1) and elastic parameters

SeismicBase
(version 2) as observations (both as absolute values or changes MISMATCH
minimisation + MISMATCH
minimisation
Geostatistical

Seismic
Inversions
between two surveys). It contains a PEM module.
Optimisation issues (WP2): the traditional least square Well observations
History-production
Impedances

formulation has been retained, with a standard optimisation at Reservoir scale

algorithm, extended to include a non-diagonal weight matrix. The Fig 2: HUTS workflow
balance between terms have been partially solved by focussing on
best possible covariance matrix error estimate. Determination and New Objective Function
use of this seismic correlation matrix has been implemented. The History matching is a minimisation problem (Cf. Appendix
gradzone analysis technique23 has been extended to the new 4D A), where the objective function is the mismatch we aim to
term. reduce as much as possible. The aim of the HUTS project is to
combine the data traditionally used in history matching
Inversion techniques (WP3): the seismic inversion was
(production and pressure data) with time-lapse seismic data:
investigated separately by the partners. Acoustic impedance
3D cubes for base and monitor surveys have been produced by J (m) = α ⋅ J prod (m) + β ⋅ J seis (m), (1)
ENI for the Adriatic Sea fields and jointly by Total and where the two terms are respectively production and seismic
NorskHydro for the North Sea field (two sets of observations). mismatch, and each term can be expressed like equation (A-1)
A general formulation of PEM has been developed taking into given in Appendix A. Although it is highly recommended to
consideration immediate and future applications. It has been add a prior information term in the objective function, this has
implemented in an HUTS-PEM software, and partly in a PEM not been done in our real cases. We put two scalar coefficient
module inside the commercial fluid flow simulator27. α and β to allow some flexibility, but it should be used with
Demonstration cases (WP 4&5): a first application to a caution, as a correct estimates of the weight matrix should be
synthetic case was necessary to develop our approach. We used sufficient to balance the terms.
the PunqS3 case. We validated both pressure/saturation and The two terms are similar, as follows:
impedances matching on this example with a good synergy
1
between production and seismic observations. We also got useful J prod (m) = ( p(m) − d )T W p ( p(m) − d ), (2)
indications which appeared to be valid for the real cases. For 2
application to the field cases (Oseberg - Alpha main part in 1
J seis (m) = (s (m) − e ) Ws (s (m) − e ),
T
ORELN2 formations, and Cervia and Amelia in Adriatic Sea) we (3)
achieved to obtain satisfactory results, to validate the approach. 2
where s(m) is the seismic derived predicted results.
Commercial Software (WP6): During the project,
However, the difference in nature between these data types
development versions of the commercial fluid flow simulator requires that the weight matrices, Wp and Ws, are carefully
and history matching software with 4D features were made chosen. On the one hand the production and pressure data (d)
available. This software was released commercially shortly are characterised by a very low resolution in space, and a high
after the end of the project27,28. resolution in time. On the other hand, time-lapse seismic (e –
e.g. difference of inverted impedances) has a very low
resolution in time (several years between each seismic
survey), a medium to low resolution in space vertically, and a
high resolution in space areally. Consequently, a correct
formulation of the objective function not only requires a
proper weighting between production/pressure data and
seismic data, but also involves upscaling (horizontally) and
downscaling (vertically) the seismic data in order to compare
with the corresponding simulated data. With respect to
defining the data to be compared, inverted seismic data such
as impedance and Poisson’s ratio has been chosen. This
requires methods for 4D inversion and a petroelastic model to
calculate the elastic parameters from simulated pressures
and saturations.
4 SPE 84464

It can be shown22 that if model errors and measurement errors impedances below and above the reservoir were used. The
can be assumed Gaussian, minimising the least square data were depth converted by mapping the inverted seismic
objective function corresponds to finding the most likely cubes to a previously defined layer model of the reservoir [See
solution in a probabilistic setting. Furthermore it follows that also Ref. 2].
the weight matrix, W, should be chosen as the sum of the To produce an alternative set of 4D impedance cubes, the
inverse model error and the measurement error seismic workflow includes the following key elements:
covariance matrices. • Amplitude-preserving time-lapse pre-processing, including
Although it may be questionable, the model error is neglected spectral matching and multiple attenuation.
in this work, and we thus assume that the weight matrix is • True amplitude prestack depth migration (using time-lapse
given by the data covariance matrix. Consequently, the main specific velocity models where appropriate). Although this
focus has been how to estimate this covariance matrix, i.e., the is potentially less robust than time-migration, it is
production data and seismic data covariance matrices. In our theoretically more accurate in complex media, and can be
applications, the production weight matrix is diagonal regularised to avoid acquisition-related artefacts25.
(wi=1/σi, where σi are the measurements errors). The way to • QC of pre-stack amplitudes to check for potential problems,
construct the “seismic” or “elastic” covariance matrix error followed by horizon-oriented AVO inversion for attribute
has already been presented in two previous papers19,20. analysis.
Sensitivity Analysis • Time-lapse geostatistical inversion: we have developed a
Finally, the selection of the parameter set, m, is a crucial part method for performing simultaneous geostatistical
of history matching. One way of doing this is through a inversion on a pair of seismic volumes26 to calculate two
gradzone analysis23. The problem of applying the gradzone impedance attributes for each cell. These impedances may
analysis when seismic data is included is not at all a be linked by a prior joint probability distribution (PJD)
theoretical issue. The same procedure applied to gradient and representing the possible outcomes, e.g. due to fluid
Hessian, considered separately or globally (whether only one substitution and pressure changes. In time-lapse inversion
type of observations is taken into account or both together). there are two stages: an initial inversion for the background
However the most important part of such gradzone analysis is model which best fits both datasets, and a subsequent
experience and how to play with this tool. For production data, inversion to reveal the time-lapse difference, which
we have now acquired sufficient experience, but it appears not typically uses a different variogram and is constrained by
so easy for 4D seismic data and more experience is needed. the PJD. The fact that this is a geostatistical inversion
Alternative methods have also been used (Cf. section allows us to choose the vertical cell size to match the
Applications) reservoir simulator, which simplifies the scaling problem.
Optimisation Algorithm The calculation of a number of realisations should allow
In order to solve the minimisation problem, an optimisation estimation of the posterior statistics and thus the major
algorithm is needed. In comparison to usual production history component of the metric for the seismic history match.
matching based on gradient techniques, as soon as ∇s (m) is Covariance matrix
available, the only difference may come from the seismic term As mentioned above, this is an important part of the
weight. Since the seismic data covariance matrix may be very observation data. We investigated only one way to obtain such
large (data in every simulator grid cell) and not necessarily information based on post inversion correlation analysis of a
diagonal, the problem of inverting this matrix for use in the single realisation19,20, but other ways could be considered in
objective function also becomes an important issue. Details the future, based on multiple realisations giving standard
have already presented in a previous papers [Appendix of deviations and correlations. The current constraint is to
Ref. 20]. transform this correlation matrix into correlation lengths to
4D Data allow the the use of our inversion algorithm.
The inverted impedance (or velocity, or Poisson’s ratio)
4D cubes were an essential input of the project, to feed the Petro-Elastic Modelling
HUTS loop. But no general recommendations were made. The As already mentioned, our approach, like any other which
partners of the project used their own approach of tends to make the link between the fluid and rock properties –
time-lapse inversion. which matter for flow behaviour – and elastic and seismic
For the North Sea field, the quality of the seismic data is properties – which govern the wave propagation – needs a
generally good, and the 4D data are very consistent with the petro-elastic modelling. This link intervenes in both forward
production effects21,2. To obtain data which are better suited and inversion procedures. In our loop we perform the forward
for reservoir characterisation and history matching, the moddeling from reservoir domain to elastic one. This set of
seismic data were inverted using a special time-lapse elastic relationships is based on general theoretical principles applied
inversion technique. Here, amplitude versus angle information to the specific field under study. It consists of both analytical
is used to obtain estimates of both the acoustic impedance and formulae and empirical laws calibrated with lab
Poisson’s ratio. Based on the relation between elastic measurements29,30,31,32,34 (Fig. 3).
impedance and the VP/VS ratio given by Connolly24, the near,
mid and far offset stacks were inverted using the same layer
model. This was done for both surveys, again using the same
layer model. For the second inversion, the first inversion
SPE 84464 5

Ip, Is, ν properties, P, Sw, Sg, rs and rv, with respect to m. As for the
well properties sensitivity, we exploited an additional time-
lapse functionality provided us by the simulator. Then, the
Vp, Vs, ρ petro-elastic model has been devised to compute the
derivatives of impedances with respect to fluid properties and
K, G, ρ rock properties (porosity, ntg and lithology). This requires the
computation of derivatives of phase densities with respect to
independent fluid properties, according to the PVT tables,
Gassmann which rule the phase behaviour in the black-oil simulation
Kfr, Gfr : empirical laws Kfr, Gfr, Kmix, Km, φe, ρmix, ρm ρm : average
Vshale, Km, P, φe ρsand, ρshale
model. The last derivatives product at the r.h.s. of equation (6)
gives the sensitivity of acoustic impedances to the history
match parameter m, independently from the effect of m on
Km : Voigt, Reuss or Hill Kmix : Wood or Wood/Brie fluids properties. This product can differ from zero only if the
Ksand, Kshale Ko, Kw , Kg
rock properties α, which entered in the petro-elastic
Fig 3: PEM simplified workflow overview modelling, depend on the parameter x. The additional
advantage of computing gradients is the possibility of using
Once a petro-elastic model of the porous medium has been such PEM code for sensitivity runs during feasibility phase.
established, the acoustic impedances may be computed,
through a chain of equations, as functions of rock and fluid
properties, the latter usually given by the fluid flow
simulation, the former from the interpretation of the available
data:
I p / s = I p / s ( P , S w , S g , rs , rv , α rock ) , (4)
where αrock represents generically the properties of the
reservoir rock, going from porosity φ and net to gross ratio,
ntg, to shale and sand contents Vsh and Vsand respectively.
During the project we have compiled literature and previous
field case studies, and defined a general PEM with alternative
rules to fit the cases under study and to prepare future
evolution6. For minimisation purpose of our gradient-based
technique, we have also introduced derivatives calculations.
We took advantage of the gradient option of the commercial
fluid flow simulator27 we used in this work, which gives for a
generic set of well properties, from bottom-hole flowing
pressure to gas-oil ratio, the sensitivities w.r.t. to a prescribed
set of parameters m. This makes quite easy the set-up of the 92-88 99-88
Fig 4: Simulated Acoustic Impedance Changes
production Jacobian. More complex is the construction of the
seismic sensitivity matrix GS. This involves the computation The Fig. 4 gives an example of simulated impedance changes
of the following type of derivatives: after 4 and 11 years.
∂∆I iC  ∂I iC   ∂I C  At the end of the project, the partners share an HUTS-PEM
= − i  (5) which is suitable for further evolution (stand-alone or module
∂m j  ∂m   ∂m 
 j  t =t
M
 j  t =t
B
within the iteration loop). Inside the commercial fluid flow
where i labels the cell and j the parameter. Dropping for simulator27 a PEM option has been implemented with the main
simplicity the subscript j and applying chain derivatives rule to same features. The commercial history matching software28
the equation (4), we can write the derivative of the acoustic has no plug-in yet for a user’s specific PEM, but it should
impedances computed in the i-th cell as: become necessary in the future.
∂I iC ∂I C ∂pi ∂I C ∂S wi ∂I C ∂S g i
= + + + Application
∂m ∂pi ∂m ∂S wi ∂x ∂S g i ∂m Four application cases have been selected for feasibility
T (6)
∂I C ∂rSi ∂I C ∂rVi  ∂I C  ∂α i and demonstration purpose: a synthetic case, PunqS3, a North
+ + +   , Sea case, Oseberg field, and two Adriatic Sea cases (Amelia
∂rsi ∂m ∂rvi ∂m  ∂α i  ∂m
field is the the only described in this paper). The results
where we assumed that the thermodynamic behaviour of the obtained on Punq-S3 have already been presented and
fluids is simulated in the dynamic model using an extended discussed, and the North Sea field case has been partly
black-oil formulation which describes vaporised oil and presented in previous papers3,6,20.
dissolved gas through rv and rs, as allowed by the formulation Synthetic case
of the commercial fluid flow simulator27. The computation of
The PUNQ-S3 test case has been described fully elsewhere,
impedance sensitivity as in equation (6) requires preliminarily
and all of the data are also available on the TNO web site. In
the availability at time t of the gradients of simulated fluid
brief, the reservoir model contains 19x28x5 grid blocks of
6 SPE 84464

which 1761 are active, with petrophysical properties, porosity, plain deposits in the lower part (LN) and fluvial deposits in the
horizontal and vertical permeability, set by running standard upper part (UN). The Tarbert formation, on the top, is
sequential Gaussian simulation (SGSIM) conditioned at wells. composed of a series of transgressive sheet sands.
The reservoir initially contains a gas cap and an oil rim. It is The production started in 1988 from the A-platform, which is
bounded to the south and east by faults, while to the north and connected to the (drilling) B-platform. Later production also
west there is an extensive aquifer. There are six producing started from the C-platform to the North. During autumn 1997,
wells located near the initial gas-oil contact. the oil production from the Oseberg Field went off-plateau.
Simulated production data for the first 8 years, including The Oseberg field was chosen as the HUTS field case because
pressures, water cuts and gas-oil ratios for each well, were of the existing 4D seismic data based on seismic surveys
available for history matching. The Punq-S3 dataset provided acquired in 1992 and 1999, respectively. There is also a
by TNO does not include time-lapse seismic data of any kind, seismic survey from 1982, but these data has not been used
but a realistic petroelastic model 20 can be used to generate because of less good data quality. For more information about
values of acoustic impedance based on cell pressure, the seismic surveys and interpretation is referred to refs. 2,
saturation and porosity. and 21.
Base Line History Match
In order to get a model of reasonable size, an area covering the
Production Misfit
Production Term vs Iterations
3.5
Southern part of Alfa ORELN2 was chosen as the model area.
3.0 Production
Production
drive A fine scale geological model was built using the stochastic
Seismic
2.5
Seismic drive modelling techniques of the RMS modelling software. This
Both Prod. & Seis.
2.0
BOTH model was then upscaled to a reservoir simulation grid of
1.5
dimension 55x39x11.
1.0

0.5

0.0 Oseberg C
0 5 10 15 20 25
# iter

Seismic Misfit
Seismic Term vs Iterations
180.0

170.0

160.0
Production
Production drive
150.0 Oseberg Vest
Seismic
140.0 Seismic drive
130.0
Both Prod. & Seis. Oseberg B
120.0

110.0
BOTH
100.0

90.0

80.0
Oseberg A
70.0

60.0

50.0

40.0
1 3 5 7 9 11 13 15 17
Fig. 6. Location map.
# iter

Fig 5: Objective function (PunqS3) Base line history matching to production data only was
performed using the commercial software27,28. Two different
This synthetic case has been used to test the HUTS approach, approaches were utilised: i) A comprehensive history match
first matching pressure and saturation changes in conjunction with a total of 63 parameters (pore volume, horizontal, and
with production data3, then matching elastic changes6. Finally vertical transmissibilities) defined through a gradzone analysis
investigations were made on the impact of covariance matrix technique23. The result is described in Ref. 20. ii) A simpler
added to objective function when the observations are noisy20. case using only a single history matching parameter: The pore
The main lessons of this case were the good synergy (Cf. Fig. volume of the entire model. The final multiplier is 1.27, a
5) of 4D data with production data leading to a better model, value which is also agreed upon by the field geologists. The
where the fluid distribution is better conditioned with time quality of the match was very similar to the previous one. This
lapse data. The second important result is the importance of a is also confirmed by visual inspection of the well production
good estimate of data correlations, as the iterative process can histories. It should be noticed that the solution in case 1
converge to a wrong model20. Finally, the relatively low contains several much more extreme values for the parameters
decrease of seismic mismatch stresses the need to build an than case 2. This exercise thus shows the importance of not
initial model which takes into account the 3 D seismic (Base using a parameterisation technique, like the gradzone
survey), but that was not considered in the Punq-S3 case. analysis uncritically.
North Sea case Case 2 was chosen as the basis for the further studies
The Oseberg field is located in Blocks 30/6 and 30/9 in the including seismic data.
Norwegian sector of the North Sea some 140km WNW of Mapping of seismic data to simulation grid. Up-
Bergen, see Fig. 6. The reservoir comprises sandstones from and downscaling
the Middle Jurassic Brent group in eastward-dipping, tilted Quantitative history matching to seismic data requires the
fault blocks situated on the Eastern edge of the Viking graben. definition of an objective function measuring the mismatch
The lower part of the Brent group consists of coarse grained, between the measured and simulated seismic. In the HUTS
fairly thick, massive sandstones belonging to the Oseberg, project, acoustic impedance and Poisson’s ratio have been
Rannoch and Etive (ORE) formations. Above these we find chosen as the data to be matched. The definition of the
the Ness formation composed of marginal marine and delta
SPE 84464 7

objective function requires that the measured and simulated

data are represented on the same grid, i.e., the simulation grid.
In the Oseberg case, the result of the inversion is inverted
seismic data are represented on a layered grid2. The horizontal
gridding corresponds to the resolution of the survey, i.e.,
25x37.5m. The vertical layering is determined as a part of the
inversion process. Cell thickness after depth conversion is
around 10m. On the other hand, the simulation grid cell size is
100x200m horizontally. The vertical layer thickness follows
the geological model and varies from almost zero to a
maximum of around 25m. That is, an upscaling of the seismic
data is required horizontally. In the vertical direction, both an
upscaling and a downscaling is required, depending on the
layer thicknesses in each point.
Several ways of mapping the data were investigated, and it
turned out to be critical for a good result that the noise in the
data were filtered out before upscaling. That is, the data were
smoothed by applying a moving average technique, and then ∆AI (99 –92)
upscaled using a standard arithmetic average. Downscaling is -0.15 0.05

performed through direct sampling from the coarse grid. A (after 7 years)
Fig 7: Simulated vs. inverted impedance changes
proper up- and down-scaling of the data covariance matrix is
also required. More details can be found in Ref. 20.
History Matching With Seismic Data
Several regression runs were performed using different
parameterisations, and different formulations of the objective
function (seismic term only, production term only, both terms,
rescaled seismic data, etc.). Two sets of observations have
been used and two series of regressions performed.
Firstly using inverted Poisson’s ratio. Although the upscaled
time-lapse seismic seems to contain significant information
about the spatial distribution of gas and gas displacement, it
turned out to be difficult to get a very good match of the data
using the regression algorithm, with pore volume multiplier.
Sensitivity runs showed that pore volumes were more
sensitive than transmissibilities. One of the reasons may be
that the static model does not agree with the inverted
impedances (simple cross correlation analysis). That leads the Fig 8: Residuals of ∆Ip
gradzone analysis to be not efficient, as it led to select too
many gradzones with no apparent correlations to gather the The gradzone analysis requires time consuming runs with
cells. A match obtained by tuning the parameters by hand many gradients (sampling of cells) and was not yet
apparently matched the data better than both the base case systematically investigated, but better results were obtained
model and the results of the regression runs, but the seismic compared to the previous one based on pore volumes. So
RMS deviation between this model and the data was not any another visual analysis of observations (residuals, Cf. Fig 8)
smaller than the other models. In the most successful lead to the definition of 34 parameters (Cf. Fig. 9)
regression run, 353 pore-volume multipliers were adjusted
based on a “regular box” parameterisation. In the manual
match, a selection of multipliers was adjusted using a
parameterisation based on visual inspection of the seismic
data. Details are given in the previous paper20.
Secondly using inverted impedances on a restricted area (Cf.
Fig. 7) and after the use of geostatistical inversion tool. For
this second series, the pore volume adjustments using absolute
impedance values (first monitor survey, as the base vintage is
not available) were difficult as well, but the transmissibility
multipliers appear to be a more efficient matching parameters,
using time-lapse differences of acoustic impedances.

Fig 9: Gradzones based on Residuals of ∆Ip

8 SPE 84464

We finally got decreases in both global and separate For reservoir management purposes, a gas-water dynamic
production and elastic mismatch terms, with the expected model has been set-up to simulate field production. The model
synergy. The Fig. 10 summarises the results showing consists of 79x26x5 corner point cell, with onlyb 8190 active.
separately the production and seismic objective function The first layer models the PL3F, while the remaning four
versus iterations, but as a results of regression runs minimising model the PL11F. The sealing shale layer has not been
both terms. The same figure illustrates the good impact of explicitely modelled..
introducing a correlation matrix in the seismic term. The This model was calibrated with respect to the available
matrix was defined with correlation lengths estimated from the production data, shut-in pressures and monthly well water
first dataset20 (2x2x11 gridcells), which may not be adequate rates collcted till June 2002, in a conventional history
for this alternative dataset. However, the improvement is clear, matching (CHM) project.
as we compare to uncorrelated RMS. Preliminary Analysis on 4D Data
Base and Monitor surveys were re-processed for time-lapse
Influence of correlated Seismic data vs uncorrelated Seismic data purposes, thus generating two way times (TWT) cubes of
13 1.7 inverted absolute acoustic impedances.
1.65
Acoustic impedances were converted from the TWT domain
12.5
1.6
into depth,and sampled in a fine scale grid, aligned with the
Seismic rms acquisition directions, with a lateral spacing of 12.5x12.5
UNCORRELATED RMS

12
Without
1.55
metres. This is consistent with the typical distances between
correlations 1.5
seismic receivers during the two acquisition campaigns. The
11.5
With correlations 1.45 vertical spacing of the The depth spacing of the fine scale grid
Production rms 1.4
is given by the layering of the dynamic model. Taking into
11

Correlated, Prod Correlated, Prod

account the typical seismic resolution in this reservoir,
approximatey 10 m, we can infer that the seismic information
1.35

Uncorrelated, Seis Correlated, Seis

10.5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
1.3
has been downscaled and, consequently, a vertical correlation
Iteration #
affects the mapped data.
Fig 10: Objective function – Oseberg case (Effect of correlations) A first qualitative analysis of the 4D effect can be carried out
This case certainly needs more work, to improve the by looking at the normalised impedance variations ∆znorm,
parameterisation strategies, the covariance matrix estimates defined as
and the initial static model. It is expected to improve the z − z adria
∆z norm = Amelis ,
quality of the matched model, not only in terms of mismatch z adria
decrease, but in terms of better predictivity.
shown for the first and second layer of the fine scale grid in
figure 11.
The Amelia Gas field
The Amelia field consists of a sequence of dry gas
accumulations lying from 2740 m to 4100 m depth ssl.
Production started in 1973 with more than 150 vertical wells
perforated from 4 platforms.
Two 3D seismic surveys were shot over the field area:
Adria in 1992, mainly for structural reasons, and Amelia in
1998, for reservoir purposes.
The surveys were reprocessed for time-lapse purposes,
deriving acoustic impedance cube by analysisng andd a good
4D effect has been found for two pools, namely the PL3F and
the PL3F1. The aforesaid layers are located on top of the
sequence, between 2740m and 2780m of depth ssl, and sealed
by a thin , 1 m thick, shale layer. Even tough sealed, PL3F and
PL3F1 share both the same lateral acquifer and a pre-
production gas-water contact (GWC) …
Gas bearing sands were found mainly in the PL2F and in
the upper part or the PL2F1, while the deeper part of the P1 is
mostly below the GWC.
Both PL2F and PL2F1 has been put on stream in
commingle since 1973, but most of the production, about
56%, occurred between the dates of the two 3D acquisitions.
Notably, thse two pools are two of the most importat
reservoirs of the field, with… Good pressure support has been
provided by the acquifer, with many wells watered out in the
late 90’.
Fig. 11: Normalised impdance variation from fine scale depth
converted data, first layer (top) and third layer (bottom).
SPE 84464 9

A clear increase of the impedance, up to 7%, from base to Conventional History Matching
monitor dates can be seen close to the well area. This effect t The model achieved at the end of the CHM project was
can be due to the water encroachment from the lateral used as baseline model of this 4D calibration workflow. This
acquifers, with a possible contribution give by the pressure model has been obtained by means of a systematic application
depletion which is more pronounced around wells. of gradient based methodologies to history match static
From fine to Coarse scale pressures and monthly water rate.
Depth converted impedances have not been directly mapped to The quality the conventional history matched model can be
coarse scale grid. Rather, a 2D moving average technique, quantified by looking at the value of the root mean squares ,
based on 21x21 cells window, has been applied to discern the RMSprod for the production misfit, defined as follows :
main trend from the residual noise in the fine-scale data. 2
Following the procedure recommended by Aanonsen et al.17,18, RMS prod = J prod ,
N prod
average correlations along NW-SE and NE-SW has been
computed, and isotropic correlation length of about 5-6 where Jprod is the production objective function defined in
seicmic cells, see figure 12, has been estimated. equ. (2) and Nprod is the number of production data.In addition
1.0E+00 to production RMSprod, misftit for shut-in pressures (RMPSp)
and water rates (RMPSwpr) can be defined. Baseline model
8.0E-01
RMSp, RMSwr and RMSpr values are reported in Table 1.
6.0E-01 NW-SE
RMSprod 3.504
NE-SW
RMSpr 0.7981
correlation

4.0E-01

RMSwpr 3.601
2.0E-01
Table 1: Production RMS values for the baseline model
0.0E+00 provide at the end of the CHM
0 5 10 15 20 25 30 35
-2.0E-01
A Gassman-based petroelastic model, calibrated using CPI
and P-log data, has beed developed and code as a subset of
-4.0E-01
cell
rule in the HUTS-PEM software. Computed impedance
variation ∆zc, water saturation variation ∆sw and pressure
Fig. 12: Average correlations for residual (after moving averaging) variation ∆p for the third layer of the model are shown in
impedance variations (Amelia-Adria). figure 14 (∆zc), figure 15 (∆sw) and figure 16 (∆p).
The main trend, in other words the filtered data, has been
horizontally upscaled to the dynamic grid.

Fig. 14: Baseline model synthesized impedance variation for the

third layer of the model. Color are scaled as shown in figure 13.
Fig. 13: Depth converted impedance for the third layer of the
coarse scale grid after moving averaging.Cells below the GOC
have been shadowed.

We may notice, figure 13, hat the filtering procedure, as

expected, highlights the main trend in the time lapse data, that
is acoustic impedance increase toward wells area.
Nevertheless some strong decrease in the acoustic impedance
far from the reservoir area can still be noted, for instance
toward South-East. Moreover, dynamic grid cell cover about
8.5x8.5 seismic cell, hence a lateral correlation spanning
aèproxuimately 5 seismic cells can be neglected.
As far as the vertical correlation is concerned we can
assume that, due to the downscaling of the data from seismic
to reservoir model, some correlation exists, and, by using the
same argument presented for Oseberg data, it reasonably spans Fig. 15: Basline model simulated sw variation for the third layer.
about 2 layers.
10 SPE 84464

multipliers were constrained between 0.7 and 1.4, while

transmissibility multipliers can vary between 0.05 and 50.
After some sensitivity regressions, it was realised that a
reasonable way to improve both the time-lapse and the
production misfit was to calibrate first transmissibility
multipliers, and, next, use a plausibly good results in terms of
transmissibility multipliers to calibrate the pore volume
multipliers. Calibrating transimssibility and pore volume
alltogether did not give any appreciable results.
To vealuate the role that correlations in the seismic data
may play in the optimisation, regressions have been carried
out by considering three different schemes in the definition of
the seismic covariance matrix:
Fig. 16: Baseline model pressure variation for the third layer.
• Only vertical correlation spanning two cells – C(0,0,2),
• 3D isotropic correlation spanning two cells. – C(2,2,2),
The pressure drop shown in figure does not seem to have • no correlation – C(0,0,0).
any impact on ∆zc. Computed Impedance variation from PEM Transmissibility calibrations In the transmissibility
is ruled by the saturation variation, as it can be seen by noting calibration we have evaluated the impact of correlation by
the clear footprints of ∆sw on ∆zc: the variations in the acoustic considering different possible correlation length, including
impedance follow the ecnroachment of the water. It is pure diagonal correlation.
reasonable to infer that the petro-elastic model devised for Transimissibility calibrations for the different correlation
Amelia allows explaining time-lapse effect due to fluid schems can be compared on the basis of the production RMS
substitution during the drainage process, while pressure values versus the number of reservoirsimulation for the
effects, if any, can not be accounted. Moreover, the computed production (figure 18) and of the uncorrelated RMS seismic
ranges are broader than observed ranges. values versus the number of reservoir simulations (figure 19)..
Parameterisation 3.600
The parameterisation used in the 4D history matching of
the reservoir under issue are based on Gradzone analysis23 3.400
C=(0,0,0) RMSp
applied to the hessian of an objective function J as defined in 3.200
C=(2,2,2) RMSp
equ. (1), with α=β=1.The seismic objective function Jseis is
production RMS

3.000 C=(0,0,2) RMSp

defined using a correlation matrix based only on a vertical
correlation spanning 2 cells. 2.800

In this way, two sets of parameters, one consisting of 24 2.600

horzontal transmissibility multiplier and one consisting of 31
pore volume multipliers, have been defined. Examples of the 2.400

type of zonations are shown in Figure 17 for the third layer. 2.200

2.000
1 6 11 16 21 26 31 36 41
eclipse simulations

Fig. 18: Transmissibility calibrations - RMSp versus

reservoirsimulations.

Considerable reduction in the production RMS can be

noticed in all three cases, with a faster convergence for case
C(2,2,2).
If we consider now the behaviour of the uncorrelated
seismic RMS, it is worth noting that the relative reduction in
the seismic misfit is much lower than the reduction observed
for the production misfit. The use of a 3D correlation model,
case C(2,2,2), seems to slightly penalise the match of the
seismic.. However, the differences are so small that it is
Fig. 17: Pore volume (top row) and horizontal transmissibility difficult to give an interpretation.
(bottom row) multipliers in the third layer of the dynamic model.
The impact of the correlations is more evident if we pay
Optimisation attention to the optimised values of the 24 transmissibility
Given a set of parameters m, the baseline reservoir model multipliers (figure 20). The optimal multipliers for the
was optimised using an implementation of the Levemberg- refression with correlation schems C(0,0,2) and C(0,0,0),
Marquardt minimisation strategy suited for weighted least- largely differ from the base model, with most of the values
squares problems. Plausible ranges for the parameters have equal to the lower bound and one multiplier far above 2. The
been set by by defining upper and lower bounds. Pore volume use of a 3D correlation model, scheme C(2,2,2), leads to more
SPE 84464 11

realistic values, with only a small set of the multipliers close the one found in the previous transmissibility calibrations, as
to the lower bound 0.05, and no values above 2.0. we may see by comparing with Note that, using a relatively
The considerations made by looking at the multipliers large stencil for the correlation in the seismic observation,
values give the plausible indication that the use of a case C(2,2,2), led to a slightly worst results if compared with
correlation C(2,2,2), even if it has not been warranted by the the other cases.
data, partly prevents the occurrences of extreme values in the
parameters. The reservoir model obtained as a result of the 2.2400
C(2,2,2) case regression has then been chosen for the further
2.2350
step of optimisation, that is the calibration of the pore
volume multipliers. 2.2300

1.920

Production RMS
2.2250

C=(0,0,0) 2.2200
1.900
C=(0,0,2)
Uncorrelated Seismic RMS

2.2150
1.880 C=(2,2,2)
C(0,0,0)
2.2100
1.860
C(2,2,2)
2.2050
C(0,0,2)
1.840 2.2000
1 6 11 16 21

1.820 eclipse simulations

Fig. 21: Pore volume calibrations – RMSp versus

1.800
reservoirsimulations. Different correlation models for the TL term
1 6 11 16 21 26 31 36
eclipse simulations
were investigated.

Fig. 19: Transmissibility calibrations – RMSTL versus

reservoirsimulations. Different correlation models for the TL term 1.86E+00
were investigated
C(0,0,0)
1.84E+00
C(2,2,2)
Uncorrelated RMS seismic

8.0

C=(0,0,0) C(0,0,2)
1.82E+00
6.0 C=(0,0,2)
Multiplers values

C=(2,2,2)
1.80E+00

4.0
1.78E+00

2.0
1.76E+00
1 6 11 16 21
eclipse simulations
0.0
1 5 9 13 17 21 Fig. 22: Pore volume calibrations – RMSTL versus
i-th parameter reservoirsimulations. Different correlation models for the TL term
were investigated
Fig. 20: Transmissibility calibrations – Optimal transimissibility
multipliers for Different correlation models for the TL term were In the calibration of the transmissibility, we have figured
investigated out a connection between the variations of the multipliers and
the use of a certain correlation model. The use of an3D
Pore volume calibrations
isotropic correlation model, scheme C(2,2,2), seemed to
Starting from the model optimised in the regression C(2,2,2),
reduce the tendency of the transmissibility multipliers to hit
we have run three representative regressions by using three
the boundaries of the feasibility region during the regression.
correlation models already considered in the calibration of the
Moreover, this scheme provide with a better weighting of the
transmissibility, labelled as C(0,0,0), C(0,0,2) and C(2,2,2).
production with respect to the seismic mismatch. In the case of
To evaluate the results of the regressions, let us consider
the calibration of pore volumes, the same considerations can
first the behaviour of the production RMS values versus the
not be drawn: Although the use of a 3D correlation model
number of reservoir simulations (figure 21). We may note that
prevents the deterioration of the production misfit, the optimal
the production misfit has not been improved in any of the
values of the matching parameters (pore volume multipliers)
three regressions. Moreover, for C(0,0,0) and C(0,0,2) cases,
for the regressions at hand (see figure 23) seems to be
the quality of the production match is meaningfully
reasonably independent from the correlation models.
deteriorated. Such results can be compared with the behaviour
of the uncorrelated seismic RMS values (see figure 22). All
regressions run to calibrate the pore volumes gave an
improvement of the uncorrelated seismic RMSs greater than
12 SPE 84464

1.4 C(0,0,0)

C(0,0,2)
Pore volume multipliers

1.2 C(2,2,2)

1.0

0.8

0.6
1 5 9 13 17 21
HMPORVM
Fig. 24: Impedance variation in the second layer of the P2 pool for
Fig. 23: Pore volume calibrations – Optimal multipliers. Different baseline model (top) and for OC(0,0,2) model
correlation models for the TL term were investigated
Notably, only the model OC(0,0,2) leads to a slight
After the three calibration of the pore volumes, further tests improvement in the match of the pressure profiles (about
were done to improve the match of the time-lapse seismic 1.5%), while the cases OC(0,0,0) and OC(2,2,2) degrade the
without degrading the conventional production data match. pressure match, but not that much
Unfortunately, these tests did not give any appreciable results.
In Table 2, the matches of the three models are reported in Conclusions
terms of relative variation of the different RMS values with The outcomes of this 2-year project are mainly demonstration
respect to corresponding baseline values. of the feasibility of our approach, acquisition of experience,
and availability of new industrial software tools.
OC(0,0,0) OC(0,0,2) OC(2,2,2) Through both synthetic and real cases, the following points
RMSp were demonstrated:
-36.30% -36.42% -36.76% A proof of concept on a synthetic case with confirmation
RMSs that a good production matched model could in fact be
-15.41% -15.35% -15.12%
worse in terms of fluid distribution, so we need to take both
RMSwr -36.49% -34.27% -36.96% information into account in a global process (Fig. 2);
For synthetic and real cases and almost all regression runs,
RMSpr 5.95% -1.52% 8.78% the synergy between minimisation of both types of
observations was effective and the gradient-based
Table 2: relative (with respect to the baseline model) RMS values
regressions successful to reduce both mismatches (Fig. 6,
variations fro the three optimal models.
10); Hm4D and its PEM module proved to be efficient;
It is worth noting that the improvements in the seismic The petro-elastic model is a key element of success and this
mismatch, about 15%, even though not negligible, are far from is obviously a challenge in real applications; the coding of
being considered satisfactory. This can be confirmed by the PEM software had the merit of focusing on theoretical
comparing the impedance variation in the third layer and general insight of the physics involved, by compilation
computed using model OC(0,0,2), figure, with the baseline of literature, internal expertise and previous studies; the
impedance variation, figure… gradient computations, needed for the minimisation
The 15% of uncorrelated seismc RMS reduction has been procedure, allow us to run sensitivity studies to almost any
obtained mainly by systematically reducing the 4D effect due static and dynamic reservoir parameters;
the GWC movement from ’91 to ’98. Outside the area covred The Oseberg case, strongly stressed the need for an initial
by the GWC movement, iimpedances have been slightly static geological model consistent with the seismic base
decrease by increase the ciomputed pressure depletion which survey, not only in terms of structure but also of
makes the fluid mixture at reservoir condition lighter.The petrophysics properties populating. We recommend to first
main 4D feature, acoustic increase close to the wells, has not adjust the pore volumes using the absolute impedance values
been captured by the optimisation, at time t=0 (base survey), with a loop on initialisation of the
More significant can be considered the improvement in the reservoir simulator followed by petro-elastic modelling (and
calibration of the conventional production data, particularly as using gradient minimisation) – but this was not done for
far as historical water rates are concerned. Oseberg A pre-production survey from 1982 was available,
but not used because of less good data quality;
Quantification of the seismic inversion quality using multi-
realisations or auto-correlation of a single realisation to
obtain a correct estimate of uncertainties is crucial to
weight the seismic term in the objective function, in order to
ensure convergence (demonstrated on Oseberg case, Fig. 10)
SPE 84464 13

to a better match solution (demonstrated on PunqS3 case20); successful on real fields. The improvements of match models
To avoid the vertical downscaling issue, an inversion obtained using the 4D data are undeniable but under our
process which permits a better vertical seismic resolution expectations for the 3 real fields, partly because of time
is preferred (for instance strained by well information and constraints – all our ideas have not been yet explored because
stratigraphic grid); reaching the end of the contract.
Parameterisation is a key issue in history matching and a There are various reasons for this limited success on real
successful match can only be obtained if the pertinent applications, for instance:
reservoir parameters are chosen. The gradzone analysis The North Sea field is largely reported for effective use of
technique (using 1st and 2nd order derivatives) were 4D seismic, but the area chosen for the project suffered from
successful for the PunqS3 case, but less easy to use for the acquisition problems, the local static model was very far
real cases. The provisory conclusion is that it is still from the base survey and the inverted impedances not
applicable but can largely benefit when associated with a calibrated on all well logs.
qualitative analysis of all available data. This was the way The specific PEM (set of rules established from lab
taken for the real cases and this a domain where further experiments) was not sufficiently accurate. This stresses the
works are needed. importance of the PEM.
Recommended methodology: The inversion performed in Poisson’s Ratios was noisy.
• First, at the earliest stage, consider the implication of the Recent runs using time-lapse acoustic impedances (and
PEM in the 3D geo-modelling workflow, and build an initial another inversion package) led to improve the results.
model making full use of the 3D seismic base survey when Next steps
populating the model with petrophysical properties. The end of the HUTS project is actually a starting point. We
• Perform a petro-elastic study and include these aspects in the are now ready to apply our approach and tools to other
upscaling phase of the reservoir model. operated fields and to interact with the operational assets,
• Use the coupling of flow simulator and PEM for while improving our tools and consolidating our methodology.
feasibility/sensitivity study (using gradients if available), and Tests of the first commercial 2003 release of commercial
determine the most relevant seismic derived attributes. software27,28 with 4D option and specifications for further
• Perform a time-lapse AVO inversion, in depth, possibly with evolution.
multiple realisations, possibly with a vertical reservoir grid Strategies of parameterisation/regression and guidelines.
resolution and define the associated covariance matrix Scaling issues at the level of the PEM and Seismic
(standard deviations and correlation lengths); upscale elastic observations – the usual PEM laws apply a small scale and a
values and standard deviations. preliminary upscaling phase is needed (e.g. anisotropic bulk
• Use of monitor survey in a qualitative way for preliminary moduli).
adjustments of reservoir properties if possible or at least to Investigation of a new workflow to build an initial static
select sensitive parameters in any history matching model consistent with the base-survey, taking into account
procedure (manual or automatic). PEM needs at early stage.
• Perform sensitivity analysis using gradient, Hessian and Extensions of PEM to geomechanics and thermal effects.
gradzone analysis combined with any other qualitative New thoughts on objective function formulation.
approach, considering various types of observations – time- Use of no-gradient optimisation when the fluid flow
lapse differences/absolute values, production only, simulator do not provide this output.
impedances only or both. Acknowledgements
The authors are grateful for the permission of European
• Adjust the porosity values using absolute impedances, in an
Commission, Total E&P UK PLC, Norsk Hydro, ENI and
iterative loop without production period (wells shut), for
Schlumberger, and the Oseberg Licence partnership for
broad adjustments, not detailed.
permission to publish this paper. The work has been partly
• Finally perform the HUTS loop with separate or global
financed by the European Commission under contract number
objective function (sequence strategy to be consolidated).
ENK6-CT-20000-00061. We also acknowledge all the people
Main deliverables
who have contributed to the HUTS project for their technical
1) The methodology suggested by the project (HUTS public
support during the last two years.
domain reports).
2) Development of a general petro-elastic model, to be
References
internally maintained, amended and improved by each 1. Marsh, J.M., et al.: “The Use of 4D Seismic in Reservoir
partner. A corresponding software (PEM), as a module Management,” EAGE 63rd Conference and Technical Exhibition,
used in conjunction with the commercial software27,28 or as Amsterdam, June 2001.
a stand-alone program for sensitivity study. 2. Rutledal, H., et al.: “Time-Lapse Elastic Inversion at the Oseberg
3) An internal prototype Hm4D for further R&D Field,” EAGE 64th Conference and Technical Exhibition,
developments if necessary, by HUTS partners. Florence, June 2002.
4) A version of the commercial fluid flow simulator27 and 3. Gosselin, O., Cominelli, A., van den Berg, S., and Chowdhury,
history matching software28 with 4D options. S.D.: “A Gradient-Based Approach for History matching of Both
Other lessons learnt from the real cases: Production and 4D Seismic Data,” Proc. 7th European
Conference on the Mathematics of Oil Recovery, Baveno, Italy, 5
In 2-year time the project aimed to develop a methodology, – 8 Sept. 2000.
implement it in industrial software tools and prove both to be
14 SPE 84464

4. Arenas, E.M., Oldenziel, T. and van Kruijsdijk. C.P.J.W.: at the SPE Annual Technical Conference and Exhibition, San
“History Matching a Reservoir Model to Time-Lapse Seismic Antonio, Texas, 29 Sept – 2 Oct. 2002.
Using the Pilot Point Method,” EAGE 63rd Conference and 19. Aanonsen, S.I., Cominelli, A., Gosselin, O., Aavatsmark, I, and
Technical Exhibition, Amsterdam, June 2001. Barkve, T.: “Integration of 4D Data in the History Match Loop
5. van Ditzhuijzen, R., Oldenziel, T. and van Kruijsdijk. C.P.J.W.: by Investigating Scale Dependent Correlations in the Acoustic
“Geological Parameterization of a Reservoir Model for History Impedance Cube,” Proc. 8th European Conference on the
Matching Incorporating Time-Lapse Seismic Based on a Case Mathematics of Oil Recovery, Freiberg, Germany, 3 – 6
Study of the Statfjord Field,” paper SPE 71318, presented at the Sept. 2002.
SPE Annual Technical Conference and Exhibition, New Orleans, 20. Aanonsen, S.I., Aavatsmark, I., Barkve, T., Cominelli, A.,
Louisiana, 30 Sept - 3 Oct. 2001. Gonard, R., Gosselin, O., Kolasinski, M. and Reme, H.: “Effect
6. Gosselin, O., van den Berg, S., and Cominelli, A.: “Integrated of Scale Dependent Data Correlations in an Integrated History
History matching of Production and 4D Seismic Data,” paper Matching Loop Combining Production Data and 4D Seismic
SPE 71599, presented at the SPE Annual Technical Conference Data,” paper SPE 79665, presented at the SPE Reservoir
and Exhibition, New Orleans, Louisiana, 30 Sept - 3 Oct. 2001. Simulation Symposium, Houston, Texas, 3 – 5 Feb. 2003.
7. Arenas, E., van Kruijsdijk. C.P.J.W. and Oldenziel, T.: “Semi- 21. Rutledal, H. et al.: “Oseberg 4D Study – The Search for
Automatic History Matching Using the Pilot Point Method Remaining Oil,” EAGE 63rd Conference and Technical
Including Time-Lapse Seismic Data,” paper SPE 71634, Exhibition, Amsterdam, June 2001. Rutledal, H., et al.: “Oseberg
presented at the SPE Annual Technical Conference and 4D Study. The Search for Remaining Oil,” EAGE 63th
Exhibition, New Orleans, Louisiana, 30 Sept - 3 Oct. 2001. Conference and Technical Exhibition, Amsterdam, June 2001.
8. Kretz, V., Le Ravalec-Dupin, M. and Roggero, F.: “An Integrated 22. Tarantola, A.: Inverse Problem Theory, Elsevier (1987).
Reservoir Characterization Study Matching Production Data and 23. Bissel, R.C.: “Calculating Optimal Parameters for History
4D Seismic,” paper SPE 77516, presented at the SPE Annual matching”, Proc. 4th European Conference on the Mathematics of
Technical Conference and Exhibition, San Antonio, Texas, 29 Oil Recovery, Røros, 7 – 10 June, 1994
Sept – 2 Oct. 2002. 24. Connoly, P.: “Elastic Impedance,” The Leading Edge (April
9. Lumley, D.E. and Behrens, R.A.: “Practical Engineering Issues 1999), 438 – 452.
of 4D Seismic Reservoir Monitoring,” paper SPE 38696, 25. Baina, R., Thierry, P., Calandra, H. and Devaux, V., 2001, 3-D
presented at the SPE Annual Technical Conference and preserved amplitude PsDM & AVA relevance, 71st Ann.
Exhibition, San Antonio, Texas, 5 – 8 October, 1997. Internat. Mtg: Soc. of Expl. Geophys., 300-303.
10. Pagano, T.A., Fanchi, J.R. and Davis, T.L.: “Integrated Flow 26. Barens, L., Rebert, T., Williamson, P., Rowbotham, P. and
Modelling: The Fusion of Geophysics and Reservoir Swaby, P., 2002, Joint geostatistical inversion of time-lapse data,
Engineering,” paper SPE 63137, presented at the SPE Annual Expanded abstracts of the 64th. meeting of the EAGE, Florence.
Technical Conference and Exhibition, Dallas, Texas, 1 – 4 27. "Eclipse Technical Description", 2003, Schlumberger GeoQuest,
October, 2000. 1060pp
11. O´Donovan, A.R., Smith, S.G. and Kristiansen, P.: “Foinhaven 28. . "SimOpt User Guide", 2003, Schlumberger GeoQuest, 312pp
4D Seismic – Dynamic Reservoir Parameters and Reservoir 29. Wang, Z, “Fundamentals of Seismic rock physics”, Geophysics,
Management,” paper SPE 63294, presented at the SPE Annual 66, 398-412, (2001).
Technical Conference and Exhibition, Dallas, Texas, 1 – 4 30. Batzle, M, Wang, Z, “Seismic Properties of pore fluids”,
October, 2000. Geophysics, 57, 1396-1408, (1992).
12. Fagervik, K., et al.: “History Matching of Reservoir Flow models 31. Brie, A., Pampuri, F., Marsala, A., “Shear Sonic Interpretation in
Using 4D Seismic,” EAGE 63rd Conference and Technical Gas Bearing Sands”, paper SPE 30595 presented at the Annual
Exhibition, Amsterdam, June 2001. Technical Confenerence and Exibition, Houston, October 1995.
13. Aggio, A. and Burns, S.: “Seismic Visualization for Dynamic and 32. Mavko, G., Dvorkin J., Mukerji J., “ The Rock Physics
Static Reservoir Characterization,” EAGE 63rd Conference and Handbook”, Cambridge University Press, (1998).
Technical Exhibition, Amsterdam, June 2001. 33. Doyen, P.M., Psaila, D.E., Astratti, D., Kvamme, L.B., and Al-
14. Hatchell, P., et al.: “Comparing Time-Lapse Seismic and Najar, N.F., 2000, Saturation mapping from 4-D seismic data in
Reservoir Model Predictions in Producting Oil and Gas Fields,” the Statfjord field, OTC 12100.
EAGE 64th Conference and Technical Exhibition, Florence, June 34. Tad M. Smith, Carl H. Sondergeldz, and Chandra S.
2002. Raiz,“Gassmann fluid substitutions: A tutorial”, Geophysics, Vol.
15. Landa, J. and Horne, R.N.: “A Procedure to Integrate Well Test 68, No. 2 (March-April 2003); p. 430–440.
Data, Reservoir Performance History and 4D Seismic
information Into a Reservoir Description,” paper SPE 38653,
presented at the SPE Annual Technical Conference and
Exhibition, San Antonio, Texas, 5 – 8 October, 1997.
16. Huang, X., Meister, L., and Workman, R.: “Reservoir
Characterization by Integration of Time-Lapse Seismic and
Production Data,” paper SPE 38695, presented at the SPE Annual
Technical Conference and Exhibition, San Antonio, Texas, 5 – 8
October, 1997.
17. Huang, X., Bentley, L.R. and Laflamme, C.: “Integration of
Production History and Time-Lapse Seismic Data Guided by
Seismic Attribute Zonation,” paper SPE 68819, presented at the
SPE Western Regional Meeting, Bakersfield, California, 26 – 30
March, 2001.
18. Waggoner, J.R., Cominelli A., and Seymour, R. H.: “Improved
Reservoir Modelling with Time-Lapse Seismic in a Gulf of
Mexico Gas Condensate Reservoir”, paper SPE 77514, presented
SPE 84464 15

Appendix A: Computer-Aided History matching Nomenclature

(CAHM) G = Jacobian or sensitivity matrix dr/dm
We remind that the main idea is to consider the history B = constraint matrix
matching procedure as a minimisation problem, as a parameter C = covariance matrix
estimation problem. Certain parameters in a reservoir H = (approximate) Hessian matrix
simulation model (vector m) are adjusted until the difference J = objective function
between the predicted results of the model (vector p(m)) and Jprod = production objective function
observed data (d) are sufficiently low (minimised). This Jseis = seismic objective function
minimisation requires the formulation of an objective L = left Cholesky factor of the correlation matrix R
function, which measures the mismatch. Traditionally, this W = weight matrix
objective function is chosen as a weighted least squares Wp = production term weight matrix
function Ws = seismic term weight matrix
1 Sw = water saturation
J ( m) = ( p(m) − d )T W ( p(m) − d ), (A-1) Sg = gas saturation
2 P = Pressure
where W is a weight matrix. Vp = P-velocity [m/s]
Focus is crucial on the following aspects: Vs = S-velocity [m/s]
data (d) and associated weight matrix (W), I = impedance [kg/(m2·s)]
C
I = simulated impedance
selection of parameters (m) – restriction of search space,
Ip = acoustic impedance
the optimisation algorithm. Is = shear impedance
The observations might need a pre-processing (averaging, de- VSh = volume of shale
noising, smoothing) and the weight takes into account the Vsand = volume of sand
uncertainty (measurement errors, possible correlations). We Kfr, Gfr = Frame moduli
believe that the relative weight given to some observations versus Ksh, Ksand = shale and sand bulk modulus
others should be completely covered by an appropriate estimates Km = mineral bulk modulus
of the covariance matrix (C=W-1). Kmix, = fluid bulk modulus
The parameter selection is based on derivative information (1st Ko, Kw, K g = phase bulk modulus
and 2nd order): d = data vector (generally or specifically production)
e = seismic derived data
∇J = ∇p ⋅ W ⋅ r (A-2) m = matching parameter vector
p(m) = simulated results (or specifically production)
J ≈ ∇ 2 J ≅ ∇p ⋅ W ⋅ ∇p (A-3) s(m) = simulated seismic derived results
Where r=p(m)-d is the residual vector. r = residual vector
The gradient wrt m gives an estimate of sensitivity of p to rs = vaporised oil
changes of m. The Gauss-Newton approximation of the Hessian rv = dissolved gas
gives the quadratic trend of J and the spectral decomposition ntg = net to gross
allows to define gradzones (correlated cell properties). This σ = production measurement error
approach needs a flow simulator calculating gradients. α, β = scalar weight to balance production and seismic
The optimisation algorithm is a gradient based technique in our αrock = generic rock properties
approach (like Levenberg-Marquadt). ρ = density
All these techniques and methods can be extended to additional
ρm = mineral density
terms in our objective function, provided we still are able to ρmix = fluid density
ρsh, ρsand = shale and sand density
compute ∇p , which means in our case, we can propagate
gradient computations through the PEM.