Using Ensemble-Based Methods for More Reliable Reservoir

Predictions: Applications on Fractured Reservoirs

How Emerson Process Management uses ensemble statistics and its Roxar reservoir modeling
and reservoir engineering software to estimate the future state of a reservoir and capture the
associated uncertainty to support decision-making.

By Simon Fillacier, Jarle Frette, Richard Hammersley and Arne Skorstad, Roxar Software Solutions, a business
unit of Emerson Process Management.
Large companies traditionally rely more intensively on proba-
bilistic analysis than smaller ones that rely more on deter-
ABSTRACT
ministic methodologies. However, with the improvement in
computational capabilities versus costs, even mid-size to small
Predicting oil and gas field behavior is a difficult yet companies can start affording more robust probabilistic analy-
crucial task for multidisciplinary teams that consist of sis tools.
geophysicists, geologists, reservoir engineers and
The traditional approach to predicting oil & gas field behavior
senior decision-makers. Furthermore, the integration and updating reservoir models relies on a small number of
of different types of data (not only production and well scenarios (base, high and low cases) and introduces determin-
data but also seismic data) in a flexible and updatable istic steps that do not always fit with modern reservoir man-
workflow is central to reliable predictions. To this end, agement guidelines. It also means the lack of an automated
workflow making model updates time consuming.
the reservoir model remains the cornerstone on which
crucial field development decisions are made and a key Modern, ensemble-based approaches, on the other hand,
tool for quantifying uncertainties, analyzing results and automate the workflow and lead to easier model updates.
integrating the work of different disciplines. Percentiles can be derived from the ensemble members and
their validity discussed in light of experience gained and the
reliability of the probabilistic assessments. For any reservoir
This paper will look at the move towards probabilistic simulator output such as production or pressure, a range of
analysis tools and the better quantification of risk; values associated with a measure of confidence can then be
innovations in history matching and the use of proxies; calculated (see Figure 1).
and how ensembles can generate useful reservoir
management information. The paper will also look at
closing the loop between reservoir modeling and
simulation as well as the benefits of an ensemble-based
workflow. An example of how a fracture network can
be calibrated with dynamic data will also be included.

Introduction: Predicting the Future

Predicting the future and generating information on future oil

& gas field behavior is key to present and future reservoir man- Figure 1: Possible production values with attached confidence interval.
agement decision-making. Technological developments, such
as Emerson’s reservoir modeling and engineering solutions, are To effectively manage oil and gas related risks it is essential to
now ready to support robust, statistically sound methods such have uncertainty information. The evaluation of risk means
as ensemble-based approaches. An ensemble-based method having a set of possibilities each with quantified probabilities
uses results from a large set of models to analyze the uncer- and quantified losses / gains.
tainty in the predicted values.

About the Authors: Simon Fillacier is Technical Product Manager in the Reservoir Engineering domain, Jarle Frette is Senior Product
Specialist in the Reservoir Engineering Domain, Richard Hammersley is a Principal Developer of Tempest ENABLE, and Arne Skorstad is
Technical Product Manager in the Uncertainty domain. Together, they represent decades of experience in the industry and are continuing
to push the boundaries of reservoir modeling and uncertainty management.

Roxar, a business unit of global Fortune 500 company Emerson, is a leader in reservoir management software. Their seismic to simulation
solution helps operators make informed decisions across the prospect lifecycle on where to drill, what production strategies to adopt,
and how to maximize oil & gas recovery from their assets.
2 | Uncertainty Management

The strategy and tactical questions that are to be addressed For example if we observe the cumulative produced volumes
during a field’s life all require a risk assessment: For example, of fields after 10 years of production, what is the likelihood
should the company apply for this licence? Should more data that the production was driven by, for instance, the aquifer’s
be acquired? What kind of production strategy will maximize strength?
the recovery and optimize the field’s economics? Should we
drill one risky horizontal well or three slanted ones? And so on. Estimating the likelihood of a model, given the observations,
Figure 2 illustrates a quantifying of such risk with percentiles is carried out using a Bayesian framework that updates the
displayed. prior knowledge (see Equation 1) of the reservoir. As a result,
the models assimilate the available data, and the likelihood
- the probability of the observations taking into account the
parameters - is used to calculate distributions describing the
uncertain parameters.

Posterior ~ Prior * Likelihood

P(Parameters|Observations) ~
P(Parameters) *P(Observations|Parameters)

Equation 1: Bayes’ law relating the posterior and prior distributions using the
likelihood function.

Many history matching methods based on Equation 1 and

ensembles of runs exist. This paper, however, will focus just on
the Proxy approach.
Figure 2: Possible field gas cumulative production with associated levels of
confidence.
The Proxy Approach
Better History Matching Proxies can be used to estimate simulator response as inputs
are changed. This means that complex reservoir models are
The oil & gas industry has now reached a stage where a large emulated via simpler representations leading to more cost-
fraction of producing fields can provide several years of
effective evaluations.
operational and production history. This includes not only
production data (well data) but also seismic data that should
be shared through the same models. Models that capture the
reservoir uncertainties and integrate the acquired data will
build more accurate predictions of the future.

Figure 3: A typical ensemble-based history matching workflow

Through an ensemble-based history matching workflow

(Figure 3), uncertain inputs are identified and characterized in
order to parameterize the reservoir model and give reservoir
engineers the chance to constrain those parameters with the
observed “real” measurements. How to estimate the probabil-
ity of a hypothesis? That is the challenge reservoir engineers Figure 4: Accelerated history match using proxy technology
face in their daily work.
3 | Uncertainty Management

Figure 4 illustrates the concept of a proxy model with one strongly relies on the engineer’s knowledge of the
parameter and one observed value. The plot shows only one field, and its’ governing behavior.
uncertain parameter on the horizontal axis and the associated • Finally, reservoir engineers can examine the sensitivity of
response on the vertical axis. The proxy is a combination of a results using Tornado Charts (illustrated in the example in
response surface (red line) interpolating the simulation result Figure 11).
(dots) and some Bayes Linear Estimation techniques applied on
it to quantify the uncertainties between simulator responses
(thin blue lines). Ensemble-Based Workflows for Reliable
Predictions
The reservoir engineer wants to use each measurement (or
observation) to weight the simulation runs with a proxy similar One way of predicting the field’s behavior and estimating its
to the one built. In Figure 4, an ensemble of runs is created by uncertainty is to generate a sample of a reasonable number of
sampling the prior uniform distribution in blue on the uncer- runs (or realizations) that cover the range of possible updated
tain parameter axis. The corresponding simulator response models.
is illustrated by the blue dots, also called scoping runs. Then,
for the illustrated observation, the proxy will be improved, Often, P10, P90 and P50 are used to cover the space of uncer-
integrating the measurement (and its associated error) and tainty associated with models. But let us take a step back and
estimating the likelihood. Figure 4 show that there are more think carefully. Is there such a thing as a P50 run?
simulation results (green dots) closer to the best history
match, i.e. where the simulator gives the closest response to Statistically, it is not correct to select low/medium/high rep-
the observation. This type of run, which improves the proxy is resentative runs, but there may be runs that match the users’
called refinement. criteria for low/medium/high runs in the ensemble of runs.
Using an ensemble of realizations clearly shows that a realiza-
Once a good agreement between the proxy and the simulator tion that is a P50 at a given time might well be a P90 earlier in
response has been found, the uncertainty parameters’ poste- time (Figure 6 illustrates this statement). Using ensembles can
rior distribution (in green on Figure 4) is sampled using Markov help answer this question: How reliable is my P50 case through
Chain Monte Carlo techniques. time? Selecting P10, P50 or P90 from an ensemble differs from
selecting three runs representing P10, P50, and P90.

Figure 5 : Prior and posterior ensembles

Figure 5 shows the results of the history matching process

using proxies: The blue ensemble sampled from the prior
distributions illustrates the initial state of knowledge of the
reservoir. The green ensemble sampled from the posterior,
updated parameter distribution shows how the knowledge of
Figure 6 : Realistic – envolving with time – percentiles (top) vs Idealistic
the system has been increased. ranked percentiles (bottom).

Using proxies for history matching brings significant benefits.

These include:
Presenting Ensemble-Based Results

• The number of simulation runs is reduced compared to The display of uncertain data in space and time is something
traditional AHM methods. others industries have developed and managed to make un-
• The parameterization can be easily changed. derstandable. It’s up to the oil & gas sector to do the same.
• Different and new measurement points can be added dur-
ing the history matching process. This is where engineer-
ing judgement matters. The proxy-based approach
4| Uncertainty Management

Using a posterior ensemble for subsurface modeling provides a interval at any time of the time series, but added value is
quantitative assessment of the uncertainty associated with the gained by displaying statistics in a map view. Standard de-
predictions (Figure 7). viation, mean, and percentiles maps are very useful tools to
address strategy and tactical questions across time. Emerson’s
reservoir modeling software, Roxar RMS offers the possibil-
ity to calculate and display statistics for multi-realizations, as
illustrated in Figure 8.

Building Integrated Workflows for Flexible

and Efficient Data Assimilation

The role of geophysicists, geologists and reservoir engineers is

often very compartmentalized, with uncertainty captured at
each stage of the modeling chain. Instead of flowing through
the workflow, the uncertainty information is lost when the
geophysicist delivers his “base case” interpretation to the
geologist, who will eventually deliver a limited number of volu-
Figure 7 : Possible field gas cumulative production with associated levels of metric scenarios to the reservoir engineer.
confidence displayed using distributions showing the probability of a run given
some prior uncertainties and some observed data, in Emerson’s history match-
ing software, Roxar Tempest ENABLE. Relying on a common vision and consistent representation
through shared models ensures efficient cross-disciplinary
The result is that spatial uncertainty maps (for example, of collaboration. This leads to a closing of the loop and greater
moveable oil – see figure 8) can be calculated and the predic- integration between reservoir modeling and simulation work-
tion uncertainty also calculated for any simulator output at any flows with ensemble-based methods playing a crucial role in
time. Typically seen as 2D graphs, the challenge still remains in achieving this.
regard to 3D models and the mapping of data.

Ensemble workflows typically require hundreds of members in

Closing the Loop with Roxar Tempest and
order to estimate the probability density of the models. Roxar RMS

The ability to deliver a comparative display of multiple maps Traditionally, Emerson’s reservoir modeling software, Roxar
to select one instance in particular for example to identify an RMS has been used to create, edit and manage the static data
outlier - is fundamental in regards to quality control. required for reservoir simulation. It can on top of that, how-
ever, be a fundamental part of any assisted history matching
Ensemble predictions contain a huge amount of information as workflow using any reservoir simulator, thereby maintaining
can be seen in the oil saturation map output from the individu- the consistency between the geology and the dynamic data.
al ensemble members on Figure 8. Roxar RMS can also be used as a shared subsurface model that
is continuously updated by new knowledge acquired through
the field’s life.

Including Roxar RMS and its powerful geological modeling fea-

tures within the reservoir simulation and the ensemble-based
workflow allows the information from the dynamic data to be
used to constrain the geology.

Sensitivities through proxy-based history matching show that

an important part of the uncertainty is contained within the
structural framework. To this end, Roxar RMS can capture this
uncertainty for a better history match and better predictions.

Example on Fracture Reservoirs

One example of closing the loop using the ensemble-based

workflow is in calibrating a fracture network with dynamic
Figure 8 : A map showing the probability of finding moveable oil superimposed
data. This is possible with the latest version of Roxar RMS -
with an average pressure map, in Roxar RMS.
which can operate alongside Emerson’s history matching and
uncertainty estimation software, Roxar Tempest ENABLE and
How we analyse this information is key. We can plot how a well
its simulation engine Roxar Tempest MORE.
will perform with time and give an estimate of the confidence
5| Uncertainty Management

Figure 10: A view from the top of the fault position and the three fracture sets.
The fault related set (yellow) is attached to the uncertain fault position.

In addition, running a sensitivity study using Roxar RMS and

Roxar Tempest ENABLE together allows the reservoir engineer
to quickly establish which parameter influences simulation
results (such as cumulative produced volumes) the most.
In order to run such appraisal workflows, we recommend gen-
erating an ensemble of runs sampled from the prior distribu-
tions describing the uncertainly of the reservoir.

Figure 9: A 3D view of the model. In the upper part the arrows show how the
fault position could vary. In the lower part the corresponding fault envelopes
are displayed.

Fractured reservoirs have become increasingly important. As

fields mature, fractures will affect not only individual well pro-
duction, but also sweep and total recovery rates, making them
a key element in maximizing recovery and planning second-
ary and enhanced recovery programs. Fractures result from
the deformation of the reservoir meaning that the structural
uncertainty should be captured and represented as accurately
as possible.

Since there is usually a high degree of uncertainty on many

structural properties, the ability to interactively modify and Figure 11: A Tornado chart showing the effect of the different uncertain param-
refine properties is important not only in generating a robust eters on the field cumulative oil production uncertainty.
initial model but also in keeping that model consistent with the
changing knowledge and data. Coming back to our example, Figure 11 shows that in such a
case, the fracture apertures of the homogeneous fracture sets
Uncertain parameters for fractured reservoirs may include play the largest role when explaining the uncertainty in oil pro-
fracture aperture, fracture length, orientation and density. duction. It also shows that the fault F1’s position is an uncer-
Some fracture sets are also fault-related (see Figure 10), so it tain parameter that contributes to the uncertainty attached to
is important to be able to capture the fault position and shape the cumulative oil production of the well. Capturing the fault
uncertainty as well as the intrinsic fracture uncertainty. In this position uncertainty and the associated effects on the fracture
example, the faults’ positions are also uncertain parameters. network is a unique ability offered by Emerson’s solutions.

Again, the ability of Roxar RMS to capture and manage struc- Calibrating a Fracture Network by History
tural uncertainty helps associating fault sets with fault position
uncertainty within a smooth workflow.
Matching
History matching with Tempest ENABLE produces a proxy
model. From this we can generate a statistically valid Pos-
terior ensemble. Figure 12 shows the Posterior ensemble in
blue together with the Prior ensemble used for the appraisal
workflow.
6| Uncertainty Management

Figure 14: Prediction ensemble and attached reliability assessment.

Conclusions
Figure 12: Prior and Posterior ensembles: The history matching process pro-
duces an ensemble of history matched runs that can be used for predictions Whether as input to field development and operational plans,
investment proposals, or future divestments, the ability to
During the history matching process, the knowledge of the accurately generate future production estimates, quantify
field has been updated. For instance, Figure 13 shows how the uncertainty, and minimize financial risk is one of the industry’s
observations have constrained the probability posterior dis- greatest challenges.
tribution of two parameters. To achieve a satisfactory history
match, the F1 fault should be moved to the East and the most Through an integrated reservoir modeling and reservoir engi-
likely value of the fracture aperture for the uniform fracture neering workflow and the ensemble statistics-based approach,
set 1 is 110 µm, meaning that this fracture set should show a operators can achieve a greater quantification of risk and gen-
smaller aperture than initially thought. erate crucial information to support decision-making.

Our prior, incomplete, knowledge has been updated with new The example described in this paper has been run us-
information coming from the production data, and since the ing the latest version of Roxar RMS and Roxar Tem-
history match is acceptable, we can proceed to predictions,
using posterior parameter distributions to sample uncertain pest. The history matching and uncertainty estimation
parameters values. workflow is run in the Tempest ENABLE module and
the simulation runs are performed in Tempest MORE.

Further Reading
Slotte P.A., and Smørgrav, E., 2008; ‘Response Surface Meth-
odology Approach for History Matching and Uncertainty As-
sessment of Reservoir Simulation Models’; SPE Paper113390.

Fillacier S., Fincham A. E., Hammersley R. P., Heritage J. R., Kol-

Figure 13: Parameter posterior distributions: the knowledge of the reservoir has
bikova I., Peacock G. and Soloviev V .Y.; ‘Calculating Prediction
been increased and the uncertainty on the parameters decreased
Uncertainty using Posterior Ensembles Generated from Proxy
A prediction ensemble, typically of 100 members, is used to Models’; SPE Paper 171237-MS.
derive statistics such as prediction percentiles for any result
variable at any time, as illustrated in Figure 14. The authors would also like to thank Statoil for their
work in sponsoring the development of the ensemble
workflow.

To find out more contact Roxar Software Solutions.

Email: [email protected]

ROXAR AS, GAMLE FORUSVEI 17, PO BOX 112, 4065 STAVANGER, NORWAY TELEPHONE +47 51 81 8800 FAX +47 51 81 8801 WWW.ROXARSOFTWARE.COM