4Q16

Building geologically realistic facies models with the

new MPS method
Facies modeling with the new MPS method
Senior geologist Laetitia
Macé shows how to use the new MPS method for generating
facies models in property modeling workflows.

Laetitia Macé This NewsClick aims at introducing the new MPS method available since Petrel 2015.4. It provides a short summary of
Product Analyst the MPS theory, as well as the MPS workflow and the advantages of the new MPS method. Several use cases also
Petrel Modeling highlight the potential of MPS simulation in different geological environments.

Background theory of MPS

MultiPoint Statistics (MPS) simulation is a facies modeling technique that uses conceptual geological models as 3D training images to generate
geologically realistic reservoir models conditioned to well data.
It combines the ability to reproduce geological shapes, similarly to object-based methods, with the speed and exact data-conditioning of variogram-
based techniques. Thus MPS enables you to produce models with considerably more geological complexity and perfect data conditioning.
MPS extracts facies patterns from a conceptual training image that displays the type of facies elements believed to be present in the reservoir, and
then reproduces in the model the patterns that fit the reservoir well data. The training image is the main geological concept introduced in this
approach.
• The training image concept: The new MPS method is derived from the original Snesim algorithm, and is based on the use of search
mask, patterns and multi-grids concepts.
A training image is a 3D conceptual model of the reservoir geology. It aims at capturing the range of possible dimensions and shapes of
the facies elements believed to be present in the subsurface, as well as the spatial relationships among the different facies.
The difference between a training image and the MPS model is the conditioning to the well data. Training images are not conditioned to
any data, they are purely conceptual geological models that only provide geometrical information about facies elements. A MPS simulation
is the process that performs the conditioning by building models that honor well data.
Training images can be generated in many ways in Petrel:
– Simulating an unconditional facies model using any facies modeling methods such as object modeling, Truncated
Gaussian Simulation with trends...
– Drawing directly on your training image property with the interactive facies modeling tools
– Creating user-defined objects with the User-defined objects dialog, and then running an object modeling unconstrained
simulation.
You can estimate facies dimensions from well log data and good quality seismic. You can also obtain them from analogs (outcrop data)
and reservoir databases.
It is critical to correctly define inter-facies relationships, it can be inferred from core data analysis and geological interpretation.
• The MPS simulation workflow: For MPS simulation, the mandatory input data is a grid and a set of well log facies upscaled in the grid.

Page 1 of 9 Building geologically realistic facies models with the new MPS method
 • First, you will create a training image that displays the type of facies elements expected to be present in the subsurface.
• Second, you will define optional proportion and geometrical constraints to impose non-stationarity features (local variations of
facies element azimuth for example) in the reservoir, if needed.
• Finally, you will run the MPS simulation: it extracts facies patterns from the training image and reproduces the patterns that fit
the actual reservoir well facies data, while honoring any imposed proportion and geometrical constraints.

Fig 1. The MPS simulation workflow

Advantages of the new MPS method

• New intuitive user interface
Once the training image is built, there is only one unique step needed to create an MPS model. The new MPS method can be found in
the Facies modeling process.
• Targeted for geologists, not for MPS experts
The new MPS method is distance-based: it is using project units, in opposition to the MPS method with pattern objects which is using cell
numbers. This is essential to acknowledge when using different grids for the training image property and the simulation property as the
pattern will be reproduced using the real distance not the number of cells. Prior to running any MPS simulation, it is recommended to
check that the real size of the training image bodies are the ones you want to reproduce.
• Geologically realistic facies models
Non-stationary conceptual model can be reproduced using external local constraints:
• Proportion constraints to control the spatial distribution of the facies elements, that is, to decrease or increase locally the fraction
of a given facies.
• Geometrical constraints to modify locally the geometry of the facies patterns to be simulated.
• MPS realization result
• Honor the target fractions (within tolerance): The target fractions can be honored within 1% with the new MPS method
• Reproduce the TI patterns
• Fast performance

Page 2 of 9 Building geologically realistic facies models with the new MPS method
Handling non-stationary MPS model
MPS, as any geostatistical algorithm, generates stationary models by default. Therefore, unless dealing with a dense well environment, the facies
patterns borrowed from the training image can be reproduced anywhere in the MPS simulated model. Since actual reservoirs are not stationary,
facies proportions and geometries often vary in space. You must define external local constraints to impose non-stationary features in the MPS
model. There are two types of such constraints: proportion constraints to control the spatial distribution of the facies elements, that is, to decrease or
increase locally the fraction of a given facies; and geometrical constraints to modify locally the geometry of the facies patterns to be simulated. It is
possible to combine proportion, probability and geometrical constraints.
• Proportion constraints:
– The vertical proportion curves are 1D constraints that contain the target average facies proportions for each K layer of
the simulation grid. The vertical proportion of each facies code can be controlled vertically by a function. You can also
use Vertical proportion curves from the Data analysis process.
– The proportion maps are 2D constraints that contain the target average facies proportions for each IJ column in the
simulation grid. The proportion of each facies code can be controlled by a 2D surface with values between 0 and 1.
The surface is recommended to cover the entire modeled area.
– The proportion region properties are 3D continuous properties that contain the target average facies proportions for
different regions of the simulation grid. The proportion of each facies code can be controlled by a 3D continuous
property. Using proportions regions generates smooth, transitional region boundaries in realizations (whereas a
simulation zone by zone, or segment by segment creates sharp boundaries between regions). The proportion region
concept is to have a constant proportion defined for a given facies and for a given discrete region. It is not
recommended to use highly discontinuous proportion field as proportion region properties.
• Probability constraint: The probability properties are 3D continuous properties, considered as a soft constraint, that contain facies
probabilities in each cell of the simulation grid. The probability of each facies code can be given by a 3D continuous property. You can
also use Attribute probabilities from the Data analysis process. An additional parameter, the relative weight of the 3D probabilities (within
the range [0, 1]), is used to balance the soft and hard-data derived facies probabilities, according to the Tau-model mathematical
expression. The greater that weight, the greater the influence of the soft probability cube compared to the training image derived facies
probabilities. The default value is 0.5, which equals the influences. If the weight is 1, the facies probabilities inferred from the training image
are ignored. If the weight is 0, the soft facies probabilities have no effect.

Fig 2. MPS facies model (right) built using Multipoint simulation with probability constraints (bottom left) and braided channels training image (upper left).

• Geometrical constraints: The variable azimuth property allows constraining the local orientations of the simulated facies patterns. The
azimuths are provided in degrees, measured clockwise from the North (value of 0). Also the reference azimuth must be provided in the
training image: the local azimuth direction of each cell is then aligned according to this reference.

Page 3 of 9 Building geologically realistic facies models with the new MPS method
 Fig 3. MPS facies model (right) built with Multipoint simulation with geometrical (upper left) and probability (bottom left) constraints and a straight channels training image.

How much non-stationarity can be handled by the new MPS method?

The training image used in the new MPS method can be non-stationary. But how much non-stationary can be handled? To answer this question,
one has to recall the basis of the MPS simulation to understand the amount of non-stationarity that can be included in the training image.
At the beginning of the simulation, MPS scans the training image to extract the different facies patterns; then in a second stage MPS will reproduces
those patterns in the simulation property.
Thus if a given facies exhibits very different patterns either vertically and/or horizontally in the training image, then those patterns will be mixed in the
pattern tree. Thus, in order to better reproduce, it is recommended to add new facies codes to isolate the different patterns vertically and/or horizontally
and prevent them from being mixed with the other ones in the pattern tree.

Fig 4. Example of a non-stationary training image

Page 4 of 9 Building geologically realistic facies models with the new MPS method
How to use multiple training images with the new MPS method?

Region option is not enabled for the new MPS method. A workaround is possible to replicate the use of regions, see next slide.
Note though that the new MPS method allows to use non-stationary TI especially when combining with external constraints. Thus this remove the
need in most use cases to use multiple TIs and thus regions.
In the legacy MPS method of Petrel, there is an option to use a region property and for each code, it was possible to specify a different TI. It is also
possible to create similar MPS model using multiple TIs with the new MPS method. It cannot be done in one step though, one has to run as many
MPS simulation as there are TIs. So it is a bit more difficult to set up than in the legacy MPS but the performance will be the same.
The workflow to perform is as follow, let’s assume we have only N TIs to be used in N different regions:
1. Run MPS using TI1 only in Region1
2. Run MPS using TI2 for (Region1+Region2) with the main property used as additional hard data
3. Run MPS using TI_i for (sum of Region1 to Region_i) with the main property used as additional hard data
Note that it is not possible to use hierarchical facies option for the region property as only the cells within the specified region are taken into account
during simulation.

Fig 5. MPS model (right picture) built using two different training images (upper and middle left pictures) in sequence in two different discrete regions (lower left picture).

Best practices
• The dimensions of the training image grid are recommended to be at least twice as large as the dimensions of the largest facies elements.
The training image grid can be smaller than the simulation grid. The number of facies in a MPS simulation is recommended not to exceed
6 or 7. Generally, the training image should be as simple as possible. You should not create patterns that are too detailed and with too
many facies. The facies body thickness and width is recommended to be at least few cells. Consider reducing your grid cell size if this is
not the case.
• The number of cells in the training image grid is recommended not to exceed 1 or 2 million.
• The training image is recommended to have the same resolution (average cell size) as the simulation grid (only valid when using the
'Multipoint simulation' method) as the new MPS method uses project units, not cell numbers.
• The facies proportions in the training image are recommended to be within +/-10% to the target facies proportions in the MPS model.

Performance
Improving the overall workflow efficiency by generating fast MPS models allows testing quickly different input parameters. To increase the simulation
performance, perform one of the following:

Page 5 of 9 Building geologically realistic facies models with the new MPS method
 • Decrease the value of the 'Maximum number of data used' parameter (in the Expert tab) from its default value 40 down to a smaller value
(for example 20). This usually decreases the simulation time by a factor of 10 or more. (Only available when using the 'Multipoint simulation'
method).
• Alternatively, reduce the size (number of defined cells) of the training image property, by selecting a smaller training image zone.
In both cases, the drawback is a lower quality reproduction of the training patterns in the MPS model.

Data search parameters for MPS

In the new MPS method, an Auto-estimated option is available for the data search parameters, it does a simple analysis of the training image and
provides an estimate of the data search parameters, horizontally and vertically. When dealing with highly anisotropic bodies in the training image,
such as channels, it is advised to check the estimated data search parameters displayed in the Message log, and, if needed, it is possible to specify
those manually. Those parameters are comparable to the search mask parameters of the MPS method with pattern objects. Remember that the
new MPS is using project units, not cell numbers.

Impact of the distance-based behavior

When using a separate grid for the training image property, it is essential to understand the distance-based behavior of the new MPS method. When
running an MPS realization, the real size of the bodies in the training image will be reproduced in the simulation grid, thus the cell size of the TI grid
is a key parameter. The new MPS method does not reproduce the bodies in regards to the number of cells (or pixels) of the TI grid, it reproduces the
bodies in term of the size of the bodies present in the TI.
It is thus possible to use a training image grid independent from the simulation grid, and to create a library of training grid that captures adequately
the real relationship, size and geometries of the geological bodies present in a given environment.
In case, one wants to reproduce easily similar results created with the method “Multipoint simulation with pattern objects”, the training image grid is
recommended to have similar average cell size as the simulation grid.

In the below example, the simulation grid has constant cell size (50m * 50m * 50m) while the training image grid has varying cell size: picture A on
the left shows a TI grid with cell size of (50m * 50m * 50m) while pictures B and C in the middle and on the right show a TI grid with cell size of (25m
* 25m * 25m).
In case A, both grids have identical cell size, thus the bodies are reproduced in the simulation grid with similar size (diameter of 500m) and number
of cells as in the TI grid.
In case B, the cell size of the TI grid is half the cell size of the simulation grid and the bodies are also half the size compared to TI grid from case A
diameter of 250m). When running MPS, the bodies are reproduced in the simulation grid with similar size as in the TI grid, but not the same number
of cells.
In case C, the cell size of the TI grid is half the cell size of the simulation grid but the bodies have the same size as in the TI grid from case A (diameter
of 500m). When running MPS, the bodies are reproduced in the simulation grid with similar size, but not the same number of cells.

Page 6 of 9 Building geologically realistic facies models with the new MPS method
Fig 6. Impact of the cell size of the training image grid during the MPS simulation

Fig 7. Detailed view of Fig 6. Illustrating the impact of the cell size

Page 7 of 9 Building geologically realistic facies models with the new MPS method
Examples of MPS simulation results
The geological environment of a carbonate ramp is used to illustrate the results that can be achieved with the new MPS method. Simple examples
are provided, all based on synthetic data and using a non-stationary training image. Proportion constraints are used to guide the reproduction of
specific facies in preferential location in the field (Fig 8. and Fig 9. ). Using such constraints in the below unconditional simulation highlights the impact
of using secondary data to constraints MPS.

Fig 8. Use case of a carbonate ramp: unconditional MPS model using proportion constraints from scenario 1

Fig 9. Use case of a carbonate ramp: unconditional MPS model using proportion constraints from scenario 2

Page 8 of 9 Building geologically realistic facies models with the new MPS method
In Fig 10. , two conditional simulations are run using only few wells. Note that the vertical relationship of the different facies is consistent with the one
observed in the training image, thus all data used are consistent. Without using any additional constraints, the MPS simulation replicates in
satisfactory manner the training image pattern and produces a geologically realistic facies model

Fig 10. Use case of a carbonate ramp: conditional MPS model with two input data scenarios

Conclusion
MPS is a very powerful algorithm that can reproduce any type of geological environment. The new proposed MPS method in Petrel has key attractive
advantages: ease of use and accessibility to non-MPS experts. Using non-stationary training combined with secondary constraints, it is now quick
and easy to build geologically realistic facies models.

Page 9 of 9 Building geologically realistic facies models with the new MPS method