STOCHASTIC SIMULATION

And
RESERVOIR MODELING WORKFLOW

C&PE 940, 21 October 2005

Geoff Bohling
Assistant Scientist
Kansas Geological Survey
geoff@kgs.ku.edu
864-2093

Overheads and other resources available at:

http://people.ku.edu/~gbohling/cpe940

1
Stochastic Simulation

Stochastic simulation is a means for generating multiple

equiprobable realizations of the property in question, rather than
simply estimating the mean. Essentially, we are adding back in
some noise to undo the smoothing effect of kriging. This possibly
gives a better representation of the natural variability of the
property in question and gives us a means for quantifying our
uncertainty regarding what’s really down there.

The two most commonly used forms of simulation for reservoir

modeling applications are sequential Gaussian simulation for
continuous variables like porosity and sequential indicator
simulation for categorical variables like facies.

The basic idea of sequential Gaussian simulation (SGS) is very

simple. Recall that kriging gives us an estimate of both the mean
and standard deviation of the variable at each grid node, meaning
we can represent the variable at each grid node as a random
variable following a normal (Gaussian) distribution. Rather than
chooses the mean as the estimate at each node, SGS chooses a
random deviate from this normal distribution, selected according to
a uniform random number representing the probability level.

2
For our six-point example in the porosity data, ordinary kriging
gave a mean estimate of 12.93% with a standard deviation of
0.49%. In this case, if we happened to generate a uniform random
number of p = 0.665 for this grid node, then the assigned porosity
would be 13.14%, the corresponding value of the cumulative
normal probability function:

3
So, the basic steps in the SGS process are:

• Generate a random path through the grid nodes

• Visit the first node along the path and use kriging to estimate
a mean and standard deviation for the variable at that node
based on surrounding data values

• Select a value at random from the corresponding normal

distribution and set the variable value at that node to that
number

• Visit each successive node in the random path and repeat the
process, including previously simulated nodes as data values
in the kriging process

We use a random path to avoid artifacts induced by walking

through the grid in a regular fashion. We include previously
simulated grid nodes as “data” in order to preserve the proper
covariance structure between the simulated values.

Sometimes SGS is implemented in a “multigrid” fashion, first

simulating on a coarse grid (a subset of the fine grid – maybe every
10th grid node) and then on the finer grid (maybe with an
intermediate step or two) in order to reproduce large-scale
semivariogram structures. Without this the “screening” effect of
kriging quickly takes over as the simulation progresses and nodes
get filled in, so that most nodes are conditioned only on nearby
values, so that small-scale structure is reproduced better than large-
scale structure.

4
For SGS, it is important that the data actually follow a Gaussian
distribution. If they do not, we can use a normal score transform:

5
Here are six sequential Gaussian simulations of our porosity data,
using the spherical semivariogram model and a 16 nearest-
neighbor search:

6
Sequential indicator simulation (SIS) is very similar to sequential
Gaussian simulation, expect that indicator kriging is used to build
up a discrete cumulative density function for the individual
categories at each case and the node is assigned a category selected
at random from this discrete CDF.

Very briefly, an indicator representation for a categorical variable

such as facies would be formulated as

1 if facies k is present at uα
i(uα ; k ) = 
0 otherwise

where you would have one indicator variable for each of the K
different facies. We can then use kriging (based on indicator
semivariograms) to produce a set of facies membership
probabilities at each grid point, build up a CDF from the
probabilities, and select a facies at random from the CDF:

7
For a continuous variable such as permeability, indicator variables
are built by comparing data values to a set of thresholds, zk :

1 if z (uα ) ≤ z k
i(uα ; k ) = 
0 otherwise

We might define thresholds, for example, at the 10th, 25th, 50th,

75th, and 90th percentiles of the data distribution, for example. In
this case, kriging the indicator values for the k th threshold, zk , gives
estimates of P[Z (u ) ≤ z k ] at each estimation point. Since this
already is a cumulative probability, we don’t need to go through
the process of summing to get a CDF, although we will need to
correct any violations of the expected order relationships,
P[Z (u ) ≤ z k ] ≤ P[Z (u ) ≤ z k +1 ], that happen to occur.

In this case, SIS assigns each node to a corresponding range (e.g.,

upper 10%) by random selection from the CDF and the resulting
indicator vector gets added to the conditioning data for the
remaining nodes.

8
Typical Reservoir Modeling Workflow

Basically, work from large-scale structure to small-scale structure,

and generally from more deterministic methods to more stochastic
methods:

• Establish large-scale geologic structure, for example, by

deterministic interpolation of formation tops; this creates a
sete of distinct zones

• Within each zone, use SIS or some other discrete simulation

technique (such as object-based simulation) to generate
realizations of the facies distribution – the primary control on
the porosity & permeability distributions

• Within each facies, use SGS (or similar) to generate porosity

distirubtion and then simulate permeability distribution
conditional to porosity distribution, assuming there is some
relationship between the two

Porosity and facies simulations could be conditioned to other

secondary data, such as seismic. Methods also exist for
conditioning to well test and production data, but these are fairly
elaborate and probably not in very common use as yet. More
typical (maybe) to run flow simulations after the fact and rank
realizations by comparison to historical production & well tests.

For more on reservoir modeling workflow and assessment of

uncertainty from multiple realizations, see

C. V. Deutsch, 2002, Geostatistical Reservoir Modeling, Oxford

Univeristy Press, 376 pp.

9
Dealing With Trend: Zone A Thickness Data

Here are the Zone A thickness values measured in the same 85

wells as our example porosity data:

The SW-NE trend affects the omnidirectional semivariogram:

10
Approach 1: Fit a global linear trend and krige the residuals

11
12
Approach 2: Fit variogram in trend-free direction and use kriging
with first-order trend in X & Y

Here are the directional variograms (directions are azimuths from

north) computed with a directional tolerance of 22.5°:

The model shown is that fitted to the semivariogram for N 135° E,

which seems to be reasonably trend-free. The model is
exponential with a sill of 6.1 m2 and a range of 5292 m.

13
We then krige using a first-order trend model in X & Y and the
presumably trend-free semivariogram model for N 135° E:

14
Accounting for Porosity-Permeability Correlation

Permeability data are available from 42 of the 85 Zone A wells:

There is a fairly strong correlation between the log(perms) and

porosity at the 42 wells:

15
We want to account for the observed permeability-porosity
relationship for two reasons: 1) to preserve this relationship in our
modeled permeability and porosity distributions, and 2) to take
advantage of the more abundant porosity data in our estimation of
the permeability. To do this, we will work with the LogPerm
residuals (actual LogPerms minus those predicted from the
porosity) and add the kriged or simulated residuals back into a
mean LogPerm grid predicted from the porosity grid.

The LogPerm residuals at the wells seem reasonably normally

distributed:

16
and may or may not show a bit of trend:

The semivariogram model is exponential with a sill of 0.026 and a

range of 11800 m. The fitted sill matches the global variance quite
closely, so I am going to treat the residuals as trend-free.

17
Here are six sequential Gaussian simulations of the LogPerm
residuals based on simple kriging with a 16 nearest-neighbor
search:

18
And here are those six simulations of the LogPerm itself based on
adding the simulated residuals to the LogPerm values predicted on
the basis of the six porosity simulations we developed before,
using the regression equation developed on the well data:

19
Here are the crossplots of simulated LogPerm versus simulated
porosity, along with the regression line developed from the well
data. The simulated values show a somewhat higher correlation
overall (0.79) than seen in the original well data (0.72), probably
due to the small positive correlation (0.19) between the simulated
LogPerm residuals and simulated porosities.

20