Discussion of Variogram

Principle of Variogram calculation

Interpretation of Variograms
Variomap calculation
Modeling of Variograms

1
 Variogram
Variogram overview

Variogram = statistical measure of geological heterogeneity

A variogram is a quantitative description of the variation in a property
as a function of separation distance between points
Based on the principle that two points close together are more likely to
have similar values than points far from each other
Two Main aspects to a variogram:
1. How similar are two values right next to each other?
2. How far apart do points have to be before they bear no relation to each other?

The variogram is used to model the way two values in space or time are correlated. Most
people intuitively know that two values in space that are close together tend to be more
similar than two values farther apart. Univariate statistics cannot take this into account. Two
distributions might have the same mean and variance, but differ in the way they are
correlated with each other.

Property modeling is normally used to describe the natural variation in a property. The
variogram should therefore describe this natural variation, rather than broad scale trends that
you see in your data. Identify any regional trends in data analysis before you begin the
variogram analysis.

When a discrete property is populated deterministically by using regular kriging (Indicator

Kriging), or stochastically by using either Sequential Indicator Simulation (SIS) or Facies
Transition Simulation, a variogram is needed as an input.

Similarly, a variogram is needed when a continuous property is populated deterministically by

using regular kriging (Kriging or Kriging by Gslib) or stochastically by using Sequential
Gaussian Simulation (SGS).

2
 Variogram
Introduction to Variogram

Measure variability with distance

Larger distances = larger variability

Calculated in 3 directions:
– Horizontal Major
– Horizontal Minor
– Vertical

Points = Sample variogram

Line = Model Variogram

The points that are posted in the Figure are referred to as an experimental or sample
variogram. These are computed from the data and each point shown represents a measure
of average variation at a given separation distance.

The line (shown in red) is the variogram model and is the result of a curve fitting exercise
based on the experimental variogram points.

The variogram analysis should always be performed on transformed data (transformation of

data will be explained when discussing the Data Analysis process).

3
 Variogram
Parameters

Variance
Average degree of difference between pairs
of points; A measure of how different
the members of a collection are from each other

Variance
Sill Sill
That Variance where the summary plot flattens
out to random similarity.

Range Nugget Range

Distance to the sill; That distance beyond which
data points no longer exhibit any statistical Separation
similarity 1 2 3 4 5 distance (lag)

Nugget
Degree of dissimilarity at zero distance: Sum of
geological microstructure and measurement and
experimental error

In summary: The closer two points are to each other, the more similar they are expected to
be (small variances for low separation distances). The nugget represents small scale
variation. This is often due to measurement error, which should not be taken into account
when establishing the variogram model. Small scale variation which should be taken into
account when modeling is, for example, lamination, which could cause rapid changes in
porosity over short distances.

In the Data Analysis process in Petrel, histogram columns will indicate the number of pairs on
the right-hand Y axis that define each variogram sample point and the semi-variance on the
left hand Y –axis.

4
 Variogram
Parameters

Variogram Parameters and their influence on the kriging/simulation result

Nugget Effect
Very sensitive; A high nugget gives random result
(even if the range is high!). Nugget due to
measurement error should not be modeled. High

Variance
nugget leads to averaging
Type Sill
Big differences are possible when a Gaussian
variogram is used
Range Nugget Range
Bigger range means more continuity
Separation
Variance + Sill 1 2 3 4 5 distance (lag)
Have no influence on the estimation result (only the
kriging variance is affected)

The Nugget effect must be modeled carefully. It is often that the pure nugget effect
is due to sparsely and/or poor quality, not due to lack of correlation.

When making a 3D model by using either kriging (explained in Chapter four:

Kriging) or simulation (explained in chapter five: Gaussian Simulation) the
algorithm need a variogram as an input. Changing the settings for the different
parameter gives different results. Examples can be find under the next two chapters
(Kriging and Gaussian Simulation).

5
 Variogram
Variogram Principle

A Variogram describes the variance between data

points as a function of the separation distance
between the data points Lag h.
Nk
1
γ (hk ) =
Nk
∑[ z( x ) − z( y )]
i =1
i i
2

hk ∈ [hMin , hMax ] (lag distance)

The Lag defines a distance range given by a

minimum and maximum distance. N is the number of
data pairs. The variance of a lag is the mean
variance of all data points being separated by this
lag.

A variogram is a tool to split the overall variance into a spatially related variance. The closer
the measurement points (borehole/log samples) are to each other, the less the variance.

6
 Variogram
Calculating an Experimental Variogram – general procedure
2

∑ ((Φ ) − (Φ ))
Nh
1
γ (h ) = (i + h ) i
2Nh i =1

1. The user defines a lag increment and search radius

1. Petrel will compare all points separated by the lag distance

2. For each pair, calculate the difference and then square it. => h-scatterplot (histogram bins in
Petrel)
3. Petrel will sum up and divide by twice the number of pairs (N)
4. And then do the same for other lag distances
5. Petrel plots the variogram value versus the lag distance for the user.

1. Define a lag increment (or spacing between any two points in the data), say 100
ft.
2. From the measurements, take all pairs of values separated 100 ft. apart.
3. For each pair, calculate the difference and then square it. => h-scatterplot
(histogram bins in Petrel)
4. Sum up all the differences and divide by twice the number of pairs (N). This
gives you the value of the variogram for that particular lag increment or distance
(one dot in the semivariogram above).
5. Do the same for other lag distances, say 200ft, 300 ft, 400 ft, and so on.
6. Plot out the variogram value versus the lag distance.
=> Experimental variogram (also called sample variogram or simply the
variogram)

Note: The given formula is the same as the previous slide, just divided by two. This gives a
Semi-Variance on the Y-axis rather than the original Variance.

7
 Variogram
Calculating an Experimental Variogram

Variance for 1 lag distance

φ1 N1 2

∑ ((φ ) − (φ ))
2
( φ2 - φ1 ) 1
φ2
)
+
γ (h1 ) = i +1 i
φ3 ) (φ3 - φ 2 )
2 2 N1 i =1
… +
φi
… h1: 1 lag distance
φi +1 ) ( φi++1 - φi )
2

… +
…

This slide illustrates a concrete example of how the experimental variogram value is
computed for a lag distance of one.

The variance for one lag distance [Gamma(1)] is equal to the sum of the squared differences
for all corresponding data pairs divided by two times the number of data pairs (NI).

8
 Variogram
Calculating an Experimental Variogram

Variance for 2 lag distance

φ1
( φ3 -+ φ1 )2
2
φ2)
φ ) (φ 4 - φ 2 ) γ (h 2 ) =
1 N2 2

∑ ((φ i + 2 ) − (φ i ))
…
3
2N2 i =1
+
φi … 2
φi +1 (φi+2 - φi )
φi+2 ) +
h2: 2 lag distance
… …

This slide illustrates how the experimental variogram value is computed for a lag distance of
two.

The same equation as shown in the previous slide is applied to data samples separated by
two lag distance units.

9
 Variogram
Calculating an Experimental Variogram - Exercise

The variogram can be calculated experimentally as:

2

∑ ((Φ ) − (Φ ))
Nh
1
γ (h ) = (i+h ) i
2Nh i =1

EXERCISE
Suppose that in a well we observe the string of porosity
Φ =3 values in depth steps of 1m: 3, 5, 7, 6, 4, 1, 1, 4
Φ =5 Calculate the variogram values for γ(h) 8

Φ =7 lags 1, 2, 3, and 4 metres Sill

7

Φ =6 respectively. 6

Φ =4 Plot the variogram. Is there 5

a pattern? γ(4)=7.124
4

Φ =1 3

Φ =1 γ(3)=7.1 2

Φ =4 γ(2)=5.75 1

γ(1)=2.214 0
0 1 2 Range
3 4 h
5

In this example the semi variance value (gamma) for 1 lag distance will be:
Gamma1 = 1 / (2* 7) * { (5-3)2 + (7-5) 2 + … } = 1/14 * {4 + 4 + 1 + 4 + 9 + 0 + 9} = 31/14 =
2.214

Variance value for 2 lag distance:

Gamma2 = 1 / (2*6) * { (7-3) 2 + (6-5) 2 + …} = 1/12 * {16 + 1 + 9 + 25 + 9 + 9} = 69/12 = 5.75

Variance value for 3 lag distance:

Gamma3 = 1 / (2*5) * { (6-3) 2 + (4-5) 2 + …} = 1/10 * {9 + 1 + 36 + 25 + 0} = 71/10 = 7.1

Variance value for 4 lag distance:

Gamma4 = 1 / (2*4) * { (4-3) 2 + (1-5) 2 + …} = 1/8 * {1 + 16 + 36 + 4} = 57/8 = 7.124

10
 Variogram
Calculating – Data search – The “search cone”

Because of irregular spacing of input points, a search area must be defined in the
search for points lying the distance range given by the Lag
North

Lag tolerance

Points within this area

are accepted as lying a
distance h from origin

Lag distance (h)

Angle
Angle tolerance
Band
width

East

Direction, angular tolerance, bandwidth, lag, and lag tolerances define lag “bins”. Data pairs
are identified based on a lag bin methodology. All data pairs to the same base lag contribute
to the experimental variogram value for that respective lag distance.

Bandwidth: A distance cutoff used to prevent the lag bin search area from becoming too
large (i.e., wide) at lag distances far from point of origin.

Angle Tolerance: It would be too restrictive to expect all pairs in a given direction to lie along
the exact line representing the selected direction. This tolerance provides some leeway so
that data pairs can be identified that approximate a given direction without being too
restrictive.

Lag tolerance: If the data is irregularly spaced (which it usually is!), there is no way you can
get pairs of values exactly 100 or 200 meter apart. What you can do, is you define a lag
tolerance, normally taken to be half the lag increment used. This ensures that you can get
enough pairs for a particular lag distance. The bigger your tolerance, the more pairs you can
define, and the smoother looking the variogram.
In short, we cannot afford to lose a data pair as most analysis on upscaled well log data is at
the absolute edge of having enough data to start with.

11
 Variogram
Variogram Model Types

Spherical: good general algorithm; Use this if you don’t have a good reason for
something else!
Exponential: Produces the most “noisy” result.
Gaussian: Will give the smoothest result.

Variogram Model Types.

When you want to fit the model variogram curve to the experimental variogram (sample
variogram) you need to define the model type (shape of variogram model curve).

In Petrel there are 3 options:

Spherical: simplest and has a linear behavior at shorter distances with a sharp transition to a
flat sill. Spherical variograms are most robust and stable in terms of the Kriging-equation
system to be solved. Influence of the data points are limited by the Range.
Exponential: Has a steep behavior at shorter distances with an asymptotic approach to the
sill at longer distances.
Gaussian: Reserved for phenomena that show high degree of continuity at short distances
and then begin to transition to more of an exponential behavior at longer distances. Rarely
used for porosity, avoided for discrete data types, and used with caution for permeability
when justified. Data points beyond Range have (some) influence on each grid node value.

The Exponential and Gaussian models reach their sill asymptotically. The `effective range', is
the distance where the variogram reaches 95% of its maximum.

Note: Exponential and Gaussian variograms using a zero nugget effect can cause problems
solving the kriging-equation system. A small nugget effect will help in this case.

12
 Variogram
Applied Variogram Modeling
1
1. First, calculate the experimental variogram

2. Fit a variogram model to the experimental variogram

spherical, gaussian, exponential

Interpretative process
should take geological knowledge into account
Vertical variogram model
usually plenty of data
easily estimated
Horizontal variogram
can often not be calculated due to limited amount of data 2
usually implied from geological knowledge
can be derived from correlated data source (e.g. variogram from
seismic when modeling porosity) or take from analog field or outcrop

Often well data is far too sparse to facilitate variogram modeling in the horizontal direction.
For example: what if you have a single exploration well; This situation supports only vertical
variogram analysis, no pairs are available in the horizontal plane!

In this case, it is common to turn to a correlated secondary source of data. Given a

reasonable correlation, one can justify the horizontal analysis on the secondary data. This
data is used as a proxy or substitute for the purpose of interpreting the direction and major
and minor range values.

Note: For carbonate reservoirs (with many horizontal production wells) the situation is often
vice versa and means a lot of data in horizontal direction and less data in vertical direction
due to many horizontal wells. Creating horizontal variogram will then be easier than creating
vertical variograms.

13
 Variogram
Applied Variogram Modeling

Variogram behavior example: Cyclic log porosity data with a trend

Cyclicity = Hole effect Trend in data

Search
radius:
130 m

Search radius

Varogram behaviour
A variogram shows normally larger variability with larger distances but sometimes it shows
cyclicity e.g. for fluvial systems.

A cyclic trend can be seen in the example zone porosity data. The cyclic “humps” of the
sample variogram is called Hole Effect and yields a repetition of samples caused by geologic
processes.

By looking at all facies within a zone, such Hole Effects can be seen and the search radius
will limit how much data is compared. The amount of lags are also important when searching
for cyclicity. It is more difficult to see this effect when you are using many lags.

14
Variogram
Applied Variogram Modeling

Trend should be removed before Example: A vertical trend

variogram modeling

The modeling should proceeds as follows if a

trend occurs:

Model trends
Model residual using variograms

Add trend and residual to obtain estimates

Example, fine upwards sequence gives

variograms without an apparent sill There is no apparent sill

From CV Deutsch, 2002

15
 Variogram
As a tool for data characterization, Comparison Histogram - Variogram
Log Log
very noisy – noisy - smooth very noisy – noisy - smooth

In general it is more meaningful to calculate a variogram on smoothed data.

16
 Variogram
Anisotropy – Graphic Example
In geostatistics, we define anisotropy as a characteristic of a set of data values. If there is a clear
difference in how data values change in one direction versus how they change in another direction,
then the data set is said to be anisotropic.

Example: The variability

of particulate size across
the channels will be
much higher than along
the channels.

If you suspect this kind of directional bias in your data source, you should incorporate that information
in the variogram to get a more accurate model

If a variable exhibits different ranges in different directions, then there is geometric

anisotropy. For example, in a shoreface deposit, permeability might have a larger range
along the shoreline compared to the range perpendicular to the shoreline.
If the variable exhibits different sills in different directions, then there is a zonal anisotropy.
For example, a variogram in a vertical wellbore typically shows a bigger sill than a variogram
in the horizontal direction.

17
Variogram
Anisotropy – Applied

Note that in this example, we have a major and minor direction of anisotropy, which does not affect the
searching, but results in an elliptical weight function because of a directional variability in the horizontal direction.
The system incorporates a weight in the major and minor directions as well as vertically.
The minor axis runs N/S and causes equidistant weights to be higher in the N/S direction than in the E/W
direction.
North

Elliptical
weighting Data points

Grid node to be
calculated

18
Variogram
Two methods for Determining Anisotropy

Assuming Anisotropy:
How do you determine if directional bias exists in your data?
How do you measure the direction of bias in the data?
There are essentially two ways to do it::
1. Trial and error with directional horizontal variograms.
2. Using the variogram surface (map), if available.

Y axis (North) ce
ran
tole 60
lag = 2 .0 g=
or : an
e ct
nv 5.0
e ctio h=
d ir id t
dw
ban
0

ce
=6

tan 6
dis La
g
a ng

lag
g5
La
g4 = 22.5
La
n ce
g3 er a
La tol
ang
g2
La X axis (East)
g1
La

19
 Variogram
Horizontal Variogram Generation in Petrel
Variogram Map Sample Variogram
Good for visualizing anisotropy and its Good for finding Major and Minor
direction. Range horizontally.

Both variogram maps and sample variograms can be made from the settings dialog of a
property object in Petrel.

20
 Variogram
Variogram Map – Computation

Y axis
Y Range
Number of
Y Lags

Number of X Lags
X Range

X axis

A variogram map is a way to present variograms that have been computed in several
different directions. This is done in an automated fashion (I.e., the various directions are
computed for you in one execution).

The center of such a variogram map represents 0.0 lag distance. Out from this center, the lag
distances will show an increase in several directions. Such maps are usually displayed as a
surface. The grid geometry of such surfaces is in +/- lag space and is not located in the
project coordinate area. For this reason, such maps are usually rendered along in a map
window and not displayed along with structure in a traditional project base-map display.

21
 Variogram
Variogram Map – Geometric Anisotropy

Variomap. The arrows show the

major (blue) and minor (red) Variogram for major and
direction of anisotropy minor direction

Different range for major and minor direction

Geometrical Anisotropy – Symmetric in both directions in Petrel. Theoretically: Sill = same,

Range = different)

22
Variogram
Geometric Anisotropy
Intermediate
direction
Direction of N
Max. continuity γ(h)

γ(h)

W E
Direction of
Min. continuity

γ(h)

The rose of ranges h

follows an ellipse S

23
 Variogram
Variogram Map – Zonal Anisotropy

Sample variogram in major and minor direction show different sill.

Zonal Anisotropy – One-directional (Theoretically Sill = different, Range =same). Zonal

anisotropy is handled like isotropic data because the range of the major and minor direction
is the same.

It is a limited case of geometric anisotropy. The range of correlation in one direction will
exceed the field size. This gives a variogram that does not appear to reach the sill/variance.

24
 Variogram
Sample Variogram – Computation from an object’s setting tab

To compute a sample variogram it is important to use a densely sampled

and correlated property.
In the object’s Variogram tab, define the orientation, search radius and
number of lags.

Sample variogram
The output is a sample variogram in the specified orientation. To generate two orientations
(Major and Minor direction) the Execution must be done twice in the respective directions.
Afterwards a model variogram must be matched with the sample variogram to create a best
fit model variogram curve (see next slide).

25
Variogram
Sample Variogram – Modeling

Fit a variogram model (e.g. Spherical model) to the sample variogram

Important model parameters:

Range
Sill • Model type
• Nugget
• Range
• Anisotropy: Azimuth; given
Nugget from Variogram map

26
Variogram
Variogram Model – Limitations

Nugget, sill and model type

MUST be the same for the
experimental variograms along
the major and minor anisotropy
axis.

Note:
Sill has no influence on
simulation/kriging estimation result

27
 Variogram
Other Variogram Types

Petrel offers in addition to the ‘classical’ variogram:

Pairwise relative
Logarithmic
Semimadogram
– Objective of these variograms is to help identify the range
Limitation
The variogram model should be based on the ‘classical’ variogram!
Sill and Nugget of all other variograms cannot be used by
kriging/simulation. Only the range can be derived from any
variogram type!

Other Variogram Types for point data and surfaces

Gives different description of spatial continuity. The objective is still to help identify the
Range. Each type affects the Range in different ways.

Classical Variogram – used as default in Petrel.

Pairwise relative – each pair is normalized by the square average.
Logarithmic – logarithmic values used instead of original values.
Semimadogram – uses the absolute difference instead of the squared difference.

28
Variogram
Other Variogram Types

Semimadogram
‘Classical’ Variogram

Logarithmic Variogram
Pairwise Relative Variogram

29
 Variogram Modeling in Petrel – 3 options:
From an object’s Settings tab - OR - In the Data Analysis process

- OR -
Simply by
entering the
Advantage: range, nugget
Option to create Variogram map and azimuth • More options
to find anisotropy directly into the • Can calculate variograms for all three directions
Property • Can easily see effect of search cone settings
Modeling dialogs • On up-scaled well logs, raw well logs or 3D property data

There are essentially three options for defining the variogram settings in Petrel:
1. You can do a horizontal variogram analysis from the Object’s settings dialog
(from the Variogram tab). Here, you may compute a variogram map and
compute sample variograms for the two horizontal directions.
2. A more robust way of creating a variogram is to use the Data Analysis process
in Petrel. This requires that the property to analyze must exist as a property
under the active 3D grid. You may do the analysis on raw well logs data, on up-
scaled well log data or on the entire 3D property. You should specify the
transformations of the property before doing the variogram analysis.
3. If for some reason you don’t want to compute the variogram (for example, if
you just want to run quickly through your model to see the effect of different
settings, or if you are not able to compute a proper variogram), then it is
possible to simply enter the variogram values (range, nugget, azimuth) directly
into the Facies Modeling or Petrophysical Modeling processes. There are some
default variogram settings already specified – the only reason for this is that
these variogram settings must be specified to run these processes. You should
never use the default settings unless you have verified that they are appropriate
for your field!

30
Variogram
Variogram Modeling Summary

Why variogram modeling?

Requirement for geostatistical algorithms
Kriging needs a smooth, continuous variogram that fulfills specific mathematical
properties. A variogram model that approximates the sample variogram as good as
possible fulfills these conditions.

Variograms are useful as data analysis tools

Determine layer thickness
Determine directions/degree of anisotropy

Used as QC to judge the quality of a model

31
Exercise

32