Rolling the Dice

R. E. (Gene) Ballay, PhD

WWW.GeoNeurale.Com
The only certainty in most of our formation evaluations is the presence of uncertainty and
how that issue is (or is not) addressed. At the simplest level one may estimate the Best and
Worst Case, for each input attribute, and then bound the evaluation with the resulting extreme
values, even as we recognize that the simultaneous occurrence of multiple “best” or “worst”
values is an unlikely event.
It is in fact relatively simple to address the uncertainty question in a comprehensive, realistic
and quantitative fashion, and to further identify where to focus time, and money, in search of
an improved evaluation.
At the simplest level our Sw Figure 1
estimates are compromised
by uncertainty in the various
Archie equation attributes.
Sw n = a Rw / (Φ m Rt)
In an earlier article (Risky •Monte Carlo simulation of Sw(Archie) for Monte Carlo Distribution

Business) we took the above user-specified distributions 600

Sw

derivative of Archie’s equation •“a”, Rw and Rt are assumed to be ‘well 500

known’, but may be varied by a straight- 400

Frequency
(the same approach will forward extension of the technique 300

suffice for a shaly sand •With these specifications, there is a 95% 200
probability that Sw is bounded by 0.357 +/–
equation), and calculated the 2(0.0382) Î 0.28 < Sw < 0.43
100

0
individual impact of each •The Best / Worst case would have 0.24 < Sw 0.00 0.10 0.20 0.30 0.40 0.50
< 0.50, the difference being the unlikely
term’s uncertainty upon Sw to chance of all Best (or Worst) events occurring
Sw

identify where the biggest simultaneously 2000 Simulations

bang for the buck, in terms of

a core analyses program or suite of potential logs, was to be found. At that time we noted the
‘link between parameters, in that the relative importance of a single attribute, can be
dependent upon the magnitude of another attribute, so that the characterization must be
done for locally specific conditions.
An alternative approach is Monte Carlo simulation, which can be implemented with routine
Excel spreadsheet functions. The Monte Carlo method randomly assigns values, according to
user specified probability distributions, to each of the input parameters and then calculates the
result. When the simulation is repeated a statistically significant number of times (results
herein are based upon 2000 passes, which Excel handles without a problem), one is able to
determine the likely outcome within any specific probability band, and to further identify
which parameter is dominating the uncertainty (and hence where time and money is most
efficiently directed for an improved result).
As an example, with the specifications tabulated in Figure 1, there is a 95% probability (+/‐ two
standard deviations) that 0.28 < Sw < 0.43, whereas the Best / Worst approach would bound
the results with 0.24 < Sw < 0.50; the difference being the unlikely event of multiple,

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

simultaneous Best or Worst events. Not only does Monte Carlo give us a more realistic
summary, but by varying the input standard deviations (uncertainties), one is able to identify
where to most efficiently concentrate time / money in an effort to improve results.
Monte Carlo Technique
The Monte Carlo method relies on repeated random sampling of user specified input
probability distributions to model expected results. This approach is attractive when it is
infeasible or impossible to compute an exact result with a deterministic algorithm.
An advantage of Monte Carlo is that any type of distribution can be used to characterize the
uncertainty specification of input parameters, for example normal, log normal, etc; an issue
since the phenomena governing frequency distributions in nature often favor log‐normal
(Limpert et al, 2001).
As illustrated in Figure 1, Monte Carlo also allows one to quantify the upside and downside
better than a Best / Worst approach, and to recognize which distribution (parameter) is
dominating the result uncertainty.
A limitation of Monte Carlo is that special software is typically utilized (commercial add‐ons to
Excel, etc), and is often not even an option in commercially available petrophysics s/w
packages. Common oilfield distributions, however, such as Normal, Log Normal and Triangle are
available in Excel and it is straight‐forward to implement Monte Carlo within the Excel
framework. In this approach, one remains in the familiar Excel environment, and actually
leverages their Excel skill set via the additional hands‐on experience within the platform.
A discussion of the Monte Carlo method can be found in Decision Analysis for Petroleum
Exploration by Paul Newendorp & John Schuyler, and a collection of articles addressing
exploration risk can be found in The Business of Petroleum Exploration published by the AAPG,
Tulsa, Oklahoma.
Additional information may be found in the References, with useful on‐line reference material
to be found at the following links.
• http://www.enrg.lsu.edu/pttc/
• http://www.mrexcel.com/
• http://people.stfx.ca/bliengme/exceltips.htm
• http://office.microsoft.com/en‐us/excel/HA011118931033.aspx
• http://www.statsoft.com/textbook/stathome.html
• http://en.wikipedia.org/wiki/Monte_carlo_simulation
• http://www.ipp.mpg.de/de/for/bereiche/stellarator/Comp_sci/CompScience/csep/csep
1.phy.ornl.gov/mc/mc.html
• http://www.sitmo.com/eqcat/15
• http://www.riskglossary.com/link/monte_carlo_method.htm
• http://www.chem.unl.edu/zeng/joy/mclab/mcintro.html

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

 The Gaussian or Normal Distribution
For illustration purposes, we focus here on the bell shaped Gaussian distribution. Log Normal or
Triangular distributions are easily handled with a simple change of Excel functions.
Gauss was a child prodigy and perhaps the greatest mathematician since antiquity
(http://en.wikipedia.org/wiki/Carl_Friedrich_Gauss). When the dwarf planet Ceres was
discovered, and observed for only a few days before vanishing, Gauss was able to
mathematically predict where it would be found a year later, and missed the mark by only half
a degree.
The bell shaped, or normal, probability distribution, is the most widely used family of statistical
distributions and came to be referred to as Gaussian because he analyzed astronomical data
within that context.
Two parameters characterize the Gaussian distribution, the ‘mean’ and ‘variance’: Figure 2
Figure 2 Excel’s NormDist function
•Petroleum applications typically use “normal”, “log normal”, and “triangular” [NORMDIST(x, mean,
statistical distributions.
standard_dev, cumulative)]
•Probably the best known statistical distribution is the “bell shaped” normal
distribution, whose probability density function is described by: calculates both the probability
density, and cumulative
probability, for a specified “x”
value with given mean and
•Two parameters characterize the distribution
standard deviation (standard
•the “mean” value µ,
•the variance represented by σ2.
deviation being the square
•The square root of the variance is the standard deviation. root of the variance): Figure 3.
http://www.enrg.lsu.edu/pttc/ Setting the logical variable
“cumulative” to “false” in the preceding expression will yield the “probability mass function”,
while setting “cumulative” to “true” returns the “cumulative” distribution.
Be aware that Excel 2007 is used for these illustrations and while there is a Compatibility Mode
for earlier Excel versions, it is conceivable that screens could differ. For greater clarity, the
graphics / text in Figure 3 are color coded, and in each case (true & false) the Excel cursor has
been placed in a calculation cell, so that the functional form of NormDist is displayed at the top
of the respective screen capture.
As an illustrative interpretation of the cumulative distribution, we recognize (Figure 3, right
side) that the Cumulative Probability (vertical axis) has reached 0.50 (50%) when the Probability
Distribution (horizontal axis) is 0. As expected, when the mean value is specified to be 0, there
is equal probability of any single value being higher, or lower.

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

Departing from the mean, Figure 3 Figure 3
(graphical display and Excel cell
values) reveal that a cumulative
probability of
• 16% has been reached @ x= ‐1.0,
• 31% @ x=‐0.5,
• 69% @ x=+0.5,
• 84% @ x=1.0, etc. •Syntax: NORMDIST(x, mean, standard_dev, cumulative)
•There are two variations of NORMDIST, according to the value assigned cumulative
There is then a 31% chance that “x” •Cumulative is a logical parameter that determines the form of the function
is at least ‐0.5 in magnitude, a 69% •If True, it returns the cumulative distribution function.
•If FALSE, it returns the probability mass function.
chance that “x” is at least 0.5 in
magnitude, etc.
The NORMINV [NORMINV(Probability, Mean, Standard_Dev)] performs the inverse operation
by returning the “x” value for a given cumulative probability of normal distribution with
specified mean and standard deviation: Figure 4.
Figure 4 This calculation indicates the
•The NORMINV Cumulative probability is
[NORMINV(Probability, Mean,
Standard_Dev)] performs the inverse
operation by returning the “x” value for
• 16% when “x” attains the value
a given cumulative probability of ‐0.99,
normal distribution with specified
mean and standard deviation • 31% when “x” attains the value
•Cumulative probability is 16% ‐0.50
when “x” attains the value -0.99
•Cumulative probability is 31% • 69% when “x” reaches 0.50.
when “x” attains the value -0.50
•Cumulative probability is 69%
when “x” attains the value 0.50

Monte Carlo Modeling of Sw(Archie)

With a basic understanding of what the Excel Gaussian Distributions options are, one is able to
model the Archie equation within that framework. For illustration purposes, we regard “a”, Rw
and Rt to be well‐known, and Φ, “m” and “n” subject to uncertainty as specified in Figure 5.
Allowance for uncertainty in “a”, Rw and Rt may be addressed by a straight‐forward extension of
the techniques presented here. Also, while the focus here is on the simple Sw(Archie), any
other model (shaly sand, for example) may be evaluated in a similar manner. Once the
concepts are understood, locally specific models are readily developed.
Each of the uncertain attributes is modeled as a random number input to NormInv, whose
mean value and standard deviation are locally appropriate. For example (Exhibit 5), the first

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

pass random estimate of porosity, Figure 5
with a distribution centered on 20
pu and having a standard
deviation of 1 pu, results in an
estimate of 21 pu. The random
values of “m” and “n”,
appropriate to the specified
distributions, are independently
and randomly determined, and Sw
calculated per the Archie relation. •As a quality control device, we determine and display the
distribution of random numbers, between zero and one, for the
number of Monte Carlo passes being used in a specific simulation
Because Excel recalculates (2000, in this example). In a perfect world there would then be
200 observations in each of the ten bins displayed
equations each time the
•Each of the uncertain attributes are modeled as a random number input to NormInv,
spreadsheet is opened, or whose mean value and standard deviation are locally appropriate. For example, the
specifications are changed, the first pass random estimate of porosity, with a distribution centered on 20 pu and
having a standard deviation of 1 pu, results in an estimate of 21 pu.
various results will change (your
line item spreadsheet values will change, each time you make a modification).
As a quality control device, we determine and display the distribution of random numbers,
between zero and one, for the number of Monte Carlo passes being used in a specific
simulation (2000, in this example). In a perfect world there would be 200 observations in each
of the ten blue bins displayed in Figure 5.
Figure 6 illustrates the relation between the magnitude of NormInv, and the distribution of
NormInv values, for different
Figure 6 standard deviations, at 90
1.50
NormInv Magnitude •Relation between NormInv Magnitude
simulations. Both distributions
1.50

1.00 1.00
NormInv Magnitude
0.50 and the distribution of
take on an approximate Gaussian
0.50
NormInv

NormInv

0.00 NormInv values appearance, with the larger0.00

•90 samples
‐0.50
standard deviation result ‐0.50

‐1.00 •The distribution of ‐1.00

‐1.50 NormInv Magnitude displaying more scatter. It is the

‐1.50
0.00 is ‘normal’ per
0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00

Rand() specified ‘mean’ & Rand()

distribution of NormInv values
Mean = 0. Std = 0.25 ‘std’, and will Mean = 0. Std = 0.50 that is driving the Sw(Archie)
approach the
NormInv Distribution expected ‘bell shape’ NormInv Distribution simulation. It’s important to
25 as yet more 10

20 simulations are
realize that each occurrence of 9
8
performed NormInv involves an 7
Frequency

Frequency

15 6
•It is the distribution 5
10
of NormInv values independent Rand() input. 4
3
5 that is driving the 2

0
The approach taken here is
Sw(Archie)
1
0
simulation
Bin Bin intended to parallel that of the
‐2.00 ‐1.50 ‐1.00 ‐0.50 0.00 0.50 1.00 1.50 2.00 ‐2.00 ‐1.50 ‐1.00 ‐0.50 0.00 0.50 1.00 1.50 2.00

LSU results (Must Read

supplemental material), which also includes Log Normal and Triangle distributions, and so can
be directly referenced if either of those distributions are required: www.enrg.lsu.edu/pttc/.

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

As an additional QC device, the statistical attributes of the simulated quantities (Φ, ‘m’ and ‘n’
in this example) are tabulated directly from the simulation population, and displayed
graphically: Figure 7.
With 2000 simulations, the Figure 7
model population nicely •Porosity is specified as a
Cross Check

replicates the input numerical Gaussian distribution,

centered on 20 pu with a
specifications, and the porosity standard deviation of 1 pu
distribution takes on the •2,000 calculations are
expected appearance (Figure 7). done, and the result
“checked’ by means of
Simulation results are reported histograming the resulting
porosity distribution and
both numerically and calculating the resulting
graphically: Figure 8. In this statistics Monte Carlo Distribution
450
particular case, there is a 95% •The MC simulation is observed to reproduce the 400 Porosity

specified inputs 350

likelihood that Sw is contained cyn 300
250

within + / ‐ 2 σ, (0.357 – 0.076)

•Random number “mean” & “Std” converge to e
u
q
e 200
rF
input values 150

< Sw < (0.357 + 0.076) Î 0.28 < 100

50
0
Sw < 0.433. 0.00 0.05 0.10 0.15 0.20 0.25 0.30

Porosity

In utilizing Excel frequency distribution graphics, one should take note of how the ‘bins’ are
populated, as they are not
Figure 8
Sw(Archie)
‘centered’. This can cause the
•One issue of interest is the
dependence of Sw upon
Uncertainty graphic to take on a shifted
individual attribute values appearance, with respect to the
/ uncertainties
numerical report (consult Excel
•With the specifications at
right Help on the Frequency function for
• Sw(mean) = 0.357 details).
• ◊→
(Sw) = 0.038
The Sw(Archie) result population is
•There is a 95% likelihood that Sw is contained
600
further affected by the nonlinear
M onte C arlo Distribution

within + / - 2 ◊ relation between the variousSw

500

y 400
c
•(0.357 – 0.076) < Sw < (0.357 + 0.076)
n
e
u
q
e
300 attributes, as discussed by Bryant
r
•0.28 < Sw < 0.433 F 200

100 et al in Understanding Uncertainty,

•Be aware of how Excel ‘bins’ data
0
0.00 0.10 Oilfield Review. Autumn 2002, who
0.20 0.30 0.40 0.50
Sw
illustrates that a normal
uncertainty distribution about a given porosity yields a log‐normal distribution for the
resulting Sw distribution. Bryant’s article is another Must Read.

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

 In Search of The Biggest Bang For The Buck
Attribute Uncertainties Specified Individually We are typically confronted with
Figure 9 Light Green Cells require User Specification
Light Blue Cells are calculated results
two issues, first to characterize the
• + / - 1 σ will encompass ~ 68% of the distribution
Individual Best Relative Uncertainty
Attribute UncertaintyEstimate On Sw(Archie)
a 0.0% 1.00 0.0000 uncertainty in the Sw estimate
• +/- 2 σ ~ 95 % of the distribution, Rw 0.0% 0.02 0.0000
Phi
m
15.0%
10.0%
0.10
2.00
0.0900
0.2121
itself, and next to identify where
•Approximate the “Phi” relative uncertainty of 15% n 5.0% 2.00 0.0224
@ 10 pu (1.5 pu) as 2 σ ~ 1.5 pu => σ ~ 0.75 pu for Rt
Sw
0.0% 40.00
22%
0.0000 time and money would be best
Monte Carlo Simulation purposes spent to reduce that uncertainty.
•That is, 2 σ encompasses 95 % of the spread in the distribution, and it is set equal to There are two ways to proceed: 1)
the 15 % uncertainty of the Chen & Fang analyses
take the derivative of Archie’s
•Approximate the “m” relative uncertainty of 10% @ 2.00 (0.2) as 2 σ ~ 0.2 => σ ~
0.1 for Monte Carlo Simulation purposes equation with respect to each
•Approximate the “n” relative uncertainty of 5% @ 2.00 (0.1) as 2 σ ~ 0.1 => σ ~ term, and compare the magnitude
0.05 for Monte Carlo Simulation purposes
of each term against one another,
•The Biggest Bang For The Buck is to be found in “m” as it has the greatest
uncertainty relative to the other attributes
for specific attribute values. 2)
Monte Carlo simulation, with the
After C. Chen & J. H. Fang. Sensitivity Analysis
The Biggest Bang for the Buck of the Parameters in Archie‘s Water Saturation input attribute distributions
Derivatives vs Monte Carlo Equation. The Log Analyst. Sept – Oct 1986 specified per locally representative
requirements.
The differential approach was illustrated in an earlier article, Risky Business, and those results
have been included in the following so as to both ‘make the connection’ and to also serve as a
cross‐check. The illustrative attribute values / uncertainties are those in the Chen & Fang (1986)
paper, so as to allow reference to that material as well.
In the case of Figure 9 attributes, the differential approach would indicate that time / money
would be best spent on “m”, as the relative uncertainty of the cementation exponent is far
greater than any of the other attributes.
The connection between Attribute Uncertainties Specified Individually
Figure 10 Light Green Cells require User Specification
derivatives and Monte Carlo is •Chen & Fang identify the attribute resulting in the
Light Blue Cells are calculated results
Individual Best Relative Uncertainty

made by recognizing that two greatest Sw uncertainty

Attribute UncertaintyEstimate On Sw(Archie)
a 0.0% 1.00 0.0000
Rw 0.0% 0.02 0.0000
standard deviations •In the case at right, “m” is the dominant issue Phi
m
15.0%
10.0%
0.10
2.00
0.0900
0.2121
encompasses 95% of the •This same issue can be addressed with Monte Carlo
n
Rt
5.0%
0.0%
2.00
40.00
0.0224
0.0000

statistical scatter, and then simulation (below) Derivatives vs Monte Carlo Sw 22%

setting, attribute by attribute, •The Base Case is at lower left, with each simulation towards the right reflecting an
2σ equal to the Chen & Fang improvement in “Phi”, “m” and “n” individual precision by 10 %, respectively.

illustrative uncertainties,
thereby forming the Base Case Attribute Mean Std Attribute
"a"
Mean
1
Std Attribute
"a"
Mean
1
Std Attribute Mean Std
"a" 1 "a" 1
for Monte Carlo: Figure 9. Rw 0.02 Rw
Rt
0.02
40
Rw
Rt
0.02
40
Rw
Rt
0.02
40
Rt 40
Phi 0.1 0.0075 Phi 0.1 0.0068 Phi 0.1 0.0075 Phi 0.1 0.0075
Monte Carlo simulations are 0.1000"m" 2 "m"
"n"
2
2
0.1000
0.0500
"m"
"n"
2
2
0.0900
0.0500
"m"
"n"
2
2
0.1000
0.0450
0.0500 "n" 2
then performed, incrementally, Monte Carlo Results Monte Carlo Results Monte Carlo Results
Monte Carlo Results
with each attribute better Sw Sw Sw Sw
Mean Std_Dev Mean Std_Dev
defined by 10% and the Mean Std_Dev Mean Std_Dev
0.226 0.0329
0.227 0.0332 0.225 0.0323 0.226 0.0297
improvement (reduced scatter)
in the resulting Sw noted: Figure 10.

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

Monte Carlo simulation reveals that a 10% improvement in definition of the cementation
exponent will yield the greatest reduction in Sw uncertainty, relative to the other attributes,
and consistent (as expected) with Chen and Fang (Figure 9).
Figure 11 We also note that the Best /
Attribute Mean Std
•One also observes that the Best / Worst numerical Worst case scenario would
"a" 1
Rw 0.02 evaluation of Sw(Archie) is considerably more significantly over‐state the
Rt 40 pessimistic than is the +/- 2 σ Monte Carlo
Phi 0.1 0.0075
simulation
95% Monte Carlo uncertainty,
"m" 2 0.1000
"n" 2 0.0500 •The Best and Worst of all parameters are because it’s unlikely (though
Sw 0.224
unlikely to occur simultaneously not impossible) that the Best
Monte Carlo Statistics
Monte Carlo Distribution
Cross‐check Specs Monte Carlo Results or Worst, of all attributes,
700
Porosity Sw
600
Sw
Mean Std_Dev Mean Std_Dev would occur simultaneously:
0.100 0.0076 0.226 0.0333
500
Figure 11.
Frequency

400 "m" "n"

300 Mean Std_Dev Mean Std_Dev
200
1.997 0.1005 2.001
Delta Sw
0.0508
0.133
Were porosity to be 25 pu,
100

0
High‐Low Numerical Statistics rather than the 10 pu of the
Attribute Low High Sw Range for Low Phi
0.00 0.10 0.20

Sw
0.30 0.40 0.50
Porosity 0.085 0.115 0.189 0.222 0.318 0.355 above example, the focus
"m" 1.8 2.2 Sw Range for High Phi
"n" 1.9 2.1 0.142 0.171 0.224 0.258 changes. Now attention is
Best Case Low m / Low n Hi m / High n best devoted to the “n”
Sw 0.224
Worst Case Max Delta Sw 0.212 exponent: Figure 12.

The uncertainty in an Sw Figure 12

estimate is a dynamic •If porosity were to be 25 pu, rather than 10, the focus changes
issue, dependent upon •“m” and “n” uncertainties have been set equal in this simulation
the relative magnitudes •The Base Case is at lower left, with each simulation towards the right reflecting an
of the input attributes improvement in “Phi”, “m” and “n” precision by 10 %.
•With the improved porosity, focus shifts to “n”, the tortuosity of the conductive
which are themselves (brine) phase in the presence of a non-conductive (hydrocarbon) phase, as offering
linked, and thus may the Biggest Bang For The Buck.
change through the •Improved “n” definition yields the greatest reduction in the standard deviation
of Sw
reservoir.
Attribute Mean Std Attribute Mean Std Attribute Mean Std Attribute Mean Std
"a" 1 "a" 1 "a" 1 "a" 1
Rw 0.02 Rw 0.02 Rw 0.02 Rw 0.02
Rt 40 Rt 40 Rt 40 Rt 40
Phi 0.25 0.0100 Phi 0.25 0.0090 Phi 0.25 0.0100 Phi 0.25 0.0100
"m" 2 0.1000 "m" 2 0.1000 "m" 2 0.0900 "m" 2 0.1000
"n" 2 0.1000 "n" 2 0.1000 "n" 2 0.1000 "n" 2 0.0900
Sw 0.089 Sw 0.089 Sw 0.089 Sw 0.089

Monte Carlo Results Monte Carlo Results Monte Carlo Results Monte Carlo Results
Sw Sw Sw Sw
Mean Std_Dev Mean Std_Dev Mean Std_Dev Mean Std_Dev
0.090 0.0132 0.090 0.0128 0.090 0.0127 0.090 0.0121

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

 Uncertainty Specification
Uncertainty arises from multiple sources, and includes (among others) the following.
• Calibration data (routine grain density & porosity, special core analyses for “m” & “n”,
etc)
• The down hole instruments which are making the actual measurements
• The assumed interpretive model
Thomas (1989), of
Figure 13. Core Analyses Uncertainty
Amoco, performed
•The samples are assumed to not change a detailed study of
between tests, so that differences reflect the accuracy and
random variations in the measurement
reproducibility of
process
repeated routine
•Deviations are interpreted as Gaussian in
nature, so that core analyses, and
•+/- 1 σ encompasses 68% of the data, provided the
•+/- 2 σ encompasses 95 % of the data resulting 99%
confidence limit:
•Tabulated confidence limits reflect the 99% limits
Figure 13. In
•A single measurement made on the same sample, which falls outside the specified
99% level is likely to be in error practice, the issue
is broader than just
The Log Analyst 30, No 2, March – April 1989 the individual core
A Statistical analysis of the Accuracy and Reproducibility of Standard Core Analysis. David C Thomas and Virgil J Pugh
measurements, and
also includes compaction adjustments to the core, volumetric differences between the core and
log measurements, etc.
In addition to the core
Figure 14. Wireline Tool Uncertainty
calibration data, there is also
uncertainty in the borehole
wireline / LWD
measurements: Figure 14.
In practice, particularly in a
field when there is legacy
data present, these
specifications will change
with time and tool type. The
6FF40 induction tool, for
example, had a skin effect
issue below about 1 ohm‐m
and a signal‐to‐noise limit
above ~ 100 ohm‐m. Newer
tools will have different Courtesy Tim Pritchard, BG Group
Illustrative Values. Consult your Service Company for locally specific attributes.
limitations.

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

The Archie exponents present an additional set of issues. Focke and Munn (1987) nicely
illustrate the dependence of ‘m’ upon carbonate pore geometry, while ‘n’ is controlled by
wettability (Sweeny & Jennings, 1960) and surface roughness (Diederix, 1982). In carbonates,
wettability (and hence “n”) may vary with pore size (Chardac et al, 1997), and pose an
additional challenge, particularly in the transition zone.
Adams (2005) cautions (and illustrates) that quantitative uncertainty definition is more than
just using Monte Carlo simulation to vary the inputs to the interpretation model. The largest
source of uncertainty may be the interpretation model itself.
Griffiths (2006) brings our attention to the challenges posed by carbonate dual porosity
systems and potential electrical ‘short circuits’.
Carlos Torres‐Verdin observes “my experience shows that the biasing of apparent resistivity
curves due to post‐processing (eg deconvolution) can be more detrimental to uncertainty than
Archie’s parameters, with a conspicuous example being thin, hydrocarbon saturated intervals
experiencing vertical resolution and invasion effects.
Voss (1998) comments on determination of uncertainty ranges as does Bowers (2003).
Summary points include
• A single interpreter should avoid making estimates on their own.
• A single interpreter often lacks the needed knowledge to correctly estimate
every parameter.
• In addition, many interpreters have a bias that smaller errors are better and they
will appear more knowledgeable about the subject.
• The error must reflect the level of knowledge about the parameters and the data
quality.
• A standard set of uncertainty ranges must be avoided because there is no standard
situation in which to apply them.
• Unusual events also pose special problems.
• Most people have a better recall of unusual events
• Therefore a tendency to overestimate the probability of such an event
• Especially if that event occurred recently
• Another very common mistake is to allow a very small amount of data to quantify the
range of uncertainty
• If data sets are small, the ranges probably need to be increased.
• Boundary Conditions
• Water saturation must lie between zero and one
• If the saturation values are too large or too small, the "best guesses” and ranges
must be reconsidered and calculations remade.
• The final and probably most difficult problem to overcome is the culture and
preconceived ideas of an organization.
Methods and ranges of uncertainty applied to any analysis must be questioned every time
they are applied.

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

These considerations (and others, included in the References) are not meant to be over‐
whelming, but rather simply realistic. Each case may very well be different, and must be
addressed on an individual basis.
Summary
There are two basic ways in which the issue of uncertainty can be characterized; partial
derivatives of the expression of interest (Sw in this situation) and Monte Carlo simulation. At
the simplest level, they complement one another, and since each are easily coded into an Excel
spreadsheet, we routinely perform both, as a QC cross‐check.
The deterministic derivative approach yields an equation, which may be easily coded into foot‐
by‐foot petrophysical analyses, in those cases for which the commercial petrophysics s/w does
not include an uncertainty characterization option. One is then able to ‘bound’ the calculated
results, foot‐by‐foot, which is an improvement over a ‘generic envelope’, given the inter‐
dependence of the result and specific reservoir values. An illustration of this method may be
found in Ballay, Risky Business, March 2005, www.GeoNeurale.com
On the other hand, an attribute specific distribution may not be Gaussian. Focke and Munn,
for example, investigated the dependence of the cementation exponent upon pore geometry.
Suppose across a given interval we are unable to distinguish between interparticle and vuggy
porosity; either is a possibility. The associated “m” distribution could then be rectangular, not
Gaussian, an issue that the Monte Carlo approach can easily address (each input attribute can
have its specific distribution, independent of the others).
In any case, it’s important to recognize the following.
• The uncertainty in Sw(Archie), and other common oilfield calculations, can be
quantitatively addressed by both differential and statistical modeling approaches.
• Excel can handle common probability distributions, and can then serve as the Monte
Carlo simulator. The derivative method will yield equations which may be easily coded
into Excel, thereby facilitating a cross‐check.
• Quantitative estimation of the uncertainty allows one to determine where time / money
is most effectively spent, and to further avoid the trap of being misled as a result of a
previous bad experience with a poorly defined parameter
• The importance of the various input parameters will change, according to the various
magnitudes. There may be a linkage in that one parameter becomes more or less
important as another parameter value is change. One size does not fit all feet.
• The equations resulting from the derivative approach may be coded, foot‐by‐foot, into
the petrophysics s/w package, thereby providing a live‐linked uncertainty estimate to
the actual, local reservoir properties.

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

 Acknowledgement
We appreciate the unidentified LSU faculty who posted their material (located via Google) to
http://www.enrg.lsu.edu/pttc/.
George Eden (BP Canada) and Larry Maple (ConocoPhillips) generously shared their thoughts,
and suggested relevant reference material, as ideas and material on this question were
gathered over the past few years.
Stefan Calvert (BG India, E&P) has kindly shared his thoughts and spreadsheet examples, as this
overview was put together.
Omissions, typos etc remain, of course, my responsibility.
References
Adams, S. J. Quantifying Petrophysical Uncertainties. Asia Pacific Oil & Gas Conference and
Exhibition, Jakarta. April 2005.
Ballay, Gene. Multidimensional Petrophysics in the Reservoir Description Division. Saudi
Aramco Jounal of Technology, Winter 2000/2001.
Ballay, Gene. Risky Business. March, 2009. www.GeoNeurale.com
Bowers, M. C. & D. E. Fitz. A Probabilistic Approach to Determine Uncertaint in Calculated
Water Saturation. Dialog; 8 April 2003. SPWLA 41st Annual Logging Symposium; June 2000.
Bryant, Ian and Alberto Malinverno, Michael Prange, Mauro Gonfalini, James Moffat, Dennis
Swager, Philippe Theys, Francesca Verga. Understanding Uncertainty. Oilfield Review. Autumn
2002.
Burnie, Steve. Error / Uncertainty and The Archie Equation. Insight : Canadian Well Logging
Society. January 2004
Case Western Reserve University. Appendix V of the Mechanics Lab Manual, Uncertainty and
Error Propagation (available on‐line)
Chardac, Jean‐Louis, Mario Petricola, Scott Jacobsen & Bob Dennis. In Search of Saturation.
Middle East Well Evaluation Review. Number 17. 1996.
Chen, C and J. H. Fang. Sensitivity Analysis of the Parameters in Archie‘s Water Saturation
Equation. The Log Analyst. Sept – Oct 1986
Denney, D. Quantifying Petrophysical Uncertainties. SPE JPT. September, 2005
Diederix, K. M. Anomalous Relationships Between Resistivity Index and Water Saturations in the
Rotliegend Sandstone (The Netherlands), Transactions of the SPWLA 23rd Annual Logging
Symposium, Corpus Christi, Texas, July 6‐9, 1982, Paper X
Focke, J. W. and D Munn. Cementation Exponents in ME Carbonate Reservoirs. SPE Formation
Evaluation, June 1987

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

Freedman, R. And B. Ausburn. The Waxman‐Smits Equation of Shaly Sands: I, Simple Methods
of Solution, II Error Analysis. The Log Analyst. 1985.
George, Bovan. A Case Study Integrating the Physics of Mud‐Filtrate Invasion with the Physics of
Resistivity Logging. MS Thesis. University of Texas. 2003. Download from UT site.
George, Bovan and C. Torres‐Verdin, M. Delshad, R. Sigal, F. Zouioueche & B. Anderson. A Case
Study Integrating the Physics of Mud‐Filtrate Invasion with the Physics of Induction Logging:
Assessment of In‐situ Hydrocarbon Saturation in the Presence of Deep Invasion and Highly
Saline Connate Water. Download from University of Texas site.
Griffiths, R. and A. Carnegie, A. Gyllensten, M. T. Ribeiro, A. Prasodjo & Y. Sallam. Estimating Sw
with a volume measurement. World Oil, October 2006
Hill, T. & P. Lewicki (2007)
Statistics, Methods and Applications. StatSoft, Tulsa, OK
http://www.statsoft.com/textbook/stathome.html
Hook, J. R. The Precision of Core Analysis Data and Some Implications for Reservoir Evaluation.
SPWLA 24th Annual Symposium, June 27‐30, 1983
Limpert, L. and W. Stahel & M. Abbt. Log‐normal Distributions across the Sciences: Keys and
Clues. BioScience, Vol 51 No 5, May 2001.
LSU. Probabilistic Approach to Oil and Gas Prospect Evaluation Using the Excel Spreadsheet.
Found with Google, Author n/a. http://www.enrg.lsu.edu/pttc/
Sweeney, S. A. and H Y Jennings Jr: The Electrical Resistivity of Preferentially Water‐Wet and
Preferentially Oil‐Wet Carbonate Rock, Producers Monthly 24, No 7 (May 1960): 29‐32
Thomas, David C and Virgil J Pugh. A Statistical analysis of the Accuracy and Reproducibility of
Standard Core Analysis. The Log Analyst 30, No 2, March – April 1989
Voss, David, 1998, Quantitative Risk Analysis: John Wiley and Sons, New York

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC

 Biography
R. E. (Gene) Ballay’s 32 years in petrophysics
include research and operations assignments in Mississippian limestone
Houston (Shell Research), Texas; Anchorage
(ARCO), Alaska; Dallas (Arco Research), Texas;
Jakarta (Huffco), Indonesia; Bakersfield (ARCO),
California; and Dhahran, Saudi Arabia. His
carbonate experience ranges from individual
Niagaran reefs in Michigan to the Lisburne in
Alaska to Ghawar, Saudi Arabia (the largest
oilfield in the world).
He holds a PhD in Theoretical Physics with double
minors in Electrical Engineering & Mathematics,
has taught physics in two universities, mentored Chattanooga shale
Nationals in Indonesia and Saudi Arabia, published numerous technical articles and been
designated co‐inventor on both American and European patents.
At retirement from the Saudi Arabian Oil Company he was the senior technical petrophysicist in
the Reservoir Description Division and had represented petrophysics in three multi‐discipline
teams bringing on‐line three (one clastic, two carbonate) multi‐billion barrel increments.
Subsequent to retirement from Saudi Aramco he established Robert E Ballay LLC, which
provides physics ‐ petrophysics consulting services.
He served in the U.S. Army as a Microwave Repairman and in the U.S. Navy as an Electronics
Technician, and he is a USPA Parachutist and a PADI Dive Master.

WWW.GeoNeurale.Com July 2009 © 2009 Robert E Ballay, LLC