SPE 128349

Latin Hypercube Sampling (LHS) for Gas Reserves

*Ekeoma, E and D.A. Appah, Institute of Petroleum Studies, University of Port Harcourt, Rivers State, Nigeria.
*Currently in Orwell international (Oil & Gas) Nigeria Ltd

Copyright 2009, Society of Petroleum Engineers Inc.

preparation of estimates that have an inherent degree
This paper was prepared for presentation at the 33rd Annual SPE International Technical of associated uncertainty. Classifications of proved,
Conference and Exhibition in Abuja, Nigeria, August 3-5, 2009.
probable and possible reserves have been established
This paper was selected for presentation by an SPE Program Committee following review of
information contained in an abstract submitted by the author(s). Contents of the paper, as to reflect the level of these uncertainties and to provide
presented, have not been reviewed by the Society of Petroleum Engineers and are subject
to correction by the author(s). The material, as presented, does not necessarily reflect any an indication of the probability of recovery.
position of the Society of Petroleum Engineers, its officers, or members. Papers presented
at SPE meetings are subject to publication review by Editorial Committees of the Society of The estimation and classification of reserves requires
Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this
paper for commercial purposes without the written consent of the Society of Petroleum the application of professional judgment combined with
Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not
more than 300 words; illustrations may not be copied. The abstract must contain geological and engineering knowledge to assess
conspicuous acknowledgement of where and by whom the paper was presented. Write
Librarian, SPE, P.O. Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952- whether or not specific reserves categorization criteria
9435.
have been satisfied. Knowledge of concepts including
uncertainty and risk, probability and statistics, and
Abstract deterministic and probabilistic estimation methods is
required to properly use and apply reserves definitions.
Latin Hypercube Sampling (LHS), unlike other Reserves estimates may be prepared using either
simulation methods such as Monte Carlo Simulation deterministic or probabilistic methods (Petroleum
(MCS), has not enjoyed wide applications in Society of CIM definitions-guidelines, 2002.).The
determination of reserve despite its capability to realize deterministic approach, which is the one most
same accuracy as with the latter even with much fewer commonly employed worldwide, involves the selection
simulation trials. The parameters used in these of a single value for each parameter in the reserves
distributions are extracted using software “digitizer” calculation. The discrete value for each parameter is
from different plot of values derived from selected based on the estimator’s determination of the
measurements of reservoir and fluid properties. This value that is most appropriate for the corresponding
work dwells on analysis of two depleted gas fields for reserves category.Probabilistic methods involve
which the ultimate recovery is known. The distribution describing the full range of possible values for each
of porosity, water saturation, reservoir thickness and unknown parameter. This approach typically consists
reserve size are input to crystalbal software to of employing computer software to perform repetitive
generate cumulative distribution function (CDF) and calculations (e.g. “Monte Carlo Simulation”) to
Probability density function (PDF) of original gas in generate the full range of possible outcomes and their
place (OGIP) for each of the fields. The accuracy of associated probability of occurrence. Estimating of
Latin Hypercube Sampling is analysed by comparing hydrocarbon volumes with probability has its most
the actual ultimate recovery and the OGIP predicted by important application when associated with major
LHS. The calculated recovery factors of 68.7% and petroleum development projects. Simulations may be
93.3%, for water driven and depletion driven gas carried out by an analytical or a probabilistic approach.
reservoirs respectively, compare favourably with those While both methods have their strengths and
of known gas reservoirs. Therefore, LHS is a good weaknesses, it is the latter method that lends itself
estimator of gas reserve. It is recommended that more easily to describing uncertainties associated with
similar study should be carried out on condensate and hydrocarbon volumes which have to be estimated. In
oil reservoir, particularly on Niger Delta fields. this respect, one should differentiate between
“recoverable hydrocarbons” (oil or gas), a quantity
Introduction which represents the maximum possible recovery
The determination of oil and gas reserves involves the essentially governed only by physical reservoir
2 E. Ekeoma and D.A. Appah SPE 128349

processes and “reserves” which is the maximum Log-normal Distribution

quantity (usually less than recoverable hydrocarbons) The log-normal distribution is frequently associated
that can be recovered with a certain development plan with process that divides large quantities into smaller
and production policy.There are three categories of ones. This is exactly the type of process that occurs
reserves; proved, probable and possible. during sediment deposition. There, log-normal
The process of estimating oil and gas reserves for a distributions are common in petroleum industry
producing field continues throughout the life of the (Harbaugh et al, 1977).
field.
The lognormal distribution is used to model
continuous random quantities when the
Distribution Types distribution is believed to be skewed. The
There are many different shapes that the CDF and lognormal density function, with parameters μ and
PDF can take depending on the data they are σ, is given by
describing. Mathematicians have developed models of
the CDF and PDF for certain distribution types that
occurs frequently in applications. ⎛ 1 ⎞ [−( x − μinx )2 / 2 (2σ inx 2 )]
F ( x ) = ⎜⎜ ⎟e
⎟
,
xσ
⎝ inx 2π ⎠
Normal Distribution
The expected value of normally random distributed is
equal to the mean. The normal distribution is also
The CDF is given by
called the Gaussian distribution and the PDF and CDF
are symmetrical in shape. ⎛ 1 ⎞ [−( x − μinx )2 / 2 (2σ inx 2 )]
F ( x) = ⎜ ⎟e dx
The probability density function has notable properties ⎜ xσ π ⎟
including:
⎝ inx 2 ⎠

• symmetry about its mean μ The parameter μ is the mean and σ is the
standard deviation of the distribution for the
• the mode and median both equal the mean μ
normal random variable ln[X], not the lognormal
random variable X.The lognormal distribution
• the inflection points of the curve occur one
arises from many small, multiplicative random
standard deviation away from the mean, i.e. at
effects, in contrast to additive random effects that
μ − σ and μ + σ.
lead to the normal distribution.
It is mathematically represented as
Uniform Distribution
(
F ( x) = 1 2πσ e 2
) [− ( x − μ )2 / 2 (σ 2 )]
, − ∞ < x < +∞ Uniform Distribution is a two parameter distribution and
is the simplest of all distribution discussed here. It is
useful when little is known about the distribution of the
The cumulative distribution function (CDF) of a subject parameter other than a maximum and a
probability distribution, evaluated at a number (lower- minimum. In this study, the uniform distribution is used
case) x, is the probability of the event that a random as an estimator for the aerial extent of the studied
variable (capital) X with that distribution is less than or reservoir.
equal to x. The cumulative distribution function of the
normal distribution is expressed in terms of the density Average (mean) is the measure of central tendency
function as follows: for a normal distribution.Variance is a measure of
dispersion for a normal distribution.
Mathematically, the CDF is given by
Skewness tests if the shape of a sample distribution is

(
F ( x) = 1 )
2πσ 2 e [− ( x − μ )2 / 2 (σ 2 )] dx
similar to that of a normal distribution. A risk
investigator prefers positive skewness or a value
between -0.5 and +0.5. Kurtosis also tests if the shape
Additionally, a plot of x vs. F(x) on probability paper will of a sample distribution is similar to that of a normal
yield a straight line if x is normally distributed. distribution. A risk investigator prefers a distribution
3 Latin Hypercube Sampling (LHS) for Gas reserves SPE 128349

with low kurtosis. normal distribution N(0,1) as n approaches ∞ (this is

convergence in distribution).
Selection of Number of Trials

One of the issues in using sampling methods is, "How

Latin Hypercube Sampling (LHS)
many trials is enough?" John Schuyler (1997) stated The statistical method of Latin hypercube sampling
that the approach is to use the standard error of the (LHS) was developed to generate a distribution of
mean statistic, plausible collections of parameter values from a
multidimensional distribution. The sampling method is
S .E.M = σ sqr (n) often applied in uncertainty analysis.

Where σ is the standard deviation of the variable of This technique was first described by McKay (1979). It
interest (e.g. reserve), and n is the number of trials in was further elaborated by Ronald and others (1981).
the simulation.For most purposes, running the Detailed computer codes and manuals were later
simulation until the S.E.M. is less than 1% of the mean published by Imam and others (1980).
is a good rule-of-thumb stopping rule.
Latin hypercube sampling begins by estimating each
parameter’s uncertainty using a probability distribution.
Central Limit Theorem For some applications such as numerical integration
The central limit theorem (CLT) states that the re- we would only use uniform distributions in each
averaged sum of a sufficiently large number of dimension since we are trying to see which fraction of
identically distributed independent random variables points fall within the object of interest and which
each with finite mean and variance will be outside.
approximately normally distributed (Rice 1995).
Formally, a central limit theorem is any of a set of We would break up the distribution into N equal
weak-convergence results in probability theory. They probability segments and a value for the parameter
all express the fact that any sum of many independent would be generated from each segment. Either the
identically distributed random variables will tend to be midpoint or the endpoints of each segment are used to
distributed according to a particular "attractor generate the N parameter value. Notice that each
distribution". segment has equal area. Midpoints of each segment
Since many real populations yield distributions with were used to generate the black points on the bottom.
finite variance, this explains the prevalence of the LHS is a stratified sampling technique where the
normal probability distribution. random variable distributions are divided into equal
probability intervals. A probability is randomly selected
Let X1, X2, X3 ... Xn be a sequence of n independent from within each interval for each basic event. During
and identically distributed random variables having sampling, a sample is drawn from each interval. In
each finite values of expectation µ and variance σ2 > 0. LHS, the samples more accurately reflect the
The central limit theorem states that as the sample size distribution of values in the input probability
n increases, the distribution of the sample average of distribution.
these random variables approaches the normal Generally, LHC will require fewer samples than simple
distribution with a mean µ and variance σ2 / n Monte Carlo Simulation for similar accuracy.The LHC
irrespective of the shape of the original distribution. sampling method reduces the number of samples and
variance. In LHS, the range of each variable is divided
Let the sum of n random variables be Sn, given by
into non-overlapping intervals (m) on the basis of equal
Sn = X1 + ... + Xn. Then, defining a new random width or equal probability. These intervals are sampled
variable according to probability density functions associated
with the variables. Rather than sampling all possible
S n − nμ combinations, the method selects only m of these
Zn = combinations. Each stratum of each variable is only
σ n sampled once without replacement.Thus, the full range
the distribution of Zn converges towards the standard
of each input variable is sampled. This can significantly
improve the accuracy and convergence rate.
4 E. Ekeoma and D.A. Appah SPE 128349

was calculated using the simanddoux equation

(Bassiouni, 1994).Based on the unknown water-gas
Advantages of Latin Hypercube sampling contact, the downdip limit was established. Some
uncertainty of the size of the reservoir exists due to
1. Despite their flexibility, MC methods can require complex faulting. The net sand thickness was
much iteration to achieve the desired accuracy. estimated from the sand thickness as determined by
the logs of seven wells that penetrated and produced
2 Latin Hypercube sampling can improve the efficiency
the reservoir.
of Monte Carlo simulations by picking the input
The CDF and PDF plots for porosity are characteristics
samples better. Whereas MC methods typically pick
of a normally distributed variable. To check whether or
points at random within the domain, Latin Hypercube
not the distribution is truly normal, a plot of porosity
sampling samples the entire domain more
versus the probability transform is made. This plot is
systematically.
equivalent to a probability plot and can be used to
3 Dapeng (2006) maintained that LHS provides better check for normality. This plot is equivalent to a
distributed samples than random sampling and also probability plot and can be used to check for normality
improves sampling quality as well as optimization .For a normally distributed variable, a straight line is
efficiency. expected. .
4 LHS strongly improves the representation of Using Crystal Ball, a normal distribution is assumed
stochastic design parameters compared to a standard with a mean of 26.41% and a standard deviation of
Monte Carlo sampling. The method presented offers 2.55%.Figures 3-5 and 3-6 show the CDF and PDF for
the same general applicability as the standard Monte the water saturation data.
Carlo sampling method but is superior in computational The PDF and CDF plots for water saturation appear
efficiency. similar in character to the log-normal distribution.To
check if the distribution is truly log-normal, a plot of
Case Studies water saturation values versus the probability
Data from two depleted gas reservoirs were collected transform on a semi –log plot is made. For a log-
and analyzed. normal distribution, this plot should be a straight line.
In each case, distributions of porosity, water saturation, Based on the above analysis, water saturation is
reservoir thickness, and aerial extent of the reservoir modeled as a log-normal distribution with mean 23.9%
were developed with the help of cores, wireline logs, and standard deviation 9.08%.This distribution is input
and seismic lines and geologic maps. Other data such into crystal Ball software for use in Latin Hypercube
as pressure-volume-temperature information were Sampling (LHS) simulation.
used to estimate the parameters needed in the The reservoir thickness is modeled with a triangular
volumetric equations. The derived distributions were distribution with minimum value of 20 feet, most likely
loaded into the crystal Ball software and 5,000 trials value of 49 feet, and a maximum value of 98 feet. The
performed using Latin Hypercube sampling for each area is approximated with uniform distribution varying
case. from 1286 acres to 1402 acres. Initial reservoir
pressure was measured to be 5911Psi and the
Case 1 – Gulf of Mexico Gas Reservoir compressibility factor is 0.93.The porosity and water
The first case study is a gas reservoir located in the saturation are checked for correlation by computing the
Gulf of Mexico. The reservoir drive mechanism was correlation coefficient, which is is -0.42.
water drive. The reservoir is at an average depth of
11,000 feet and is clean, unconsolidated sandstone. To check for the importance of the correlation between
Seven wells produced this reservoir 1962 until 1998 porosity and water saturation on the prediction of
with an ultimate recovery of 117,798 MMcf. Core data OGIP, the simulation is run assuming a correlation
was available for five of the wells. Form the cores, an coefficient of -0.42 with respect to these variables and
empirical CDF and PDF of porosity was developed. again assuming no correlation between any variables.
Because the wells were drilled with a water based The required distribution are input into the simulator
mud, reliable estimates of water saturation could not and 5000 trial calculations are made to create a
be made from the core data. As a result, the resistivity distribution of the OGIP based on the volumetric
and gamma logs were analyzed and water saturation equation correlation and no correlation.
5 Latin Hypercube Sampling (LHS) for Gas reserves SPE 128349

manifests itself in the form of distributions with larger

Case2- Gulf Coast Gas Reservoir standard deviations and larger mean values.
Case 2 is a gas reservoir located of Louisiana. The
drive mechanism was depletion drive. The reservoir is The results of the simulation when no correlation
at approximately 10,000feet and has good core data between water saturation and porosity is assumed are
from three penetrations, and the available data is very similar to the calculations of OGIP when
somewhat limited. The reservoir produced correlation is considered.
22476MMscf from 1943 to 1953. The distributions obtained for OGIP obtained from the
The core data used to construct the PDF, CDF, and Latin Hypercube Sampling (LHS) for each case can
probability transforms plots for porosity. There is a now be compared to the actual production histories for
cluster of porosity values around 30%, as can be seen each of these reservoirs. The recovery factor is the
on the PDF plot. The probability transform plots as a percentage of oil or gas in the ground actually
reasonably straight line, and it is determined that the produced.The table below shows the recovery factors
porosity is approximately normal with a mean 27.2% of each case based on the actual volumes produced
and a standard deviation of 5.4%. corresponding to the calculated in the Latin Hypercube
The water saturation data is obtained from the wireline Sampling.
log measurements in conjunction with the core porosity
data. Based on the positively skewed PDF plot and the Calculated Recovery Factors
reasonably straight line seen on the semi-log plot of CASE TYPE DRIVE UR %RECP
the probability transform, the water saturation is
assumed to be log normally distributed. The water MECH 10
saturation has a mean of 20.97% and standard
deviation of 6.5%.
1 GAS WATER 117,798 68.7
The reservoir thickness is a triangular distribution
based on the pay thicknesses seen in the existing
welbores with a minimum value of 15 feet, a most likely 2 GAS DEPL 22476 93.9
value of 41 feet, and a maximum value of 71 feet. The
aerial extent of the reservoir is estimated with uniform
distribution between 375 and 450 acres. The
correlation coefficient between porosity and water The value of OGIP when correlation was considered
saturation is -0.69. was used when calculating the recovery factors.
Based on the correlation coefficient, porosity and Real recovery factors vary from 5% to 90% depending
saturation are correlated strongly enough to make two on the reservoir fluid properties and the reservoir drive
separate simulations to determine the effect of the mechanism. The recovery factors shown above are
correlation on the OGIP distribution. The parameter reasonable for the drive mechanism of the reservoirs
distributions are input into the Latin Hypercube and the associated fluid types. None of the Latin
Sampling (LHS) simulator and 5,000 trials calculations Hypercube Sampling (LHS) results shown in Table
are made for the case of correlated porosity and water above predict greater than 100% recovery, which
saturation and also for the case assuming no would be impossible. In this respect, the results
correlation of variables. indicate that the LHS provide reliable estimates of
hydrocarbons originally in-place. There have attempts
Results and Discussion made to estimate reasonable ranges of recovery
The correlation between porosity and water saturation factors in the literature.
were so significant that it was considered necessary
during simulation. These correlations were always
between porosity and water saturation. Based on the
comparison of the distributions between the
simulations including correlation and the simulations
not including correlation, it is concluded that correlation
has a very small effect on the calculations. However,
the effect of correlation between the variables
6 E. Ekeoma and D.A. Appah SPE 128349

Standard Recovery Factors reserve.

Recovery Factor %
Recommendations
A review of the existing literature shows that no
Accumulation Drive Type Low High
analysis has been carried on the Latin Hypercube
Sampling to determine its accuracy as an estimator of
Gas Water 50 70 gas reserve except similar work done on Monte Carlo
simulation (MCS) by John, 2004. It is my opinion that
Gas Depl 80 90
since the LHS enjoys many advantages when
compared to MCS, works should be done to verify the
accuracy of the former in estimation of condensate and
Oil Water 15 50 oil reserve. I suggest that LHS should be done with
data from Niger Delta fields as depletion fields become
Oil Depl 5 30 available.

References
Anders M. J. Olsson and Göran E. Sandberg (2002):
A compilation of some of these attempts is shown in Latin Hypercube Sampling for Stochastic Finite
the Table above. These values are estimates only and Element Analysis. American Society of Civil Engineers.
there may be cases where the actual recovery factor of Submitted September 24, 1999; accepted June 6,
a reservoir is outside of the ranges given. 2001
In particular, by examining the tenth percentile(F10) of Bassiouni L: Theory, Measurement, and Interpretation
OGIP, which is by definition the proven reserve value of well logs, (1999), pp. 16-17.
of the distribution, and calculating the resulting
Cooper Energy NL - Probabilistic Reserves Reporting
recovery, the predictions of OGIP in fact quite good. Guidelines.
The recovery factors calculated against the F (10)
values of OGIP are considered normal. Harbaugh, Dovoton and Davis (1997), Probability
methods in Oil Exploration, pp.20-30.
Summary and Conclusions Henk Tijms, Understanding Probability: Chance Rules
Latin Hypercube Sampling was applied to two different in Everyday Life,: Cambridge University Press,
mature or depleted reservoirs to determine its Cambridge 2004.
effectiveness as an estimator of OGIP. The resulting
http://en.wikipedia.org/wiki/Latin_hypercube_sampling
distributions from the simulations allow us to conclude
that the procedure worked well in predicting the OGIP http://en.wikipedia.org/wiki/Latin_hypercube_sampling"
for the reservoirs studied. The back calculated
recovery factors in general fall within accepted ranges http://en.wikipedia.org/wiki/Law_of_large_numbers
of what would be expected from the reservoirs given
http://www.petrobjecs
their drive mechanism and hydrocarbon properties. By
inspection of the PDFs of OGIP, the distributions Iman, R.L.; Davenport, J.M. ; Zeigler, D.K. (1980).
appear to be approximately log-normally distributed as Latin hypercube sampling (program user's guide).
predicted by the Central Limit Theorem. The effect of OSTI:5571631.
the correlation between variables does not appear to
Iman, R.L.; Helton, J.C.; and Campbell, J.E. (1981).
have a large effect on the distributions though in cases "An approach to sensitivity analysis of computer
where correlation was considered. Additionally, models, Part 1. Introduction, input variable selection
distributions with correlation built in to volumetric model and preliminary variable assessment". Journal of
tended to be more dispersed than those with no Quality Technology 13 (3): 174–183.
correlation. Computed RF is less than 100% and the Jensen, Lake, etal.1997. Statistics for Petroleum
Calculated 68.7% and 93.3% are within the RF of Engineers and Geoscientists
water driven and depletion driven gas reservoirs John D.W.(2004),”Analysis of Monte Carlo simulation
respectively.Therefore, LHS is a good estimator of gas as an estimator of original oil in place and original gas
in place”, M.Sc Thesis, Faculty of Graduate School,
7 Latin Hypercube Sampling (LHS) for Gas reserves SPE 128349

University of Texas at Austin.

John L. and Robert A. Gas Reservoir Engineering,
SPE textbook series, USA, vol 5, 1996, pp349.

John S. (1997)”stopping rule for Monte Carlo”

http://www.maxvalue.com/tip025.htm
Definition and Guidelines for Estimating and classifying
Oil and Gas Reserves: Petroleum Society of CIM
definitions-guidelines January 25 2002 public draft
committee.
Rice, John (1995), Mathematical Statistics and Data
Analysis (Second ed.), ISBN 0-534-20934-3 Probability transform plot for porosity
Risk Analysis and Decision-Making Software Package
(1997 Version)
Wumi I. (2008): Lecture Notes on Rock Properties.
Institute of Petroleum Studies, Port-Harcourt, Nigeria.

Appendices
CASE 1

CDF for water Saturation

Gas production plot for Case 1

PDF for water Saturation

CDF for porosity

Probability transform plot for water Saturation

PDF for porosity

8 E. Ekeoma and D.A. Appah SPE 128349

CASE 2

Gas production plot

PDF for water saturation.

CDF for porosity

Probability Transform Plot for water saturation.

PDF for porosity.

Probability Transform Plot for Porosity.