You are on page 1of 10

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/261651653

Probability Logs for Facies Classification

Article in IN SITU · March 2000

CITATIONS READS

18 472

3 authors, including:

Larry Lake
University of Texas at Austin
442 PUBLICATIONS 15,234 CITATIONS

SEE PROFILE

All content following this page was uploaded by Larry Lake on 20 April 2014.

The user has requested enhancement of the downloaded file.


*-* m
,.:

-
Soclaty d Pehleum qtn-

SPE 39805

Modeling Nonlinear Phenomena


Loveena Kapur, Larry W. Lake, and Kamy Sepehrnoori, The University of Texas at Austin

Cqht 109S, See”@ d P6trdewn Enginaara, Inc.

This pa~r W= prapared for fwaaentation at tha 109S SPE Petian Brnin 04 and Gas Introduction
fl~ry Confemnm M in Wdarrd, Taxaa, 2S-27 March lW.
Accurately modeling nonlinear phenomena is always a
This papar WM salactad for pentation
information contmnad in an abdti
by an SPE Progrm Cunmitfaa fdldng
aufnn~
Mew of
~ tha a~tror( s). Contents d the fmfmr, as
challenge for engineers and scientists, As there are severat
praaanted, haua @ been radawed by the Scddy d Patrdeum Engine8m and are suti~ to nonlinear phenomena in nature, most relationships in
~i by the author(s). Tha material, - mentad, * ti nasarily reflect any
pmMon d tha Society d Pat*wn Engi-ra, Ita offiim, or mkra. Papra mented at
petroleum engineering are also nonlinear. Various statistical
SPE m.aatinw am subect to ~b~i raviaw by Editorial Cd- d the Scciety d modeling techniques are discussed in the Literature for
Petmlem Errgineara. Etirunb ra@*, diati~on, or aforaga d any wrt of ttis wr
for cunmerdal f)urpmes wit~ tfm written rnn$ent d the S- d Pefmtewn Enginaam ia developing accurate correlations. The key idea in every
~hibfiad. Permbaion to raprodm in print Is radridad to an afmtract d not mom than W
words; iflustratis mey d ba m. l?w *tract mud ccmtdn c0n5@cu0ua
technique is to eliminate noise and nonlinearity in such a way
ackntigment d wham and @ whcan h pm was -ded. Writs Librariam SPE, P.O. that optimat correlation can be obtained. We have chosen
Sax -, ~hardam, ~ 7SOSS-W U.SA., fax 01-972W-9435.
three common techniques, linear regression, optimal
transformatim, and artificial neural networks for
Abstract investigation, The f~st part of this paper compares the three
This paper discusses various ways to quantify nonlinear techniques using a noisy and nonlinear synthetic function.
relationship% in general, and an applicatim of the prtiures In the second part we selected the optimal
explored to the Atlas of Major Texas Oil Reservoirs database transformation technique to predict the oil recovery efficiency
compiled by The University of Texas Bureau of Economic reported in the AUas of Major Texas Oil Reservoirsa
Oeology. (AMTOR). Good correlations were attained only after the
A system is nonlinear if the relationship between its reservoirs were classified according to drive mechanism. The
input and output cannot M described by a straight line. correlations between recovery efficiency and reservoir
Detiting such a relationship is confounded when random properties were not as good for lithology classifications.
noise is present in the system’s output. Most relationships in
petroleum engintig are nonlinear. The first part of this Correlation Teohniqttes
work describes a numericrd experiment that uses a known but The following section briefly discusses three gentiic
noisy nonlinear functim to genaate system input. The correlatim techniques, linear regression, Optimat
function contains a parameter that specifies the degree of transformations, and artificial neural networks.
nonlinearityy and an independent white noise component. For
varying degrees of nonlinearity and noise we investigate Linear Reg-ion
which of three tiniques, conventional linear regression, the Single vwiable linear regression (LIN) is used to produce a
optimal transform method, and a neurat network, best model that will ptilct some property Y from a measurement
reproduce the input-output response. As expected, linear of another property,
regrwion works best (in the sense that it reproduces the
known tiction) when nonlinearity is small. optimal Y=p1+p2x+& (1)
transforms work best for moderate amounts of noise and for X is the predictor or expfanufory variable, Y is the response, E
nonlinear functions. No technique works, regardless of how is a random error, and ~1 and ~z are the regression model
nmlinear it is, when the s~tem is dominated by noise. parameters. X is an explunuto~ variable because knowledge
The optimal transform procedure is illustrated by of how X varies can be used to explain variations in Y.
correlating oil recovery efficiencies reported in the Attas of
Major Texas Oil Reservoirs database. The results are obtained Optimal Transformations
by performing the pr~ure on data sets that have been pre- When large variations in proWrties occur, linear regression
processed by dividing according to drive mechanisms and/or will fail. Optimat transformatims are used to obtain a best-fit
reservoir classes. additive model. The optimat transformations aid in finding
the strongest correlation between a dependent (response)
random variable and multiple independent (predictor) random
variables. The alternating conditional expectation (Am)

457
2 LOVEENA KAPUR, LARRY W. LAKE, KAMY SEPEHRNOORI SPE 39ao5

algorithm, originally proposed by Breiman and Friedman’, changing the two parameters, K and e, in the above equation.
provides a method for estimadng optimai transformations for The K parameter determines the degree of nonlinearity in the
multiple linear regression. Knowledge of such relationship between the independent variable (X) and
transformations aids in the interpretation and understanding of dependent variable (Y). K= 1 corresponds to a linear
the relationship between the response and predictors. relationship and K iarge (or smail) means a very noniinear
The basic idea behind the ACE aigorithm is that relationship. E sign~les independent white noise with a zero
linear correlations can be greatiy improved if all the variables, mean, variance Var (e)= ~c2,and E (e)= E(c, Y)=E (&A)=O.
both dependent and independen~ are transformed prior to The X is generated randomly within the range (0,1); this
correlation being sought. Simple examples of this are causes YACmd to be between zero and one aiso, but Y can be
transformations to standard normai variables or to logarithmic significandy outside this range depending on the magnitude of
space. However, tie ACE pwedure is mom flexible than 0,2. The results in this paper are from a single set of randomly
this. In fact the transformations in ACE are developed to generated X. Y is obtained from Eq. (2) by varying the degree
eventually lead to the stingest linear collations possible. of nonlinearity (K) and the noise variance (0s2).
See Breiman and Friedman ‘ for ftier details. See Xue et The standard measure of the goodness of an
al.’ for examples of the application of ACE to other petroleum
association between two variables is determined by the
engineering problems.
coefficient of detennination~, defined as,
Artificial Neural Networks Techniques and Aigorithm
Artificial neurat networks (ANN) m a class of numerical Cov(zl, Z2) 12
R2(Z1, Z2)=[ (3)
optimization algorithms originally inspired by studies of the JVar(Z1 ) Var(Z2 )
brain and nervous system. In genemi, ANNs function as
nonlinear dynamicai systems that learn to recognize patterns
where Z1 and z are any two random variables. R2is a way of
after undergoing training, The training involves presenting the
ANN with a series of input values and desired output values expressing the strength of a linear association between two
variables Z1 and u, however, Eq. 2 is most definitely not
(training patterns). In effecq the ANN implicitly &fines its
own, usually very complicated, predictive function as a result linear. Consequentiy, R2 can be different from one either
of the training. Once trained, the ANN can be applied to new because the relationship between Z] and &is not linear or the
input data, ailowing prediction of the output vaiues. system contains noise. Figure. 1 illustrates this fact by
tificial neum.i networks have two major plotting R2 (Y,X) against the nonlinearity coefficient, K, with
components: nodes (aiso caited units or neurons) and noise variance, aC2, as a parameter. Noise-free noniinear
connections (weighted links between the nodes). Each node systems can have an R2 as smatl as 0.5 over the range of K
(ex@pt those receiving input values) applies an arbitrarily Used in Fig. 1.

chosen function (known as a basis, activation or transform The second phase of the experiment consists of tests
function) to the weighted sum of the results passed from other of the model described above. The word “Model” in a
connected nodes. In this manner, input vaiues w passed subscript refers to linear regression (LIN), the ACE (Am) or
through the network and transformed into one or mom output Mlcial neural networks (ANN) as the case may be. A
values. The output values m then com~d to the desired relationship is said to & we~ly noniinear if K=l, modemtely
values (e.g. recovery efficiency) and, in accordance with the nonlinear if K=5 and strongly nonlinear if K =1000. The
particulru learning algorithm chosen, the weights m adjusted. noise is reported as normalized by the variance of Y,
This process iterates until the error between the output and var(Y)=oy2 , The noise is smail if this quantity is less than
desi~d vaiues is minimized. Details about generat ANN 0.11, moderate if it is between 0.11 and 0.98, and strong if is
theory can b found in many publications (e.g. Stephen’, and greater than 0.98. In all cases input to the respective model is
Hertz et ai.’ ). X and output Y~&,. RecaU that the miationship between Y_
We use the back propagation atgorithm to train the and YM~tshould be lin~ if the model is a good predictoc
ANNs in this study. See Kapur’ for more details. hence, R (YMJ , YM&,)should be a good measw of the
success of a correlation.
Numericai Experiment
We use a simple numerical experiment to investigate whiti of Results
the three techniques discussed above best reproduces the The results of calculating R2(YACmd
, Y~O~,Jfor a rangeof K
input-output response for a simple nonlinear system. The and 0C2are summarized in Table 1. Figures 2-5 show the
following nonlinear synthetic function is used to perform the results of the predictions for selected cases. Each case is a
experiment single-rdizatim exampl~ the results might change somewhat
x if multiple executions were to be performed.
Y=Y (2) Table 1 shows tha~ as expected, ail models perform
‘CW+S=X+K(l -X)+ E
well in the case of small noise and weakly nonlinear functims.
This is true regardless of the noise level as long as the
where Y is the output from the system defined by Eq. 2, Y~C~d
underlying function is linear. For moderately noniinear
is the noise-tiee output. The experiment is carried out by

458
SPE 39s05 MODELING NONLINEAR PHENOMENA 3

functions Am performs the kst except for cases of large Approach


noise where the ANNs seem to be preferred. For strongly We fiist tried to predict ~ using all the in&pendent variables
nmlinear tictions and low to mcderate noise, Am is still the for all 255 RSCtVOkS.The R2( ER,ACE, ER A.m~) = 0.48 when
bes4 predictw, although the R2 vatues are not particularly 127 randomly selected reservoirs m’elumped together. R* is
large, and the differenms between the ANN and ACE small because a large range of geologic and petrophysical
predictions may not be significant. For strongly nonlinear information are incorpomted in the same data set. ~ls resdta
functions with large noise, no prediction technique works in a poor correlation.
well, as expected. We also applied ANNs as an attempt to eliminate the
For all the cases expressed by Table 1, ACE is the noise and over-e nonlinearity by developing more complex
best or near the best procedure except for one case. For this correlations among the independent variables. Out of 255
reason we use ACE to interpret the ATMOR in the following reservoirs, we randomly selected 127 rmoirs for testin the
section. ANN. The coefficient of determination in this case is R4(E%
ANN, HR ACM) = 0.36. The results f~ the Prtilction of ERin
Application the 127 re=voirs using the ANNs and ACE techniques are
In this section we analyze the Atlas of Major Texas Oil cmpared in Fig. 6. In an effort to reduce this noise, we
Reservoirs’ database compiled by The University of Texas continued to work on groups of drive mechanims w
Bureau of fionomic Oeology. The objective is to develop a lithologies as given in the AMTOR.
correladon, using the ACE correlation technique, between The most common lithologies among these reservoirs
recovery efficiency and available petrophysid and geological are dolomi~, sandstone, and limestone. The drive
information given in the atlas and thereby ~tter understand mechanisms are solution gas drive, water drive, gas cap
the factors affecting oil recovery. expansion followed by water drive, and solution gas drive
followed by the partial water drive.
Atlas of Major Texas Oil Reservoirs Database We also tried to determine the most important
The Bureau of Economic Geology has compiled geologic, independent variables (XJs) that affect the ER. First, we
engineering and volumetric data for nearly 450 of the most chwked the correlations between the transformed dependent
productive oil reservoirs in Texas. Reservoirs described in the variables and a transformed independent variable. We found
AMTOR have each produced at least 10 million barrels of oil that no linear relationship is observed between transformed
and have adequate data available for useful geologic Xi’s and Y. Second, we calculated the variance of
description, as well as for the estimation of recovery transformed indeWndent variables Xi’s with the idea that if Xi
efficiency. The basic information about each mwrvoir has a small effect on Y then the Var(X[) will he a small
includes field and reservoir name, year of dlxovery, lit.hology, number. However, this procedure for determining the most
trap type, and drive mechanism. m rese~oir and fltid important variable was unsuccessful because good correlations
properties, depth, height of the oil column, porosity, were obtained even tier ignoring the variables having large
permeability, initiat water saturation, API gravity, initird gas- variance. Therefore, we decided to do a step-wise analysis for
oil ratio (OOR), inidal reservoir pressure, temperature, and all possible combinations of variables in each category of
well spacing, are also summarized in this database. Along drive mechanism and Iithologies.
with this, the data =t atso includes volumetic data, which
consists of residual oil saturation in the swept portion of the Step-wise analysis
resemoir, oil in plain, cumuktive production through 1981,
proj~ted ultimate recovery, and estimated recovery The step-wise analysis with ACE determines tie effect of each
efficiency. independent variable (reservoir property) on a dependent
variable (~) by predicting the dependent variable based on
only that one variable. The coefficient of determination, R2
Recovery Efficiency Prediction
(E% ACE, ER AcIuaI),is calculated for tie Prtiiction based on
Out of 450 reservoirs in the AMTO~ we selected the 255 each variable. The two indepen&nt variables having the
reservoirs that have the most complete data set for ftier largest R’ are grouped together for a succesive prediction ( a
analysis. We selected lithology, drive type, net pay, height of cumulative analysis) of ~. Each other variable is then
the oil column, average porosity, average peability, initiaI introduced in order of descending R2. The idea behind the
water saturation, GO~ APL initial pressure, temperature,and stepwise analysis is that the single-variable correlation
well spacing as indeWndent variables. Recovery efficiency, coefficient wilt indicate the order in which variables should be
E~, cumulative oil produced divided by initiat oil in place, added to a mdtivariable correlation to arrive ac ultimately, the
tabulated in the ATMOR is the dependent vtiable. These smallest variable set that effects a good correlation.
values are in effect as of the date of the survey (1981); many
are probably larga that what we used her~ but the absence of
sensitivity to time in most of the results given below suggest
that the reported E~ are not much different from their ultimate
values.

459
4 LOVEENA KAPUR, LARRY W. LAKE, KAMY SE PEHRNOORI SPE 39S05

Analysis by Drive Type combination of initial water saturation, with permeability and
height of oil column is notable. An excellent multivtiable
ACE comlation is achieved with the mmbination of only six
Water ~ive Mechanism
variables (permeability, height of oil column, initial water
A reservoir is classified as “water drive” if there exists an saturation, GO~ tempertaure, porosity). The R’ with these six
active hydra~ic connection between the reservoir and a variables is 0.94. A slight improvement is noticed in each step
contiguous aquifer, or if water has been injected. An ERof 70 of adding other variable, increasing the final R*to 0.98.
to 80% is possible in some of water drive reservoirs. There
are 57 r~oirs in the AMTOR database under this Analysis by Lithology
classification.
Figure 7a shows the results of the single-variable
Limestone Reservoirs
ACE analysis for the water drive category. While porosity has
the largest R’, no single-variable R’ would be considered Limestone reservoirs are formed in bedded sedimentary
significant by iklf. Figure 7b shows the results of the deposita that chiefly consists of carbonates of dcium and
cumtiative analysis. The combinations (porosity, magnesium. Them are 33 of these reservoirs with complete
temperature) m both shown since temperatm and data in the AMTOR. Figure 10(a) demonstrates the msulm of
permeability were equally important in Fig. 7a. The the single-variable analysis and Fig. lo(b) shows the
multivariable ACE correlations does not become significant cumulative analysis, R’ is 0.56 for pressure in the single-
until the~ are 9 independent variables in the procedw, variable analysix whereas, R2 is 0.38 for the (pressw, API)
adding the final variable, height of the oil column, has very combination in the cumulative tiysis. This implies that
little cmmdative effect. The fd R’ is 0.89. (pressure, API) has mom complex comlation than pressure by
itself. A significant effect is observed for the combination of
Solution Gas and Partial Water Drive Mechanism (pressure, API, permeability, net pay); and the R’ for the best
possible correlation is 0.96.
Solution gas drive refers to the displacement of the oil by gas
coming out of solution because of the reduction of reservoir
DoIomiti Reservoirs
pressure upon the withdrawal of fluids. Water drive
mechanism usually occurs as gas and oil are produced. The Dolomite reservoirs are formed in xnagnesian limestone.
complete data set is available only for 19 solution gas and Them are 63 dolomite reservoirs in AMTOR database. The R’
partial wat~ drive rmrvoirs. for the ~ prediction correlation using the ten selected
Figure 8a shows the results of the single-variable reservoir properties is only 0.41 in this case. The low value of
analysis for the solution gas and partial water drive category. R’ either implies the prewnce of the more complex comlation
Porosity and permeability have the largest R* and well spacing among the reservoir properties or the need of additional
has very small R2. Also, the single-variable R2 value for information. We selwt~ average depti as an additiond
parameter and a slight change in R’ was noticed. Finally, we
height of oil column is very close to net pay value and the
estimated a new meter, dimensionless time, for more
value for pressure is close to GOR value, but no single-
complete information. Good correlations were obtained
variable R2 is significant by itself. Figure 8b shows the results
without the consideration of dimensionless time in other
of the cumulative analysis. A gd mdtivariable ACE
We have assumed several parameters for
categories.
correlaticm is obtained with 7 independent variables in the estimating the dimensionless time. See the appendix for
procedure with R2 value of 0.83; excellent results are obtained discussion and definition. A better correlation is obtained
when tem~ature is added as the eighth independent variable, after considering dimensionless time but the R’ is increased to
increasing R2 to 0.95. A slight increase in R2 is observed on only 0.69 after this addition. The results m summarized in
adding the initial Wata saturation and is decreased to 0.92 Fig. 11.
upon addition of well spacing.
Sandstone Reservoirs
Gas tip and Water Drive MechaAsm
Sandstone reservoirs are composed of compacted detrital
In gas cap drives, as production proceeds and reservoir sedhnenL predominantly quartz grains. There are 119
pressure declines, expansim of a gas cap displaces oil sandstone reservoirs with mmplete data in the AMTOR.
downward toward the wells. At the same time, the gas cap Flgnre 12 illustrates the results for this case. R2 is only 0.44
retards the pressure decline and hence sustains oil recovery. with 10 independent variables. The correlation improves upon
As in the above case water-drive mechanism takes place as the addition of average depth and dimensimless time,
gas and oil is produ~. We have analyzed 30 of these increasing the finat R2to 0.52.
reservoirs.
Figure 9a shows the results of the single-variable Comparison of Categorical Classifications
analysis for the gas cap and water drive category. When reservoirs m classified according to drive type, the
Permeability has the largest R2, but no single-variable R’ range of R’ for ~ ptiiction using ten Eservoir properties is
would be considered significant by itself. ~gufe 9b ShOWS between 0.85-0.98. When classified according to Iithologie$
the results of the cumdative analysis. The consequence of the the R’ is 0.96 for limestones and only 0.41 and 0.44 for

460
SPE 39305 MODELING NONLINEAR PHENOMENA 5

dolomites and sandstonesusing ten rewrvoirproperties, A P~ . pressure at the wellbore, psig


slightly better R*is obtained in limestones and dolomites when q = estimated flow rate
two more variables (avearge depth and dimensiordess time) R’ = coefficient of determination
w introduced. The R* for dolomite increases to 0.69 and for r. = (well spacing/n) *n,acres
limestone is 0,52. Based on the comparison of R’ values, the r. = wellbore radius fi
best co=tations m obtained only when reservoirs w = timq years
classtiled according to drive mechanism. This an approximate bar = variance
comparison because the number of the reservoirs used for x,X1 = independent or Prdlctor random
developing the correlations in each category m differen~ variable
however the nmnber of reservoir properties uwd for x: = transformed independent or
developing the comlation is the same. predictm random variable
Y= dependent or response random
Conclusions variable
The following conclusions can be drawn from this study z,,& = random variables
PI = re~ession model parameter
1) Optimat transformations gives good correlations for a p, = regression model parameter
moderate amount of noise and a smatl degree of =
nonlinearity. Artificial neural network works are random noise or errcx
slightly better than optimal transforms for a system ;= average porosity
with large noise and moderate nonlinearity. 0s2 = noise variance
2) None of the correlation techniques is awurate for
large noise and extreme nonlinearity. Subscripts
3) Classification of reservoirs is nmssary for ACE = alternating conditional expectation
development of a good correlation of recovery Actual = actual
efficiency using the optimal transform method. ANN= artificial neural networks
4) Initiat gas-oil ratio (GOR), temperature net pay, and D= dimensionless
API are the most important parameters for solution LIN = linear regression
gas and partial water drive mdanism. Well spacing, Predicted= predicted
GOR and permeability are most immt for water
drive mechanisms. The most important variables for Operators
gas cap expansion and water drive mechanism are E(o) = mathematical expectation
initial wata saturation and GOR.
(5) Net pay, GO~ initial water saturation, and Acknowledgements
permeability are important factors for good
correlation in limestone reservoirs. The average Larry W. Lake holds the W.A (Monty) Moncrief Centennial
depth of the reservoir and time are important Chair at The University of Texas at Austin. The authors
variables for obtaining correlations in dolomite and thank F. Jerry Lucia of The University of Texas at Austin
sandstone reservoirs. Bureau of Economic Geology for valuable discussions on the
(6) GOR is important fm every category except dolomite Atlas of Major Texas Oil Reservoir database. We also
and sandstone reservoirs. Net pay and GOR are acknowledge Akhil Datta-Gupta of Texas A&M U. for
important for both limestone as well as solution gas supplying the ACE sofiware used in this work.
and partial water drive mechanism.
(7) The categorical classification according to drive Appendix - Definition of Dimensionless Time
mwhanism gives better collations than The dimensionless time is estimated using the following
classification by Iithologies. relation:
qt
Nomenclature t~=—
fluid vimsity, cp A@ L
~=
= well spacing, acres
B = formation volume factor, (RB/STB) We estimated the variables in q from the information given in
Cov = Covariane the database. Therefore, in this case
E~ = recovery efficiency q = estimated flow rate (given in next step)
h = net pay, ft t = AMTOR time(1981) – Discovexy thne (given in the
K= degree of nonlinearity database)
k = average permeability, mD A = well spacing (acres)
L= net pay, ft L = net pay (ft)
Pe = avemge pressure, psig $ = average porosity

461
6 LOVEENA KAPUR, LARRY W. LAKE, KAMY SEPEHRNOORI SPE 39305

The flow rate is estimated=

= o.oo708M(Pe – Pw)
9 Table 1. Results from numerical experiment.
pBln(~)
Degree of nonlinaity + Weakly lModerate Iarge
Oczlcryz J Noise~ I K=l K=1OO
k = average permeability (mD) Model
h = net pay (ft)
Low R2(LIN) 1.00 0.91 0.36
Pe = average pressure (psig) 0.2 /ay2 <o,11 R2(ACE) 0.99 0.99 0.49
Pw= pressure at the wellbore ( assumed zero)
R2(ANN) 0.99 0.89 0.40
= fluid viscosit (cp)
k = formation vo rume factor (RB/STB) Moderate R2(LIN) 1.00 0.84
4 ---
0.36
..-

(assumed 1.0 RB/STB) 0.11< is? Icryz <0.98 R2(ACE) 0.94 10.98 0.47
r, = generally, an external drainage radius (ft) R2(ANN) 10.99 ]0.86 0.40
(assumed =(well spacing/n) *n) Large 0.05
rW = wellbore radius ( assumed 0.5 ft) ,_a,z IGYZ20.98
. :::) % 0.08
IR2~A~ I 0.99 10.95 ~0.06
References
1. Bishop M. Christopher Neural Networks for Pattern The R2 values indicated are the sample values between Y~Cmd
Recognition, Oxford University ~ss, New York from Eq. 2 and Y~O~~l.Best model is shown in bold.
(1995).
2. Guoping Xue, Akhil Datta-Gup@ Peter Valko, and
Tom Blasingamti “ Optimal Transformations for
Multiple Regression: Application to Permeability
Estimation from Well Logs,” SPE/DOE 35412 1.00
presented at Impmved Oil Recovery Symposium,
TUIW Oklahom&l 15-130,(April 1996).
3. Hertz, J., Krogh, A., and Patmer, R.G., Intiuction
to tie ~eory of Neural Computation: Addison -
Wesley Publ. Co. (1991).
4. Isaaks, E,H. and R.M. Srivastavx Applied
Geostatisticsi Oxford University ~ss, New York
(1989).
5. Jerry L. Jenwn, Larry W, Lake, Patrick W,M.
Corbet6 and David J. Goggin: Statistics for
Petroleum Engineers and Geowientists, Prentice Hall
~ (1996).
6. Kapur Loveena: Ph.D. dissertation in progress, The
University of Texas at Austin (1998).
7. Leo Bmiman, and Jerome H. Friedman: “Estimating 0.00 11 ill I 1 I I I
o 5 10 15 20 25
Optimal Transformations For Multiple Regression &~ of Nonlinearity (K)
and Correlation”, Journal of the American Statistid Figure 1, RelstioIIslIip btween coefficient of detcmi=tloo ●nd
Association, Vol. 80, No. 391, 580-602, (September
1985).
8. Mark H. Holtz, Noel Tyler, C,M. G-tt, Jr., W.G.
White, and N.J. Banta: Atlas of Major Texas Oil
Reservoirs, Bureau of Economic Geology, The
University of Texas at Austin, Austin (1991).
9. Stephen, J.J., Neural Network Design and
Complexity of kaming: MIT Press (1990).

462
SPE 39m5 MODELING NONLINEAR PHENOMENA 7

1.Co

0.80

0.60
1.W

*
> 040

0.50

0.20

O.m

- O,al
..1. ~

-0.s0
o,(h)
..:

0.20
1

0,40
1
0.60
1

0.80
I

1.00
1 .0.20
O.co O.a 0.40 O.m O.m 1 .OJ
..
x
x
Figure 2. ComPrison of comlalion lechniquca for K.1 sad Flgure5. tipwbmdcwrclaUm &hnlquufw K4md
U**= 0.04. o,’= 0.85.

1,m .! ,-. -

0.s0 --

i- ! :+
0.60 . . .. . . . .. . . . . . .

!:
i!
0.40

0.20 . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . ..i. .. . . . . . . . . . . . ...{. . . . . . . . . . . . . . . ..j. . . . . . . . . . . . . . .

-1s0 . . . . . . . .. . . . . . . . . . . . . . . . . ..+... } . . . . . . . . . . . . . . . . . .. . . . . . .. . . . .. . ..i . . . . . . . . . . . . . . . . .

:. ‘.”:
. ;.
-Z.m 1 # 1 1 O.m
Osnl 0.20 0.40 0.60 0,s0 1.m
x
Flgurr
3.Cunbon ticmmlatlon ~hnlquu far K=l md
0“= O,u,

1.20
.—
1,m

0.80

0.60

*
0.40

O,qo

0.00

.o,~o

-0,40 r
0.00 0.20
L 1 1 1 i
0.40 0.60 0.80 1.m
x
Figum4. Comptison ofcomlatlou techniqumfor K4snd Fiwm 7(9). Mngk-m*bk smlyak of -h Ppcrty for mtcr drive
=t’vein.
O*= 0.01.

463
8 LOVEENA KAPUR, LARRY W. LAKE, KAMY SEPEHRNOORI SPE 39W5

Pasmmm

drive rmemdm.

464
SPE 39M5 MODELING NONLINEAR PHENOMENA 9

0.00
metal

Fiwrc 11. Step-tic anm~is for dolomite wcmolm.

O.CQ

465
View publication stats

You might also like