You are on page 1of 13

ARTICLE IN PRESS

Computers & Geosciences 31 (2005) 263275


www.elsevier.com/locate/cageo

Lithology identication of aquifers from geophysical well logs


and fuzzy logic analysis: Shui-Lin Area, Taiwan$
Bieng-Zih Hsieh, Charles Lewis, Zsay-Shing Lin
Department of Resources Engineering, National Cheng Kung University, Tainan, Taiwan

Received 7 July 2003; received in revised form 7 July 2004; accepted 16 July 2004

Abstract

The purpose of this study is to construct a fuzzy lithology system from well logs to identify formation lithology of a
groundwater aquifer system in order to better apply conventional well logging interpretation in hydro-geologic studies
because well log responses of aquifers are sometimes different from those of conventional oil and gas reservoirs. The
input variables for this system are the gamma-ray log reading, the separation between the spherically focused resistivity
and the deep very-enhanced resistivity curves, and the borehole compensated sonic log reading. The output variable is
groundwater formation lithology. All linguistic variables are based on ve linguistic terms with a trapezoidal
membership function.
In this study, 50 data sets are clustered into 40 training sets and 10 testing sets for constructing the fuzzy lithology
system and validating the ability of system prediction, respectively. The rule-based database containing 12 fuzzy
lithology rules is developed from the training data sets, and the rule strength is weighted. A Madani inference system
and the bisector of area defuzzication method are used for fuzzy inference and defuzzication. The success of training
performance and the prediction ability were both 90%, with the calculated correlation of training and testing equal to
0.925 and 0.928, respectively. Well logs and core data from a clastic aquifer (depths 100198 m) in the Shui-Lin area of
west-central Taiwan are used for testing the systems construction. Comparison of results from core analysis, well
logging and the fuzzy lithology system indicates that even though the well logging method can easily dene a permeable
sand formation, distinguishing between silts and sands and determining grain size variation in sands is more subjective.
These shortcomings can be improved by a fuzzy lithology system that is able to yield more objective decisions than
some conventional methods of log interpretation.
r 2004 Elsevier Ltd. All rights reserved.

Keywords: Groundwater; Aquifer characterization; Hydrogeology; Articial intelligence; Soft computing

1. Introduction

Fuzzy logic analysis of well logs has been recently


$ applied extensively in many reservoir characterization
Code available from server at http://www.iamg.org/CGE-
ditor/index.htm.
studies. For example, Fung et al. (1997) applied a self-
Corresponding author. Tel.: 886-6-275-7575  62825; fax: generating fuzzy rule extraction and inference system to
886-6-238-0421. the prediction of petrophysical properties from well log
E-mail addresses: wink@ms31.url.com.tw (B.Z. Hsieh), data, whereas Huang et al. (1999) presented a useful
zsaylin@mail.ncku.edu.tw (Z.S. Lin). fuzzy interpolator for permeability prediction based on

0098-3004/$ - see front matter r 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cageo.2004.07.004
ARTICLE IN PRESS
264 B.Z. Hsieh et al. / Computers & Geosciences 31 (2005) 263275

well logs from the North West Shelf in Australia. Fuzzy successfully used to determine hydrocarbon sediment
logic has also been used to determine hydrocarbon lithology (Cuddy, 2000). Similar to conventional com-
formation lithofacies and permeability from well log puterized well log analyses, fuzzy logic allows all
data in the southern North Sea (Cuddy, 2000). Cuddys pertinent log data, core analyses, mud analyses, etc. to
results gave near-perfect differentiation among aeolian, be examined simultaneously by the interpreter.
uvial, and sabkha rock types (the major lithofacies in Although the fuzzy logic method uses the same data as
several North Sea elds) from basic logs such as gamma- conventional log analysis, it is unlike conventional
ray (GR) and porosity logs. The techniques of fuzzy analyses which still demands qualitative determination
logic analysis from well logs can be applied to both of lithology. Instead, fuzzy logic adopts a set of rules
consolidated and unconsolidated sediments, as well as insuring objectivity in determination of soil/rock type
for water applications in oil exploration. whilst incorporating the expertise of the interpreter.
Although conventional geophysical well logging is an The purpose of this study is to construct a fuzzy
ideal method for hydro-geologic studies involving lithology identication system based on the GR log, the
aquifer characteristics, such as porosity and hydraulic resistivity logs, and the sonic log from the Shui-Lin area
conductivity (Temples and Waddell, 1996; Lin et al., of Taiwan to identify formation lithology of a ground-
1997), the identication of aquifer lithology from well water aquifer. The fuzzy logic lithology identication
log data depends upon the ability to distinguish between system can provide a more objective approach for log
soils/rocks with grain sizes varying from clay to gravel, analysts in determining lithology in the gray areas
and this method is still largely subjective in the absence (areas involving clastic rocks with grain sizes between
of core data. Well logging also provides in situ and sand and shale) of the system of interest.
continuous data, as well as yielding a number of
economic benets by saving the cost and time of core
analyses. However, well logging is limited because
lithology identication is still a subjective task that
depends largely on the experience of the log analyst. 2. Basic Theory
Identication of hydrocarbon formation lithology
from geophysical logs commonly employs lithology 2.1. Conventional Well Log Analysis
crossplots (such as MN lithology plot which requires
a sonic log, density log, and neutron log) or the The hydro-physical logs used in this study are: (a) the
combination gamma-ray neutron-density log method GR, (b) borehole compensated sonic (BHC) with sonic
(Asquith and Gibson, 1982). However, consideration porosity (SPHI) curve, (c) spontaneous potential (SP),
must be given to the idea that groundwater aquifers can and (d) phasor induction (PI). The PI includes four
be contaminated by the radioactive sources required for curves: medium very-enhanced resistivity (IMER), deep
these two types of logs, and the large hydraulic very-enhanced resistivity (IDER), spherically focused
conductivities might create an adverse environment for resistivity (SFLU), and apparent formation water
decentralized neutron and density logs (Peng, pers. resistivity (Rwa). The lithologic results of core analysis
comm., 2003). Furthermore, the lithologies involved in were also used in this study. This study limits the
water wells versus oil/gas wells might require different following explanation of conventional well log analysis
log suites. basic theory to clastic sedimentary rocks, focusing on
Identication of groundwater (shallow aquifer) for- shales and sandstones and the different responses of
mation lithology from well log data largely depends on logging tools to salt water versus fresh water zones.
expert experience and rather subjective rules, such as, The following describes the basics of the log types used
IF the natural GR reading is high and the separation in in this study to acquaint the general reader.
readings between shallow formation resistivity and deep
formation resistivity is small, then the formation
lithology is probably shale (Chapellier, 1992; Hsieh,
1997). Moreover, groundwater aquifer systems involving 2.1.1. Gamma-ray (GR) log
rocks with grain sizes ranging from clay to sand are The GR log is designed to measure the natural
often characterized by well logging methods as simply: radioactivity of soils and rocks, and is particularly useful
sands (including ne-, medium-, or coarse-grained in distinguishing between shales and sandstones and in
sands: the major components of an aquifer) and shales determining depositional environments. Shales usually
(including silts, clays, and muds: the major compo- exhibit high GR readings if they contain sufcient
nents of an aquitard). This type of analysis from well quantities of accessory minerals containing isotopes like
logging is simple and subjective. One way to reduce this potassium (40K), uranium (238U) or thorium (232Th). On
subjectivity is with the fuzzy logic technique, a type of the other hand, sands normally exhibit low GR
articial intelligence (AI) technology that has been responses (Fig. 1).
ARTICLE IN PRESS
B.Z. Hsieh et al. / Computers & Geosciences 31 (2005) 263275 265

Fig. 1. Lithology determination from gamma-ray and resistivity logs.

Table 1
Average interval transit time and velocity in rocks (after Sheriff and Geldart (1995); Chapellier (1992); Dewan (1983); Asquith and
Gibson (1982))

Lithology Transit time, Dt (ms/ft) Velocity of matrix (ft/s) Velocity of matrix (m/s)

Clays 16762.5 600016,000 18304880


Shale 167.662.5 590016,065 18004900
Sandstone 66.751.5 15,00019,500 45755950
Limestone 47.643.5 21,00023,000 64007015
Dolomite 43.538.5 23,00026,000 70157930

2.1.2. Borehole compensated sonic (BHC) log with negative deections (with salt waterooilogas), whereas
porosity curve (SPHI) clean sandstones containing fresh water might have zero
Because of the overlap in velocities between sand- or even positive SP responses. If the formation water is
stones and shales (Table 1), the primary function of a fresher than the mud ltrate, the curve will show a
sonic log is seldom determination of lithology; however, positive deection, with the amount of deection
it can sometimes provide useful information regarding proportional to the difference in salinity between the
rock type and porosity, particularly if this log is used in formation water and mud ltrate.
conjunction with other logs. For clean sandstones
saturated with oil, salt water or fresh water, the sonic 2.1.4. Phasor induction (PI) log with apparent formation
log may give similar responses, but gas usually has a water resistivity (Rwa) curve
more pronounced effect on this log. The bulk compres- The most useful log in this study for distinguishing fresh
sional wave velocity in rocks is also heavily dependent water aquifers from salt water reservoirs is the PI log. It
upon porosity, that decreases the velocity, and the consists of curves for shallow, medium and deep
primary wave velocity may depend upon degree of resistivities, along with a curve for apparent formation
consolidation or packing as well. Generally, the velocity water resistivity of the uninvaded zone where the
of acoustic waves is slower in clays than in sandstones formation water is uncontaminated by mud ltrate.
(Table 1). Although induction logs do not work well in highly
conductive muds, they can be run in holes lled with air,
2.1.3. Spontaneous potential (SP) log oil, or freshwater muds. Aquifers tend to be more resistive
The secondary potential or SP log requires a than aquitards. For a well drilled with salt water based
conductive drilling mud for best results. According to drilling mud, the resistivity of the invaded zone, that
Asquith and Gibson (1982), the magnitude of SP consists of rock, mud ltrate, formation water (either salt
deection depends upon the difference in resistivity or fresh water), and possibly residual hydrocarbons, will
between the mud ltrate and formation water, and if generally be smaller than the resistivity of the uninvaded
these two uids have the same resistivity, there is no zone containing fresh water. For this situation, porous and
deection of the SP from the shale baseline. Clean permeable sandstones are characterized by a wide separa-
sandstones containing oil, gas and salt water have tion between the shallow (invaded zone) and deep
ARTICLE IN PRESS
266 B.Z. Hsieh et al. / Computers & Geosciences 31 (2005) 263275

(uninvaded zone) resistivity curves. On the other hand,


under the same conditions above, a shale would exhibit a
small separation between the shallow resistivity curve and
the deep resistivity curve (Fig. 1). If, however, the drilling
mud is fresh water based, the separation between the
spherically focused (shallow) resistivity curve and the deep
resistivity curve will be considerably less (the invaded zone
resistivity can be approximately equal to that of the
uninvaded zone since both contain fresh water) than if the
drilling uid were salt water based. The separation between
the two resistivity curves is therefore an important
parameter in lithology determination involving a ground-
Fig. 3. Linguistic input variable model. Each linguistic input
water aquifer, provided the type of drilling mud is known.
variable is constructed from ve linguistic terms: VL (very low);
L (low); M (medium); H (high); and VH (very high).
2.2. Fuzzy lithology system

Fuzzy set theory, a method to distribute linguistic variables. The linguistic input variables include GR,
fuzzy information by mathematics, distributes a set by DR, and Dt, which are some of the most important
using a membership function, and extends the concepts basic parameters in lithology identication of ground-
of classical set theory. Fuzzy logic can be dened as: a water formations. Every linguistic input variable in-
logical system that generalizes classical two-valued logic volves ve linguistic terms, such as very low (VL), low
for reasoning under uncertainty (Yen and Langari, (L), medium (M), high (H), and very high (VH) as
1999). Therefore, fuzzy logic theory eliminates the shown in the trapezoidal membership function (Fig. 3).
problem of two-valued logic reasoning in classical set The linguistic output variable is lithology, consist-
theory (Klir and Yuan, 1995). ing of ve linguistic terms: C (clay), Z (silt), FS (ne
The major procedures in a fuzzy lithology system sand), MS (medium sand), and CS (coarse sand). The
developed in this study include (i) fuzzication, (ii) fuzzy reference boundary of the output variable linguistic term
if-then rules database, (iii) fuzzy inference system and is dened as an exponent. The grain size range is
(iv) defuzzication (Fig. 2). During fuzzication, well presented by an exponential function (of the form 2n,
log data (such as the GR reading, the separation where n is a negative integer) (Table 2). From the
between resistivity curves (DR), and the interval transit range of grain size, the exponent n for the upper and
time (Dt)) are transformed to linguistic input variables lower boundaries is adopted to dene the reference
constructed by linguistic terms and a membership boundary of linguistic terms. The membership function
function. The fuzzy if-then rules database contains adopted for the linguistic output variable is a trapezoi-
several lithology identication rules; the form of dal membership function (Fig. 4).
lithology identication rules is constructed by if A, B
and C, then D where A, B, C, and D are fuzzy sets.
Fuzzy approximate reasoning is then determined by a
fuzzy inference system. A fuzzy lithology value is 2.2.2. Fuzzy if-then rules and rule-based database
obtained by a defuzzication method, and nally the The rule-based database consists of several general
lithology of groundwater formation can be determined. lithology identication rules. The format of the lithology
identication rule is
2.2.1. Fuzzification If GR is A, and DR is B, and Dt is C, then
The fuzzy lithology system in this study contributes lithology is D.
the linguistic variables from the original domain Where GR, DR, and Dt are linguistic input variables;
lithology is the linguistic output variable; A, B, C are
linguistic terms of input variables (VL, L, M, H, or VH);
Fuzzy if-then
rules database and D are the linguistic terms of output variables (C, Z,
FS, MS, or CS).
The number of lithology identication rules depends
Fuzzification Defuzzification on the training data. For example, if all combinations
Fuzzy inference
Linguistic
system
Fuzzy between every two input variables are considered, the
Well log input lithology Lithology
rule-based database consists of a total 125 (=53) if-
reading variables value
then rules. Therefore, an appropriate reduced rule-
based database must be incorporated into the system
Fig. 2. Fuzzy lithology system. training step.
ARTICLE IN PRESS
B.Z. Hsieh et al. / Computers & Geosciences 31 (2005) 263275 267

Table 2
Grain size range of matrix and reference boundary setting

Linguistic term of output variable Grain size range of matrix (cm) Reference boundary of linguistic term
0 * 1
CS (coarse sand) 2 4AGS 42 [0, 1]
1 2
MS (medium sand) 2 4AGS42 [ 1, 2]
2 4
FS (ne sand) 2 4AGS42 [ 2, 4]
4 8
Z (silt) 2 4AGS42 [ 4, 8]
8 12**
C (clay) 2 4AGS42 [ 8, 12]
*
AGS=Average grain size.
**
2 12 represents the value of zero.

southwest of Yun-Lin, Taiwan (Fig. 5). The area is part


of the south branch of the Chou-Shui River alluvial fan
system, whose deposits consist of unconsolidated sand,
silt, and clay from the Chou-Shui River and its
tributaries. The upper section of the alluvial fan consists
primarily of gravel deposits, whereas the lower section
(Shui-Lin area) consists mainly of sand or clay. The
interbedded shale aquitard and the sand aquifer were
deposited because of alternating transgression and
regression. All of the sedimentary formations in the
investigation area are Pleistocene-Recent in age. In the
Shui-Lin area, the shale materials (silt and clay) are
Fig. 4. Linguistic output variable lithology constructed from aquitards, and the sands (FS, MS, and CS) are aquifers.
ve linguistic terms: C (clay); Z (silt); FS (ne sand); MS
(medium sand); and CS (coarse sand). 3.2. Data Collection

The geophysical logs from SL-2 well used in this study


2.2.3. Fuzzy inference system (Fig. 6) include the GR log, a PI log consisting of three
A Madani inference system was chosen for this study. usual resistivity curves (the medium very-enhanced
Madani fuzzy inference uses a linguistic reasoning curve was not used in this study) plus an Rwa curve (not
process that has been extensively applied to engineering shown in Fig. 6), and a BHC log with SPHI curve
studies (MATLAB, 2001). Because the output variable (porosity curve not shown in Fig. 6). Lithologic types
in this study is dened by fuzzy sets, the process of fuzzy from core analyses from the SL-monitoring well include
reasoning belongs to a type of linguistic reasoning. C, Z, FS, MS and CS.
Therefore, the Madani inference system is an appro- Both the geophysical logs from SL-2 well and the core
priate method for this study. analysis lithology from SL-monitoring well represent
continuous data over the depth range from 100 to 198 m.
2.2.4. Defuzzification Also, these two wells are located very close to each other
The input for the defuzzication process is a fuzzy set (Fig. 5). The distance between the two wells is about
and the output is a crisp set. The purpose is to derive a 400 m. Because of their close proximity, it is assumed
crisp value which can represent the result of fuzzy sets in that their lithologies and depths are equivalent.
linguistic output variables. The bisector of area method is
introduced into the defuzzication process (MATLAB,
2001). This method bisects the aggregate output area and 4. Procedure
obtains the output crisp value from the center of the area.
4.1. Data digitization

3. Case study For the SL-2 wells, log curves were read every 2 m for
the depth range from 100 to 198 m (drill depths with
3.1. Regional geology ground level equal to 7.1 m) and then converted to
digital data sets. A total of 50 data sets were digitized
The Shui-Lin area, used for identifying lithologies of a (Table 3). Every log data set included the GR log, the
groundwater aquifer system in this study, is located SFLU and the IDER curves from the PI log, and the
ARTICLE IN PRESS
268 B.Z. Hsieh et al. / Computers & Geosciences 31 (2005) 263275

Fig. 5. Study area and well locations.

BHC log. The input parameters used in the fuzzy rules, and by constructing the rule-based database. The
lithology system, GR and Dt, were directly taken from test data sets were used to validate the ability of system
the digitized data of the GR log and the BHC log, prediction. Based on the 80/20 rule for the total of 50
respectively. The other input parameter, DR, is the value data sets, the amount of the training data sets and the
of the IDER (deep) curve reading minus the SFLU test data sets were 40 and 10, respectively. The following
(shallow) curve reading. steps are necessary to extract the 10 test data sets: (1)
According to the core analysis from the Central arrange all sets in order of increasing depth (Table 3); (2)
Geological Survey of Taiwan, the lithology of the Shui- choose a random depth value among the 50 data sets
Lin area groundwater aquifer includes C, Z, FS, MS, (122 m was chosen at random in this study); (3) pick up
and CS. In the fuzzy lithology system, the lithology type the test data sets every 10 m spaced in the up direction
must be converted to a crisp set for system mathematical from the chosen depth value (in this study, the testing
estimation. A code number from 1 to 5, ranging from data sets started from 122 m by random choice, the 112
coarse sand to clay, respectively, was assigned for each and 102 m were extracted in the up direction) (Table 3);
lithology (Table 4). (4) pick up the test data sets every 10 m spaced in the
Therefore, the input variables used in the fuzzy down direction from the chosen depth value (in this
lithology system were GR, DR, and Dt. And the study, based upon the 122 m depth selected by random
output variable used in the fuzzy lithology system was choice, 132, 142, 152, 162, 172, 182, and 192 m were
lithology. For the depth interval from 100 to 198 m, a extracted in the down direction) (Table 3). Step (3)
total 50 datasets were collected and digitized (Table 3). involving depths of 100, 102, 104, 106 and 108 m (chosen
by random) can be ignored because no test data sets can
4.2. Data cluster be found in the up extracted direction. Step (4) can be
ignored for depths of 190, 192, 194, 196 and 198 m
The 50 data sets were clustered into two parts: (chosen by random) because no any test data sets can be
training data sets and test data sets. The training data found in the down extracted direction.
sets were used to construct the fuzzy lithology system for By the way of test data set extraction, 10 test data sets,
Shui-Lin area by carefully adjusting the fuzzy sets of the at depths of 102, 112, 122, 132, 142, 152, 162, 172, 182,
fuzzy input variables, by reducing the fuzzy lithology and 192 m, were extracted. The reason for not extracting
ARTICLE IN PRESS
B.Z. Hsieh et al. / Computers & Geosciences 31 (2005) 263275 269

Fig. 6. Geophysical logs from SL-2 well and core analysis from SL-monitoring well.

all test data sets by random is to avoid the test data sets and VH, respectively. The membership function
from becoming too concentrated in some depth intervals. adopted for linguistic input variable analysis is a
trapezoidal membership function.
4.3. Fuzzy lithology system construction The linguistic output variable is lithology, which is
differentiated by ve linguistic terms: C, Z, FS, MS and
Based on the 40 training data sets, the linguistic input CS, respectively (Fig. 4). From the reference boundary
variables, including GR, DR, and Dt, can be dened (Table 2) and the trapezoidal membership
constructed (Figs. 79). Every linguistic input variable is function, an output fuzzy set, lithology, can be
based on ve linguistic terms: VL, L, M, H, constructed (Fig. 4).
ARTICLE IN PRESS
270 B.Z. Hsieh et al. / Computers & Geosciences 31 (2005) 263275

Table 3
Digitized well log data and resulting lithologies

Depth GR* DR* Dt* Lithology** Depth GR DR Dt Lithology

100 72 10 182 4 150 82 3 177 4


102 67 20 171 3 152 87 11 182 5
104 77 5 176 3 154 88 4 183 5
106 75 9 180 4 156 88 8 184 5
108 78 17 176 4 158 90 6 190 5
110 83 12 179 4 160 79 9 170 4
112 82 14 179 4 162 69 18 165 3
114 79 10 178 4 164 68 21 166 3
116 79 5 174 4 166 82 2 155 4
118 78 13 174 4 168 82 10 174 4
120 81 11 178 4 170 78 9 178 4
122 68 15 180 3 172 74 12 173 3
124 67 16 177 3 174 82 13 170 3
126 72 14 178 3 176 82 10 179 4
128 71 17 177 3 178 83 8 182 4
130 76 14 176 3 180 83 11 176 4
132 77 14 176 4 182 85 13 177 5
134 78 10 178 4 184 79 10 177 4
136 76 11 178 4 186 75 13 171 3
138 83 9 177 4 188 75 23 168 2
140 81 7 182 4 190 73 11 176 3
142 82 7 176 3 192 69 14 175 3
144 68 23 175 3 194 67 12 167 3
146 60 20 177 3 196 53 29 179 1
148 67 14 173 3 198 62 19 178 3
*
Digitized well logs: GR: Gamma-ray log reading, API; DR: Value of deep curve reading minus shallow curve reading, O-m; Dt:
Borehole compensated sonic (BHC) log reading, ms/ft.
**
Lithology abbreviation: 5 (clay); 4 (silt); 3(ne sand); 2 (medium sand); 1 (coarse sand).

Table 4
Lithology code used in fuzzy lithology system

Lithology (Shui-Lin area) Grain size range of matrix (cm) Correlated lithology code

Coarse sand 204AGS42 1 1


Medium sand 2 14AGS42 2 2
Fine sand 2 24AGS42 4 3
Silt 2 44AGS42 8 4
Clay 2 84AGS42 12** 5

A reduced fuzzy lithology rule-based database was method derives a crisp value which represents the result
developed from the 40 training data sets. The rule-based of aggregate output area. By using the reference
database contains 12 fuzzy lithology rules (Table 5), boundaries for lithology types (Table 2), a crisp set
which are all in the form of an if-then model. A rule derived from defuzzication can be converted to a
weighting concept was introduced in this study for specic lithology; thus, the lithology of the groundwater
carefully adjusting the rule strength (MATLAB, 2001). aquifer can be identied from the fuzzy lithology system.
Every rule can dene a rule weight, which is a number
between 0 and 1. A rule weight used in this study not
only reects the strength of the rule, but also expresses 5. Results
the relative importance between rules.
Fuzzy reasoning for all rules in this study was based By using the 40 training data sets, the specic fuzzy
upon a Madani inference system. After the process of sets of input variables were constructed. A fuzzy
output aggregation, the bisector of area defuzzication lithology rule-based database, containing 12 fuzzy
ARTICLE IN PRESS
B.Z. Hsieh et al. / Computers & Geosciences 31 (2005) 263275 271

of the training data sets into the trained fuzzy lithology


system. In this study, a total of 40 training data sets were
used to check the performance of system training.
The results of lithology from this fuzzy lithology
system (named fuzzy lithology) were compared with
the results of lithology from core analysis (named true
lithology). In Fig. 10, the square marks (also connected
by a line) represent the true lithology, the star marks
represent the fuzzy lithology, the vertical axis shows the
depth interval from 100 to 200 m, and the numbers from
1 to 5 on the horizontal axis represent the different
lithologies from CS, MS, FS, Z, and C, respectively.
Fig. 7. Linguistic input variable, GR, includes ve linguistic Thirty-six training data sets were identied correctly
terms (for denitions, see Fig. 3). from the total 40 training data sets (Fig. 10), for a
success rate of 90%.
In the performance validation of the system training,
all of the sand types (CS, MS, and FS) were successfully
identied (Fig. 10). Only four layers were not well
trained. Even though the training performance was not
perfect (success rate of 100%), but, in this study, the
real performance of the system depended on the
predictive ability as well; therefore, an appropriate
training performance to avoid over-training (means the
system had a perfect training result but poor predictive
ability) was considered. On the other hand, achieving
the best predictive ability of the system was the desired
target.
Fig. 8. Linguistic input variable, DR, includes ve linguistic
terms (for denitions, see Fig. 3). 5.2. Predictive ability of fuzzy lithology system (test
results)

Ten non-trained test data sets were introduced into


the fuzzy lithology system for validating the predictive
ability of the system. Nine test data sets were predicted
correctly from the total 10 testing data sets (Fig. 11) with
90% success. The predictive ability of 90% is considered
high, and only one silt type was predicted incorrectly. It
is possible that heterogeneous and/or anisotropic con-
ditions existed at this depth between the two wells and
resulted in the wrong prediction of the silt zone. Another
possible reason could be due to some factors that were
not considered in this study such as lacking the SP log
Fig. 9. Linguistic input variable, Dt, includes ve linguistic information.
terms (for denitions, see Fig. 3).

lithology if-then rules with its specic rule weighting, 6. Discussion


was established in this study for the Shui-Lin area.
After the training work of the fuzzy lithology system The original well survey of the SL-monitoring well
was completed, the performance of training and the included the GR log, the short (spacing) normal (16")
ability of prediction were validated. resistivity and long (spacing) normal (64) resistivity
curves. Because mud recycling was not adopted in
5.1. Performance validation of fuzzy lithology system drilling, and the plastic casing was installed quickly to
training (training results) avoid well collapse in some depth intervals, the logging
data quality was of poor quality to identify lithology.
The performance validation was employed to check The vicinity well, SL-2, recycled the mud (GELMUD
the systems training performance by placing all or part consisting of brackish water and bentonite) during
ARTICLE IN PRESS
272 B.Z. Hsieh et al. / Computers & Geosciences 31 (2005) 263275

Table 5
Fuzzy if-then lithology rules after fuzzy system training

Rule 1 If GR is VL and DR is VH and Dt is (N/A)* Then Lithology is CS (1)**


Rule 2 If GR is L and DR is H and Dt is (N/A)* Then Lithology is MS (1)
Rule 3 If GR is (N/A)* and DR is H and Dt is M Then Lithology is MS (0.8)
Rule 4 If GR is M and DR is M and Dt is M Then Lithology is FS (1)
Rule 5 If GR is M and DR is M and Dt is L Then Lithology is FS (1)
Rule 6 If GR is (N/A)* and DR is H and Dt is H Then Lithology is FS (0.8)
Rule 7 If GR is H and DR is L and Dt is H Then Lithology is Z (0.6)
Rule 8 If GR is H and DR is M and Dt is H Then Lithology is Z (0.4)
Rule 9 If GR is H and DR is M and Dt is M Then Lithology is Z (0.4)
Rule 10 If GR is VH and DR is L and Dt is H Then Lithology is C (0.4)
Rule 11 If GR is VH and DR is VL and Dt is H Then Lithology is C (1)
Rule 12 If GR is VH and DR is VL and Dt is VH Then Lithology is C (1)
Abbreviation identify: VL (very Low) ; L (low) ; M (medium) ; H (high) ; VH (very high) CS (coarse sand); MS(medium sand); FS(ne
sand); Z(silt); C(clay).
(N/A)*: Rule did not use this component after system training (a reduced rule works here).
**
The rule weighting value.

Fig. 10. Comparison between true lithology and fuzzy lithology in training period.

drilling, and the well was logged by Schlumberger the invaded and uninvaded zones, thereby allowing the
Corporation. The log quality was sufcient to identify Induction Phasor log to operate. Results from the fuzzy
lithology. It should be noted that the mudcake resistivity lithology system for the SL-2 well were then correlated
(Rmc) and mudltrate resistivity (Rmf), both with with the core analysis lithologies from the SL-monitor-
values of about 4 O-m as determined from the log ing well in this study.
header, indicate that the drilling mud in the SL-2 well In the performance validation of the system training,
was just salty enough to act as a conductive uid for 36 training data sets were identied correctly from the
the SP to operate and to yield sufcient contrast between total 40 training data sets. The training performance was
ARTICLE IN PRESS
B.Z. Hsieh et al. / Computers & Geosciences 31 (2005) 263275 273

Fig. 11. Comparison between true lithology and fuzzy lithology in testing period.

90%. In the results of the system prediction ability (Rule 1) IF the GR reading is low, IF the separation
validation, nine test data sets were predicted successfully of deep and shallow resistivity curves (DR) is wide, and
from the total 10 test data sets. The predictive ability IF the interval transit time (Dt) is short, THEN the
was 90%, and it is considered high compared to Cuddys lithology of the formation is sand.
(2000) and Fung et al.s (1997) studies. Another way to (Rule 2) IF the gamma-ray GR is high, IF the
check the prediction accuracy for both training and separation of deep and shallow resistivity curves (DR) is
testing is based on the coefcient of correlation (Fung et narrow, and IF the interval transit time (Dt) is long,
al., 1997). A high value of the coefcient of correlation THEN the lithology of the formation is shale.
means the system results have high correlations to the The words used in the above sentences, low and high,
original core data. In this study, the calculated training mean the relative degree of GR reading. Usually,
correlation was 0.925, and the calculated testing maximum values of GR readings are used to infer a
correlation was also high (up to 0.928). In the (Fung shale base line, and minimum values will be used to set
et al.s (1997) study, the best training correlation was up a sand base line. The word low means close to the
0.917, and the best testing correlation was 0.865. sand base line, and the word high means close to the
Compared with their results (even though they had a shale base line. The words wide and narrow refer to
different output variable to be porosity) we can the relative degree of separation between the deep and
conclude that the prediction accuracy of our fuzzy shallow resistivity curves. The word wide means the
lithology system is acceptable. value is close to the maximum separation for the given
The lithologic results from core analysis, well logging depth interval, and the word narrow is just the
and fuzzy lithology are compared in Fig. 12, in which opposite. Also, the words short and long refer to the
the three major columns represent lithologies from the relative values of the interval transit times measured in
three methods. The groundwater formations between BHC log.
100 and 198 m (drill depths) in the SL-2 well were The well logging method can easily delineate a
divided into either sands or shales based on conven- permeable sand formation from log characteristics
tional well logging analysis of the log curve shapes and (Fig. 12), but identication of silts and determination
two basic rules: of sands with varying grain sizes (from coarse to ne)
ARTICLE IN PRESS
274 B.Z. Hsieh et al. / Computers & Geosciences 31 (2005) 263275

Fig. 12. Lithology results: comparison of core analysis, well logging, and fuzzy system.

are more subjective and difcult. This shortcoming can 100 to 198 m in this area, they were omitted. On the
be improved by our fuzzy lithology system analysis. other hand, this fuzzy lithology system cannot recognize
This studys lithology system included C, Z, FS, MS, a gravel lithology because the system did not include any
and CS, all of which are common in Shui-Lin area. experience while system training. In this case, the fuzzy
Because gravels are not found in the depth range from lithology system is not appropriate for gravel formations
ARTICLE IN PRESS
B.Z. Hsieh et al. / Computers & Geosciences 31 (2005) 263275 275

(like those close to the upper section of the Choushui groundwater monitoring network set up by Water
River alluvial fan). Furthermore, this study involves a Resources Agency, Ministry of Economic Affairs,
clastic aquifer rather than a carbonate aquifer; Taiwan. We also give special thanks to John Doveton
however, some studies have developed similar method and an anonymous reviewer for their valuable review
can be applied to carbonate reservoirs (Chang et al., comments that made our study more integrated.
1997; Cuddy, 2000). This might be of interest to middle-
Eastern oil eld analysts.

7. Conclusions References

Asquith, G., Gibson, C., 1982. Basic Well Log Analysis for
A fuzzy lithology system based on well logs from the
Geologists. AAPG Publications, Tulsa, OK 216pp.
Shui-Lin area of Taiwan was constructed for identifying Chang, H.C., Chen, H.C., Fang, J.H., 1997. Lithology
formation lithology with varying grain sizes of a determination from well logs with fuzzy association memory
groundwater aquifer in this study. The specic fuzzy neural network. IEEE Transactions on Geoscience and
sets of input variables were established, and a fuzzy Remote Sensing 35 (3), 773780.
lithology rule-based database containing 12 fuzzy Chapellier, D., 1992. Well Logging in Hydrogeology. A.A.
lithology if-then rules with its specic rule-weighting Balkema Publishers, Brookeld, MA 175pp.
was formulated for the Shui-Lin area. The conclusions Cuddy, S.J., 2000. Litho-facies and permeability prediction
are: from electrical logs using fuzzy logic. SPE Reservoir
Evaluation & Engineering 3 (4), 319324.
(1) The prediction accuracy of fuzzy lithology system Dewan, J.T., 1983. Essentials of Modern Open-Hole Log
was fairly good (90% for predictive ability and Interpretation. PennWell Publishing Company, Tulsa, OK
90% or better for the coefcient of correlation) 361pp.
based on the results of the testing performance, and Fung, C.C., Wong, K.W., Wong, P.M., 1997. A self-generating
the calculated coefcient of correlation of training fuzzy rules inference for petrophysical properties prediction.
In: Proceedings of the IEEE International Conference on
and testing.
Intelligent Processing System. Beijing, China, pp. 205-208.
(2) The compared lithologic results by core analysis, Hsieh, B.Z., 1997. Estimation of aquifers formation strength
well logging and fuzzy lithology show that the from well logging data. M.Sc. Thesis, National Cheng Kung
conventional well logging method can easily distin- University, Tainan, Taiwan, 134 pp (In Chinese).
guish a permeable sand formation from log char- Huang, Y., Gedeon, T.D., Wong, P.M., 1999. A practice fuzzy
acteristics, but identication of silts and interpolator for prediction of reservoir permeability. In:
determination of sands with varying grain sizes are Proceedings of the IEEE International Conference on Fuzzy
more subjective and difcult. As illustrated in this System. Seoul, South Korea, pp. III-1528III-1533.
research, our fuzzy lithology system can improve the Klir, G., Yuan, B., 1995. Fuzzy Sets and Fuzzy Logic, Theory
denition of grain size. Although there is some and Applications. Prentice-Hall, Englewood Cliffs, NJ
574pp.
subjectivity in the fuzzy lithology system, it enables
Lin, Z.S., Hsieh, B.Z., Cai, M.Y., Tang, Y.D., Lee, C.C., Chen,
the log analyst to make a more objective nal S.T., 1997. The Hydro-geologic characteristics estimation of
decision than by conventional well log analysis. subsidence area in Yun-Lin, Taiwan. In: Proceedings of the
(3) This methodology can be particularly useful for 2nd Meeting of the Groundwater Resources and Water
large aquifers involving multi wells where only a few Quality Protection. Tainan, Taiwan, pp. 853-864 (In
core analyses are available. Chinese).
MATLAB, 2001. Fuzzy Logic Toolbox Users Guide. The
MathWorks, Inc., Natick, MA 217pp.
Sheriff, R.E., Geldart, L.P., 1995. Exploration Seismology.
Acknowledgments Cambridge University Press, Melbourne, Australia 592pp.
Temples, T.J., Waddell, M.G., 1996. Application of petroleum
geophysical well logging and sampling techniques for
The authors thank Chinese Petroleum Corporation of evaluating aquifer characteristics. Ground Water 34 (3),
Taiwan for supplying logs from the SL-2 well. We 523531.
appreciate the core analyses furnished by Central Yen, J., Langari, R., 1999. Fuzzy Logic Intelligence. Control
Geological Survey of Taiwan, and the groundwater and Information. Prentice-Hall, Englewood Cliffs, NJ
aquifer information in the Shui-Lin area from the 548pp.

You might also like