ARTICLE IN PRESS

Computers & Geosciences 32 (2006) 947–964

www.elsevier.com/locate/cageo

Neural network prediction of carbonate lithofacies from well

logs, Big Bow and Sand Arroyo Creek fields, Southwest Kansas
Lianshuang Qi, Timothy R. Carr
Kansas Geological Survey, The University of Kansas, Lawrence, KS 66044, USA
Received 21 May 2005; received in revised form 26 September 2005; accepted 21 October 2005

Abstract

In the Hugoton Embayment of southwestern Kansas, St. Louis Limestone reservoirs have relatively low recovery
efficiencies, attributed to the heterogeneous nature of the oolitic deposits. This study establishes quantitative relationships
between digital well logs and core description data, and applies these relationships in a probabilistic sense to predict
lithofacies in 90 uncored wells across the Big Bow and Sand Arroyo Creek fields.
In 10 wells, a single hidden-layer neural network based on digital well logs and core described lithofacies of the limestone
depositional texture was used to train and establish a non-linear relationship between lithofacies assignments from detailed
core descriptions and selected log curves. Neural network models were optimized by selecting six predictor variables and
automated cross-validation with neural network parameters and then used to predict lithofacies on the whole data set of
the 2023 half-foot intervals from the 10 cored wells with the selected network size of 35 and a damping parameter of 0.01.
Predicted lithofacies results compared to actual lithofacies displays absolute accuracies of 70.37–90.82%. Incorporating
adjoining lithofacies, within-one lithofacies improves accuracy slightly (93.72%).
Digital logs from uncored wells were batch processed to predict lithofacies and probabilities related to each lithofacies at
half-foot resolution corresponding to log units. The results were used to construct interpolated cross-sections and useful
depositional patterns of St. Louis lithofacies were illustrated, e.g., the concentration of oolitic deposits (including
lithofacies 5 and 6) along local highs and the relative dominance of quartz-rich carbonate grainstone (lithofacies 1) in the
zones A and B of the St. Louis Limestone.
Neural network techniques are applicable to other complex reservoirs, in which facies geometry and distribution are the
key factors controlling heterogeneity and distribution of rock properties. Future work involves extension of the neural
network to predict reservoir properties, and construction of three-dimensional geo-models.
r 2005 Elsevier Ltd. All rights reserved.

Keywords: Neural network; Lithofacies prediction; Statistical analysis; Cross-sections; St. Louis Limestone

Corresponding author at: Chevron Energy Technology 1. Introduction and objective

Company, San Ramon, CA 94583, USA. Tel.: 1 925 842 6333;
fax: 1 925 842 2076. Since its discovery over 50 years ago, hydro-
Correspondence to. Tel.: +1 785 864 2135;
carbon production from St. Louis carbonate shoals
fax: +1 785 864 5317.
E-mail addresses: LIQI@chevron.com (L. Qi), in the Hugoton Embayment of southwestern
tcarr@kgs.ku.edu (T.R. Carr). Kansas has increased significantly, especially over

0098-3004/$ - see front matter r 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cageo.2005.10.020
 ARTICLE IN PRESS
948 L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964

the last decade (Morrison et al., 2002). The numerous studies in the petroleum industry. These
application of recent technology has had a great techniques have been successfully used in predicting
impact on exploration success for these strati- lithofacies from logs (Saggaf and Nebrija, 2003a;
graphic traps (e.g., three-dimensional (3D) seismic). Bohling and Dubois, 2003), estimating missing logs
However, the complexities of stratigraphy and (Saggaf and Nebrija, 2003b), and predicting poros-
relatively thin reservoir intervals (generally less than ity and permeability from 3D seismic (Wong et al.,
4.5 m) pose a challenge in continued exploration 1998; Trappe and Hellmich, 2000; Ali and Cha-
and efficient field development for St. Louis Lime- wathe, 2000; Lee et al., 2000; Ligtenberg and
stone reservoirs. Petrophysical properties in St. Wansink, 2001; Mathisen et al., 2003).
Louis Limestone reservoirs are mainly controlled In the Big Bow and Sand Arroyo Creek fields
by facies. Accurate discrimination of facies signifi- (Fig. 1), all available well logs were digitized and
cantly reduces the error in prediction of perme- assembled into a database. Using a neural network
ability and hydrocarbon volume (Dubois et al., approach to predict fine-scale lithofacies for improved
2003). Robust reservoir models provide insight into sequence stratigraphy interpretation and robust re-
the distribution of petrophysical properties, and the servoir characterization, core description from 10
external and internal architecture of St. Louis cored wells was related to digital well logs in
Limestone oolitic reservoirs. approximately 90 uncored wells in Big Bow Field
Because of economic constraints, it is uncommon (discovered in 1989; produced over 6.6 million barrels)
for a significant number of wells in a field to have and nearby Sand Arroyo Creek Field (discovered in
core across the reservoir. This presents a challenge 1996; produced over 0.1 million barrels).
in detailed interpretation of the lithofacies at the
well location, inference of stratigraphic geometry, 2. Geological background
and characterization and modeling the reservoir at
the field scale. In recent years, artificial intelligence The reservoir rock at Big Bow and Sand Arroyo
and neural network methods have been applied in Creek fields is composed of the St. Louis Limestone

Fig. 1. (A) Mississippian paleogeography of study area in Southwestern Kansas; (B) location of Big Bow and Sand Arroyo Creek fields,
which are part of a complex of oolite shoal reservoirs showing alignment across Hugoton Embayment. Latitude and longitude coordinates
of study area are marked southwest (36.9855,
102.04720); northeast (38.700,
100.2430).
 ARTICLE IN PRESS
L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964 949

(Meramecian Series). In the Hugoton Embayment, Hydrocarbons are produced from four informal
the St. Louis Limestone comprises approximately reservoir zones in the St. Louis (zones A, B, C, and
200 ft (60 m) of the Mississippian section. The St. D). Each zone represents a thin interval of skeletal
Louis thickens southward into the Anadarko Basin. oolitic limestone (Fig. 2). Stratigraphic traps formed
It is erosionally truncated to the north and east by the pinchout of depositional geometries of the
across the Las Animas Arch and Central Kansas thin skeletal oolitic deposits prevalent in St. Louis
uplift (Fig. 1). The upper part of the St. Louis fields are below seismic resolution, and are relatively
Limestone which is the focus of this paper has been difficult to target. It is expected that new accumula-
named the Stevens Member (Abegg, 1991, 1994). tions remain to be discovered in the St. Louis
The St. Louis Limestone was formed as part of an Limestone of the Hugoton Embayment. The recent
overall regressive succession during late Mississip- advancements in stratigraphic and reservoir model-
pian (Goebel, 1968; Lane and Dekeyser, 1980). The ing techniques provide an opportunity to better
St. Louis consists of eolianite deposits, made up of understand the complex internal and external geo-
well sorted carbonate grains with siliciclastic very- metry of St. Louis Limestone reservoirs. In this
fine-grained sand and silt, relatively thin skeletal paper, a neural network method was applied to
and oolitic limestone, peloidal limestone and some predict lithofacies from digital logs as the foundation
evaporite deposits (Carr and Lundgren, 1994). It for further stratigraphic and reservoir modeling.
was deposited across an extensive shallow-marine
carbonate shelf with periodic occurrence of a non- 3. Methods and approach
marine (eolianite) facies (Handford, 1988; Hand-
ford and Francka, 1991; Abegg, 1994; Abegg et al., The thin nature and lateral variability of subtidal
2001; Abegg and Handford, 2001). grainstone of the St. Louis makes it challenging to

Fig. 2. Late Paleozoic stratigraphy in Southwest Kansas showing informal subtidal reservoir zones of St. Louis Limestone.
 ARTICLE IN PRESS
950 L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964

recognize lithofacies patterns and correlate between Sand Arroyo Creek fields. The minimum suite of
wells since it is much more difficult to recognize digital well logs used for neural network lithofacies
carbonate lithofacies using wireline logs than prediction consists of gamma-ray (GR), resistivity
siliciclastics because of the difficulty in recognizing (deep and medium), density and neutron porosity.
micrite (carbonate mud). Neural network analysis Photoelectric (PE) logs were run in a large number
provides a method to extract additional information of the wells, and were used where available. A total
from the digital well logs in order to better recognize of 750 ft of cored section of the St. Louis Limestone
lithofacies changes. Artificial neural networks are from 10 wells was described (Qi and Carr, 2005a).
mathematical algorithms inspired by the working All 10 cored wells have complete logging suites of
pattern of human brains. A neural network works listed digital logs. Core interpretations include
as a learning process from provided information, lithofacies type, depositional fabric (described in
training the data to form certain patterns for each terms of Dunham, 1962 classification), fracture,
subject, then predicts targets with the output model. grain size, the type of porosity, color, bedding, fossil
In this study, we used a single-layer neural content, and interpreted water depth.
network to predict the lithofacies from digital logs
(Duda et al., 2001). It is one of the most common 5. Lithofacies classification and environments
types of neural network and is often classified as a
‘‘back-propagation’’ network (Oja, 1989). The basic Six major lithofacies were recognized on half-foot
elements are organized into layers: an input layer intervals in cores of the St. Louis Limestone. The
(corresponding to the selected digital well logs), a lithofacies classification, detailed core description,
hidden layer and an output layer showing the set of and interpretation of depositional environments are
lithofacies membership probabilities (Fig. 3). Basic similar to previous work (Handford, 1988; Hand-
1D and 2D statistical data analysis was used to ford and Francka, 1991; Carr and Lundgren, 1994;
select input parameters from available types of well Abegg, 1994; Abegg et al., 2001; Abegg and
logs for neural network modeling. A spreadsheet Handford, 2001). The six major lithofacies are: (1)
tool implemented as an add-in using Visual Basic quartz-rich carbonate grainstone deposited by
was used to accomplish the training, cross-validat- eolian processes adjacent to a coastline; (2) argillac-
ing, and batch processing of LAS (log ASCII eous limestone deposited in the deepest marine
standard) log files in order to predict lithofacies environment; (3) skeletal wackestone deposited in
from well logs. open-marine settings; (4) skeletal grainstone/pack-
stone limestone deposited in a restricted lagoonal
4. Available data environment or tidal flat; (5) porous ooid grainstone
indicating deposition of shallow-marine shoals on
Digital well logs were assembled from approxi- an open shelf; and (6) cemented ooid grainstone
mately 100 wells in and surrounding Big Bow and deposited in a similar environment as ooid grain-
stone, but with reduced porosity and permeability
due to late cementation. Core and facies descrip-
tions are available in Qi and Carr (2005a).

5.1. Quartz-rich carbonate grainstone

The quartz-rich carbonate grainstone lithofacies

(lithofacies 1) consists of mixed siliciclastic and
carbonate grainstone (Fig. 4A). The quartz-rich
carbonate grainstone facies ranges in thickness from
2 ft (half meter) to over 50 ft (15 m), and at an
individual location can compose up to 70% of the
stratigraphic thickness of the upper St. Louis
Limestone (e.g., ARCO ]1 Sullivan; Carr and
Fig. 3. Schematic of architecture of a neural network method.
Neural network model used in this study is a ‘‘back-propagation’’
Lundgren, 1994). An increased quartz fraction
network with an input layer, single hidden layer and output layer and low GR values are typical features for
(modified from Duda et al., 2001). this lithofacies. Crossover of neutron-density logs
 ARTICLE IN PRESS
L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964 951

indicate an increase in quartz and can be used to matrix suggest that this lithofacies was deposited
evaluate lateral continuity. Subtidal St. Louis in a low energy, open-marine shelf environment.
Limestone facies can also contain quartz grains Porosity averages less than 1%, with permeability of
and produce similar log signatures. Based on a less than 0.01 md. The skeletal wackestone facies is
variety of evidence including presence of translatent considered a non-reservoir flow unit.
stratification, root structures, high concentration of
quartz silt and very-fine-grained sand, large-scale 5.4. Skeletal grainstone/packstone or fenestral
cross-bed sets (1–9 ft; 0.3–3 m), and general absence limestone
of coarse-grained open-marine fauna, the quartz-
rich carbonate grainstone is interpreted as deposits The skeletal grainstone/packstone lithofacies
of an eolianite complex. Stylolitization is common. (lithofacies 4) is composed of very-fine skeletal
The quartz-rich carbonate grainstone facies is well grains with peloids and pellets, and rare very-fine
cemented and permeability is less than 1 md, and quartz and fine ooid grains (Fig. 4D). Observed
most commonly less than 0.01 md, and is considered thickness range of skeletal grainstone/packstone
a non-reservoir flow unit. facies is from 0.5 ft (15 cm) to 5 ft (1.5 m). Generally,
GR log values are less than skeletal wackestone and
5.2. Argillaceous limestone greater than ooid grainstone facies (below), and
porosity log values low. The skeletal grainstone/
The argillaceous limestone lithofacies (lithofacies packstone facies is difficult to distinguish on logs
2) is a thin, laterally extensive unit, characterized by from the skeletal wackestone lithofacies. The
thinly bedded, often fissile olive-green to gray fenestral limestone is interpreted as a tidal-flat
argillaceous limestone, which is primarily wackes- deposit, and the skeletal grainstone/packstone is
tone or packstone with abundant bryozoans, interpreted as a back barrier lagoonal deposit. Since
crinoids and minor brachiopods, and ostracodes these two lithofacies could not be distinguished with
(Fig. 4B). Thickness is from 1 ft (30 cm) to 10 ft digital well logs, they were combined for the
(3 m). On electric logs, the facies shows distinctively convenience of modeling. Porosity averages less
higher GR values relative to adjacent facies. The than 1%, with permeability of less than 0.01 md.
argillaceous limestone lithofacies is a laterally The skeletal grainstone/packstone facies is consid-
extensive feature, and is interpreted as representing ered a non-reservoir flow unit.
major deepening. The abundant bryozoans and
crinoids in wackestone and packstone, especially 5.5. Porous ooid grainstone
well-preserved whole fossils, indicate the argillac-
eous limestone lithofacies was deposited in a The porous ooid grainstone lithofacies (lithofa-
relatively deep-marine environment. The argillac- cies 5) is characterized by medium to very-coarse
eous limestone facies has very low porosity and size and moderately sorted ooids, and absence of
permeability, and is considered a non-reservoir flow mud, generally (Fig. 4E). Thickness range of porous
unit. ooid grainstone lithofacies is from 1.5 ft (45 cm) to
15 ft (4.5 m). The ooid grainstone facies has a
5.3. Skeletal wackestone distinctive log response with higher porosity and
lower resistivity relative to adjacent beds. GR values
The skeletal wackestone lithofacies (lithofacies 3) generally are low, indicating the absence of both
is mud supported and very poorly sorted (Fig. 4C). fine-grained organic material and clays. In the ooid
Primary sedimentary structures have been comple- grainstone facies, the main porosity is interparticle
tely obliterated by bioturbation. Thickness range porosity with minor moldic porosity. Primary
for the skeletal wackestone facies is from 0.5 ft interparticle pores were reduced by varying degrees
(15 cm) to 7 ft (2.1 m). The skeletal wackestone of meniscus, rim, syntaxial, and coarsely crystalline
facies has a slightly increased GR response. On log ‘‘spar’’ calcite cement. Lower porosity grainstone
response alone the skeletal wackestone cannot be typically contains more abundant skeletal debris.
distinguished easily from the skeletal grainstone/ Generally, greater cementation corresponds greater
packstone limestone or the quartz-rich carbonate presence of skeletal fragments, and the lower
grainstone facies. Poor sorting, large skeletal frag- porosity. The presence of ooids, superficial ooids,
ments, limited siliciclastic influx and a micritic and the absence of mud indicates that porous ooid
 ARTICLE IN PRESS
952 L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964
 ARTICLE IN PRESS
L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964 953

grainstone lithofacies accumulated in a high-energy, provides fluid barriers/baffles that could be modeled
open-marine mobile grain shoal environment. Total as objects.
porosity varies from 3% to as high as 20% and
permeability ranges from 0.1 to 2000 md. The ooid 6. Statistical data analysis
grainstone facies is considered the primary reservoir
quality flow unit. GR, resistivity logs (deep-ILD and medium-
ILM), neutron porosity (Nphi), density porosity
5.6. Cemented ooid grainstone (Dphi), and PE factor are commonly available
electric logs for wells in Big Bow and Sand Arroyo
The cemented ooid grainstone lithofacies (litho- Creek fields. Basic 1D, 2D, and multidimensional
facies 6) is composed of medium to very-coarse- statistical data analyses were performed with the
grained and moderately sorted ooids and skeletal listed curves in the cored wells, and the results were
fragments (Fig. 4F). Thickness range of cemented plotted as histogram, dotplot, boxplot, and cross
ooid grainstone lithofacies is from 0.7 ft (20 cm) to plot to evaluate the predictor variables for neural
5.5 ft (1.6 m). Pressure solution, compaction, and network modeling.
late cementation occlude interparticle voids. Total The histogram plot of GR shows that the
porosity is mainly less than 3% and permeability is argillaceous limestone facies (lithofacies 2) has a
usually less than 0.1 md. This lithofacies is con- distribution mainly ranging from 30 to 60 API,
sidered a non-reservoir quality flow unit. whereas the distributions of GR in other lithofacies
The porous ooid grainstone and cemented ooid generally show a range from 10 to 40 API (Fig. 5).
grainstone facies were both deposited as mobile In terms of resistivity, argillaceous limestone
grain shoals. Because of the reduced porosity and facies (lithofacies 2) has relatively low ILD and
permeability due to compaction and cementation, ILM values (Figs. 6 and 7). Based on boxplot and
the cemented ooid grainstone lithofacies has differ- dotplot analysis, the skeletal wackestone facies
ent rock and fluid property response and has been (lithofacies 3) and skeletal grainstone/packstone
separated as another lithofacies or flow unit from facies (lithofacies 4) have wider distribution of
porous ooid grainstone lithofacies. The distinction ILM and ILD values. Distributions of ILM and
of these two lithofacies as separate flow units ILD values of lithofacies 1 display a center-bell

Fig. 4. (A) Quartz-rich carbonate grainstone (lithofacies 1), St. Louis Limestone, Wade Allen 36-1. (i) Core photo of a quartz-rich
carbonate grainstone section with abundant inverse translatent stratification dipping at varying directions and angles up to 101,
5612–5616 ft core depth. (ii) Close-up core photo of stylolites and translatent stratification (showing horsetail structure), 5614.5 ft core
depth. (iii) Thin section photomicrograph of quartz-rich carbonate grainstone, composition of quartz debris around 50%, 5614.5 ft core
depth. (B) Argillaceous limestone (lithofacies 2), St. Louis Limestone, Wade Allen 36-1. (i) Core photo of thinly bedded argillaceous
limestone section, 5624–5629 ft core depth. (ii) Close-up core photo of muddy limestone with abundant bryozoans, crinoids, brachiopods,
and ostracods, 5626.3 ft core depth. (iii) Thin section photomicrograph of argillaceous limestone, micritic muds at lower part, well-
preserved brachiopod and large spar calcite at upper part, 5623.5 ft core depth. (C) Skeletal wackestone (lithofacies 3), St. Louis
Limestone, Wade Allen 36-1. (i) Core photo of skeletal wackestone section, 5629–5630 ft core depth. (ii) Close-up core photo of
wackestone with abundant skeletal fragments, quartz debris can be clearly observed, 5629.5 ft core depth. (iii) Thin section
photomicrograph of skeletal wackestone, relatively condensed un-abraided skeletal fragments such as crinoids, fenestrate bryozoans,
echinoderms, brachiopods, gastropods, and foraminifera float in micritic matrix, 5629.5 ft core depth. (D) Skeletal grainstone/packstone
limestone (lithofacies 4), St. Louis Limestone, Wade Allen 36-1. (i) Core photo of skeletal grainstone/packstone section, 5641–5645 ft core
depth. (ii) Close-up core photo of skeletal grainstone/packstone, algae lamination replaced with chert or calcite spar, 5645.7 ft core depth.
(iii) Thin section photomicrograph, skeletal grains cemented by calcite spar, nucleuses of some skeletal grains replaced by calcite spar,
5645.7 ft core depth. (E) Ooid grainstone (lithofacies 5), St. Louis Limestone, Wade Allen 36-1. (i) Core photo of brownish ooid grainstone
section, 5632–5637 ft core depth. (ii) Close-up core photo of ooid grainstone, subtle cross stratification defined by thin interval composed
of elongate skeletal particles with varying cementation, 5635.4 ft core depth. (iii) Thin section photomicrograph, radial concentric ooids
with skeletal and peloidal nuclei cemented by meniscus and rim cement, and usually unbroken, primary interparticle porosity preserved up
to 20%, 5635.4 ft core depth. Little pressure solution was observed. (F) Cemented ooid grainstone (lithofacies 6), St. Louis Limestone,
Wade Allen 36-1. (i) Core photo of cemented ooid grainstone section, 5647–5648.5, and 5664–5665 ft core depth. (ii) Close-up core photo
of cemented ooid grainstone, subtle cross bedding and stylolite structure shown, 5648 ft core depth. (iii) Thin section photomicrograph,
radical concentric ooid grains with skeletal and peloidal nuclei highly cemented by late calcite cement, with porosity less than 5%, 5648 ft
core depth. Less meniscus cement and more pressure solution observed.
 ARTICLE IN PRESS
954 L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964

GR Histogram of St. Louis Lithofacies

0 40 80 120
1 2 3

20

15

10

5
Percent

0
4 5 6

20

15

10

0
0 40 80 120 0 40 80 120
GR (API)
Fig. 5. Histogram of gamma-ray vs. lithofacies. Lithofacies 1: quartz-rich carbonate grainstone facies; lithofacies 2: argillaceous limestone
facies; lithofacies 3: skeletal wackestone facies; lithofacies 4: skeletal grainstone/packstone facies; lithofacies 5: ooid grainstone facies;
lithofacies 6: cemented ooid grainstone facies.

Boxplot of ILM vs Lithofacies Dotplot of ILD vs Lithofacies

800

600
ILM (ohm-m)

1
Lthofacies

400
2
3
4
200

5
0 6
1 2 3 4 5 6 0 50 100 150 200 250 300 350
Lithofacies ILD(ohm-m)

Fig. 6. Boxplot shows shape of distribution, its central value, and Fig. 7. Dotplot of deep induction log (ILD) vs. lithofacies, data
variability of ILM vs. lithofacies. Picture produced consists of plotted as dots (one dot represents one observation) above a
most extreme values (maximum and minimum values), lower and horizontal axis that covers entire range of observations.
upper quartiles, and median of ILM in each lithofacies. Lithofacies 1 displays an ILD distribution more like a normal
Lithofacies 2 displays relatively low ILM value and narrow center-bell shape while lithofacies 2 and 5 show a left-centered
distribution. Lithofacies 3 and 4 have wider range of ILM skewed distribution. Lithofacies 5 has a different distribution
distribution than other lithofacies. shape and mean value from lithofacies 6.

shape normal distribution, whereas lithofacies 2 and stone facies (lithofacies 1) has an intermediate range
5 show a left-centered skewed distribution. The ooid in ILM and ILD distribution.
grainstone facies (lithofacies 5) has a different The distribution of the average of density and
distribution shape and mean value of resistivity neutron porosity by category of lithofacies shows
values compared with cemented ooid grainstone that the ooid grainstone facies (lithofacies 5) has a
facies (lithofacies 6). Quartz-rich carbonate grain- mean of 10% and the largest range of porosity
 ARTICLE IN PRESS
L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964 955

compared with other lithofacies (Fig. 8). Lithofacies toward the left side in the distribution may be the
1 and 4 have the smallest distribution of average effect of mixed quartz grains or clay.
porosity. With the effect of cementation, porosity in Many of the lithofacies in the St. Louis Limestone
cemented ooid grainstone facies (lithofacies 6) has have relatively distinctive well log responses, but
been reduced dramatically. there are significant overlaps. The overlap in log
The distribution of the PE value in terms of response can be displayed with the cross plot of PE
lithofacies illustrates that lithofacies 1 has a lower curve vs. average porosity curve by lithofacies (Fig.
mean value compared with other lithofacies (Fig. 9). 10). The ooid grainstone facies (lithofacies 5; orange
Because lithofacies 1 contains a large proportion of dots) compared to the quartz-rich carbonate grain-
quartz grains, it has relatively lower PE values. The stone facies (lithofacies 1; black dots) shows a
common reservoir mineral reference values for PE distinct trend with different slopes, and little over-
log are: quartz 1.81; dolomite 3.14; and calcite 5.08 lap. Other lithofacies display patterns that are
barns/electron. In lithofacies 2 and 4, the small tail intermediate to these two extremes, and display
extensive overlap among the lithofacies. 3D statis-
tical methods provide useful information in terms of
Boxplot of (DPHI+NPHI)/2 vs Lithofacies
classifying the lithofacies, but have limitations in
16 dealing with multiple dimensional data, and have
high rates of incorrect classifications.
(DPHI+NPHI)/2 (%)

12 As the next step, multiple linear discriminant

analysis was performed to classify the correspond-
8 ing well log values into predefined St. Louis
lithofacies. Inputs were normalized continuous
4
variables (well logs) and categorical variables
(predefined St. Louis lithofacies). The classification
analysis results using a discriminant prediction
0
equation displayed the squared distance between
1 2 3 4 5 6
Lithofacies
groups formed by the dependent variable (Table 1).
Group 3 (skeletal wackestone facies; lithofacies 3)
Fig. 8. Boxplot shows shape of distribution, its central value, and has the smallest number of samples and the lowest
variability of average log porosity vs. lithofacies. Lithofacies 5 correct proportion from the classification analysis
has largest range of porosity with mean around 10%. Lithofacies
1 and 4 have smallest distribution of average porosity. With
(11.1%). On the other hand, group 2 (argillaceous
effect of cementation, porosity in lithofacies 6 has been reduced limestone facies; lithofacies 2) showed the highest
dramatically. correct percentage of 73.5%. Also, groups 5 and 1

Dotplot of PE vs Lithofacies
PE vs. Phi (D+N)/2 by Lithofacies
Lithofacies
1
15 2
3
Lithofacies

4
1 5
Phi (%)

10 6
2
3
4
5
5
6
2.4 2.8 3.2 3.6 4.0 4.4 4.8 5.2 0
PE (B/E) 2 3 4 5
PE (B/E)
Fig. 9. Dotplot of photoelectric (PE) log vs. lithofacies, data
plotted as dots (one dot represents one observation) above a Fig. 10. Cross plot of photoelectric (PE) log vs. average porosity
horizontal axis that covers entire range of observations. by lithofacies. Lithofacies 5 (orange color) shows a separate trend
Distribution of PE value illustrates that lithofacies 1 has a lower from lithofacies 1 (black color) with different slopes, and little
mean value compared with other lithofacies, indicating high overlap. Other lithofacies show patterns that display extensive
content of quartz grains. overlap.
 ARTICLE IN PRESS
956 L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964

Table 1
Discriminant analysis result

Linear method for response: lithofacies

Predictors: GR, ILM, ILD, OHM, DPHI, NPHI, PE

Lithofacies 1 2 3 4 5 6
Count 1113 215 81 221 207 186
Summary of classification
True lithofacies

Put lithofacies 1 2 3 4 5 6

1 724 17 2 46 10 39
2 55 158 13 6 1 4
3 132 6 9 34 11 27
4 110 0 26 89 24 17
5 15 10 2 5 124 21
6 77 24 29 41 37 78
Total N 1113 215 81 221 207 186
N correct 724 158 9 89 124 78
Proportion 0.650 0.735 0.111 0.403 0.599 0.419
N ¼ 2023 N correct ¼ 1182, Proportion correct ¼ 0.584
Squared distance between lithofacies

1 2 3 4 5 6

1 0.0000 9.2200 1.4814 1.9910 10.1624 2.3950

2 9.2200 0.0000 7.8829 10.3711 13.9126 7.8766
3 1.4814 7.8829 0.0000 0.9893 8.1698 0.7493
4 1.9910 10.3711 0.9893 0.0000 10.4562 2.3650
5 10.1624 13.9126 8.1698 10.4562 0.0000 6.5030
6 2.3950 7.8766 0.7493 2.3650 6.5030 0.0000

Classification summary illustrates correct proportion of each category variables (predefined St. Louis lithofacies) with input predictors (six
well logs). Squared distance shows difference between predefined St. Louis lithofacies.

Score Plot of Selected 6 Log Curves have reasonable results (60%). Group 1 had a more
4
distinctive difference from group 2 and 5 compared
3
with other three groups. Based on squared distance
2
Second Component

1
analysis, it remains difficult to separate group 3
0 from group 1, 4 and 6, while groups 2 and 5 can be
-1 easily separated from the other four groups. Multi-
Lithofacies
-2 1 ple linear discriminant analysis improves on simple
2
-3 3 2D statistical methods, but incorrect classifications
-4 4 and overlap of facies remain relatively significant.
5
-5 6 Principal component analysis (PCA) was per-
-5.0 -2.5 0.0 2.5 5.0 7.5 formed to reduce the dimensionality of the training
First Component data set. Based on the analysis results, the six log
curves (GR, medium resistivity, deep resistivity,
Fig. 11. Score plot of principal component analysis. On first density porosity, neutron porosity, and PE) were
component, lithofacies 5 and 2 show different directions and may
be separated from lithofacies 1 and 4. On second component,
plotted on the score plot and loading plot (Figs. 11
lithofacies 4 can be separated from lithofacies 5. It is difficult to and 12). On the first component, the training data
separate lithofacies 6 either on first or second component. can be separated by porosity logs (Dphi, Nphi) and
 ARTICLE IN PRESS
L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964 957

Loading Plot of Selected 6 Log Curves and graphics. It is similar to the S language and
0.75 environment developed by John Chambers and
PE
colleagues (Becker et al., 1988; Chambers and
Second Component

0.50 DPHI(%)
Hastie, 1992), but is available in public domain.
ILM
0.25 The output-layer nodes of the neural network model
NPHI(%) ILD were sets of predicted lithofacies probability, and
0.00 the final predicted codes were assigned to the
-0.25 lithofacies with maximum probability at that
specified depth. The neural network models were
-0.50 GR cross-validated to select the optimized neural net-
work ‘‘size’’ (number of hidden-layer nodes) and the
-0.50 -0.25 0.00 0.25 0.50 damping parameter. The neural network ‘‘size’’
First Component controls the richness of the model, and the damping
Fig. 12. Loading plot of principal component analysis. On first parameter governs the magnitude of the neural
component, training data can be separated by porosity logs network weights in order to prevent overtraining. R
(Dphi, Nphi) and medium and deep induction logs (ILM, ILD). language scripts were programmed to automatically
On second component, gamma-ray and photoelectric logs have perform the cross-validation by invoking the nnet
opposite responses.
function in the R data-analysis system. The 10 cored
well data were randomly split into 10 sets, which
included training (2/3) and prediction (1/3) subsets.
resistivity logs (ILM, ILD). On the second compo- The training subsets were trained to build the neural
nent, GR and PE logs have opposite responses. network model for each pair of network size (10, 20,
Even after the PCA, lithofacies show a high degree 30, 35, 40, 50) and damping parameter (10
4, 10
3,
of overlap. Density porosity and neutron porosity 10
2, 10
1, 100), and subsequently, lithofacies were
logs display almost the same direction on the predicted using the prediction subsets with the
loading plot, and measure different components. corresponding network model.
Thus, both logs were used to build the neural The objective function is a measure of the overall
network model. divergence between true lithofacies indicators and
the predicted lithofacies stated in terms of entropy.
7. Model building and cross-validation Another measure of misfit is the mean absolute
difference between the actual and predicted facies
The basic statistical analysis between digital well number. The objective function values and mean
logs and lithofacies described from core of 10 wells absolute difference were computed over the 10
in Big Bow and Sand Arroyo Creek fields illustrates prediction subsets with the selected six well logs for
the limitation of traditional statistical methods. Six neural network models of different network size and
well logs (GR, ILM, ILD, Nphi, Dphi, PE curves) damping parameters (Figs. 13 and 14). Based on the
were selected as input variables with digitized objective function and mean absolute difference, 35
lithofacies categories to develop a standard single hidden-layer nodes and a damping parameter of
hidden-layer neural network model. The single-layer 0.01 were selected to build the optimal neural
‘‘back-propagation’’ neural network was used to network model with the PE curve. A model with
predict lithofacies from digital well logs. The 30 hidden-layer nodes and a damping parameter of
prepared digital log curves (input-layer nodes) and 0.1 was constructed without the PE curve.
digitized lithofacies categories were input to train The neural network was trained and then used to
and build neural network models. Kipling, a neural predict lithofacies on the whole data set of the 2023
network tool implemented as a spreadsheet add-in half-foot intervals from the 10 cored wells with the
using Visual Basic, was used to build a series of selected network size of 35 and a damping
models (Bohling and Doveton, 2000). The neural parameter of 0.01. The pivot table shows predicted
network function (nnet function; Venables and lithofacies results compared to actual lithofacies
Ripley, 1999) in the ‘‘R’’ data-analysis system was (Table 2). For example, 209 intervals were predicted
called to perform the training process and optimize as ooid grainstone facies (lithofacies 5) compared to
the network weights. ‘‘R’’ data-analysis system is a 207 actual. Absolute accuracy indicates the propor-
language and environment for statistical computing tion of the correct prediction number. Lithofacies 3
 ARTICLE IN PRESS
958 L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964

Fig. 13. Objective function values for neural network models of different network size and damping parameters to cross-validate neural
network models.

possesses the smallest absolute accuracy of 70.37%. largest probability were generated for each depth
In the reservoir quality facies (lithofacies 5), (Fig. 15).
absolute accuracy is very high (90.82%). Incorpor- Based on core description information and
ating adjoining lithofacies, within-one lithofacies geological knowledge, minor modifications were
(including lithofacies 4 and 6) improves accuracy performed for the discrete lithofacies curves with
slightly (93.72%). It appears that the neural net- reference to the lithofacies probability curves as
work models effectively predict St. Louis Limestone quality control. For instance, sparse lithofacies
lithofacies. 3 samples within a thick lithofacies 1 interval would
be replaced by lithofacies 1 even though they have
the largest probability values, since the existence of
8. Lithofacies prediction a thin interval of lithofacies 3 in a thick lithofacies 1
interval was not observed in any of the cores and is
After selecting the neural network parameters considered a prediction error. Then, the LAS-
through the cross-validation and model testing, the formatted prediction lithofacies files were imported
final step is to run neural network models on wells into geological applications for demonstration of
with LAS files and output the predicted lithofacies. lithofacies prediction.
LAS log files were batch processed in Kipling, and
LAS-formatted prediction files were generated after 9. Cross-sections with interpolated lithofacies
prediction. Seven curves are contained in the output
predicted lithofacies file. For the 10 cored wells six In 10 cored wells, lithofacies curves were com-
probability curves for lithofacies and one discrete posed of description from cored intervals and
lithofacies curve representing the lithofacies with predicted lithofacies values from neural network
 ARTICLE IN PRESS
L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964 959

Fig. 14. Mean absolute difference in lithofacies number for neural network models of different network size and damping parameters to
cross-validate neural network models.

Table 2
Pivot table shows prediction effectiveness of neural network model (neural network size of 35; damping parameter of 0.01; iteration
number of 100)

Predicted lithofacies Absolute Counts within Accuracy within

accuracy (%) 1 1 (%)
Actual 1 2 3 4 5 6 Total
lithofacies

1 1074 14 4 11 4 6 1113 96.50 1089 97.84

2 15 199 1 215 92.56 214 99.53
3 11 4 57 1 5 3 81 70.37 62 76.54
4 26 1 1 172 5 16 221 77.83 178 80.54
5 9 2 2 1 188 5 207 90.82 194 93.72
6 17 1 9 7 152 186 81.72 168 90.32

Total 1152 220 65 194 209 183 2023

Percent 103.5 102.3 80.3 87.8 100.9 98.4 Absolute Within 1 Accuracy (F5, 6)
accuracy (%) lithofacies (%) (%)

Difference 39 5 16 113 12 24 209 91.05 94.17 99.75

for uncored intervals. Lithofacies curves were models. Adding the cored wells, a total of 100 wells
predicted in approximately 90 wells with available with lithofacies curves were used to construct 17
well log suites using the optimal neural network stratigraphic cross-sections of St. Louis Limestone
 ARTICLE IN PRESS
960 L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964

Fig. 15. Predicted lithofacies probability curve, predicted discrete lithofacies curve, and actual lithofacies curve for 10 cored wells, St.
Louis Limestone: (A) predicted facies probability plot; (B) predicted lithofacies plot; and (C) actual lithofacies plot.

Fig. 16. Well locations and cross-sections of Big Bow and Sand Arroyo Creek fields, T29-T30S and R38-R41W, St. Louis Limestone,
Southwest Kansas.

in Big Bow and Sand Arroyo Creek fields (Qi and cross-sections to better display large-scale facies
Carr, 2005b). distribution and depositional patterns across the
Using predicted lithofacies curves along with the study area. Two interpolated color filled cross-
traditional log markers and patterns, interpolated sections (D–D0 and E–E0 , Fig. 16) show the facies
color fill functions were applied to construct 2D distribution and depositional patterns across a local
 ARTICLE IN PRESS
L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964 961

high in Big Bow Field. Both stratigraphic cross- patterns of the six lithofacies are illustrated by the
sections illustrate the oolitic deposits (including color-filled cross-sections and illustrate facies dis-
lithofacies 5 and 6) associated with local lows and tribution at a large scale. An improved fine-scale
highs within the St. Louis Limestone (Figs. 17 and lithofacies distribution model for the St. Louis
18). Quartz-rich carbonate grainstone (lithofacies 1) Limestone can be built with more advanced
dominate in zones A and B of the St. Louis in these stochastic simulation tools.
two cross-sections (Figs. 17 and 18). All 2D cross-
sections generated in this study are available 10. Discussion
through the Internet (Qi and Carr, 2005b).
Because of the simple linear interpolation in areas Compared to conventional linear methods, neural
away from the well bores, the interpolation does not network models provide improved prediction results
generate geologically realistic lithofacies interpreta- using training patterns gained from learning
tions. However, the lateral and vertical depositional and exploring the hidden information within the

Fig. 17. Interpolated color fill cross-section of D–D0 in Fig. 16, flatten on top of St. Louis Limestone. Oid grainstone reservoir lithofacies
(5 and 6) thicken and display up-dip stratigraphic pinchouts associated with local highs. Big Bow Field, Southwest Kansas.

Fig. 18. Interpolated color fill cross-section of E–E0 in Fig. 16, flatten on top of St. Louis Limestone. Big Bow Field, Southwest Kansas.
 ARTICLE IN PRESS
962 L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964

multi-dimensional data set. As a non-parametric is emphasized when the neural network is used to
method, neural network analysis is not based on predict lithofacies from well logs.
restrictive assumptions about the probability dis-
tributions of input variables. Although it is un- 11. Conclusions
necessary to examine the correlation and
distribution of input variables when using a neural Neural network analysis coupled with detailed
network, basic statistical study of the input vari- lithofacies descriptions is applied to construct
ables is an essential step. Suitable selection of improved finer-scale facies models, and provide a
input variables will increase the predictive accuracy better understanding of the relatively thin oolitic
of the network. If training data are limited, the grainstone shoal reservoirs in the St. Louis Lime-
neural network model may be over-fitted since the stone, of Southwest Kansas. A total of 750 ft of St.
model will memorize only a few patterns. Ten cored Louis Limestone intervals in 10 cored wells were
wells (over 2000 data points) were used to build described and six lithofacies were classified and
reliable network models for the St. Louis Lime- treated as training data with corresponding suites of
stone. Cross-validations and suitable validation well logs. Correlations and relationships of well logs
parameters are keys to selecting the neural net- were explored with basic 1D statistics, discriminant
work parameters (network size and damping analysis and PCA. Six well logs were selected as
parameter). However, selection of appropriate net- input training variables along with the discrete
work parameters is subjective, since there is no lithofacies curves. Neural network parameters (net-
definite theory demonstrating the selection process work size and damping parameter) were determined
(Wong et al., 1998). for neural network models through cross-validation
Normalization of the training variables is also a with PE or without PE curves, and then used to
very important step. In this study, well logs of all 10 predict lithofacies on the whole data set of the 2023
cored wells were normalized to the same conditions half-foot intervals from the 10 cored wells with the
before the training process. Well log data from all selected network size of 35 and a damping
uncored wells were normalized to the same condi- parameter of 0.01. Predicted lithofacies results
tions as the training data. compared to actual lithofacies displays absolute
Consistent application of geological background accuracies of 70.37–90.82%. Incorporating adjoin-
and sedimentary interpretation are critical to the ing lithofacies, within-one lithofacies improves
successful neural network lithofacies prediction, in accuracy slightly (93.72%).
which descriptive information is intensively used in LAS files were patch processed by PC-based
an empirical sense. In this case, consistent lithofa- spreadsheet add-in program (Kipling.xla) and R
cies classification based on detailed core description language system (Bohling and Doveton, 2000), and
of distinctive difference is fundamental to neural neural network models were constructed and
network training and predictions. For example, lithofacies curves were predicted for 90 uncored
lithofacies 1 and 5 have distinctive differences in wells in Big Bow and Sand Arroyo Creek fields.
terms of facies description and corresponding log Interpolated cross-sections were constructed based
response. The percentages of correction for these on the neural network prediction results to better
two lithofacies are much better than other four illustrate depositional patterns of lithofacies in the
lithofacies. Lithofacies 3 and 4 were deposited in St. Louis Limestone across the study area.
adjacent environments, with similar lithologies and The following specific conclusions are drawn
log responses. The prediction results display miss- from this study:
predicted numbers for lithofacies 3 and 4. Deposi-
tional patterns observed in cores are also very 1. Described core lithofacies in conjunction with
important, and should constrain the predicted neural network models display a significant
lithofacies curves. The unreasonable occurrence of accuracy of lithofacies prediction in relatively
sparse lithofacies within intervals dominated by thin carbonate reservoirs in Southwest Kansas,
another lithofacies, which is not verified from the and provides a methodology for predicting
observed depositional patterns from core, are used lithofacies in complex carbonate reservoirs.
to revise the predicted lithofacies curves by repla- 2. Neural network models provide a robust method
cing the miss-predicted lithofacies with surrounding for predicting electrofacies from well logs in
intervals. The importance of geological constraints complex carbonate reservoirs, where it is difficult
 ARTICLE IN PRESS
L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964 963

to establish a functional relationship between Abegg, F.E., Handford, C.R., 2001. Deflation origin of
lithofacies and well logs with conventional Mississippian carbonate eolianites of Southwestern Kansas.
statistical methods. In: Abegg, F.E., Harris, P.M., Loope, D.B. (Eds.), Modern
and Ancient Carbonate Eolianites: Sedimentology, Sequence
3. Data examination, statistical analyses, and geo- Stratigraphy, and Diagenesis. Society of Sedimentary Geol-
logical constraints are key factors to neural ogy (SEPM) Special Publication No. 71, pp. 183–203.
network modeling and improved prediction in Abegg, F.E., Loope, D.B., Harris, P.M., 2001. Deposition and
carbonate reservoirs. diagenesis of carbonate eolianites. In: Abegg, F.E., Harris,
4. The neural network predicted lithofacies curve P.M., Loope, D.B. (Eds.), Modern and Ancient Carbonate
Eolianites: Sedimentology, Sequence Stratigraphy, and Diag-
derived from well logs can increase our under- enesis. Society of Sedimentary Geology (SEPM) Special
standing of the facies distribution and deposi- Publication No. 71, pp. 17–32.
tional patterns of complex carbonate reservoirs. Ali, M., Chawathe, A., 2000. Using artificial intelligence to
5. Interpolated cross-sections constructed using predict permeability from petrographic data. Computers &
neural network predicted lithofacies curves illus- Geosciences 26 (8), 915–925.
Becker, R.A., Chambers, J.M., Wilks, A.R., 1988. The New S
trate that the oolitic deposits (including lithofa- Language: a Programming Environment for Data Analysis
cies 5 and 6) associated with local highs and and Graphics. Wadsworth & Brooks/Cole, Pacific Grove, CA
quartz-rich carbonate grainstone (lithofacies 1) (702pp).
dominate in the zones A and B of the St. Louis Bohling, G.C., Doveton, J.H., 2000. Kipling.xla: an Excel add-in
Limestone. for nonparameter regression and classification. Report,
Kansas Geological Survey (69pp).
6. Using neural network modeling, the link between Bohling, G.C., Dubois, M.K., 2003. An integrated application of
prescribed lithofacies and digital logs provides a neural network and Markov chain techniques to prediction of
strong basis for fine-scale reservoir modeling and lithofacies from well logs. Kansas Geological Survey Open-
stochastic simulation in complex carbonate File Report, No. 2003-50 (6pp).
reservoirs. Carr, T.R., Lundgren, C.E., 1994. Use of gamma ray spectral log
to recognize exposure surfaces at the reservoir scale: Big Bow
Field (St. Louis Limestone), Kansas. In: Dolson, J.C.,
Acknowledgements Hendricks, M.L., Wescott, W.A. (Eds.), Unconformity-
Related Hydrocarbons in Sedimentary Sequences. Rocky
We thank the Kansas Geological Survey for Mountain Association of Geologists, Denver, CO, pp. 79–88.
financial support as part of Lianshuang Qi’s Chambers, J.M., Hastie, T.J., 1992. Statistical Models in S.
Wadsworth & Brooks/Cole, Pacific Grove, CA (608pp).
dissertation research. Special thanks to Geoff
Dubois, M.K., Byrnes, A.P., Bohling, G.C., Seals, S.C., Doveton,
Bohling for his assistance and comments on the J.H., 2003. Statistically based lithofacies predictions for 3-D
manuscript. Thanks to Marty Dubois for his reservoir modeling: an example from the Panoma (Council
discussions and comments. Reviewers Rick Abegg Grove) Field, Hugoton Embayment, Southwest Kansas.
and Jeffery Yarus provided significant input that Kansas Geological Survey Open-File Report, No. 2003-30.
/http://www.kgs.ku.edu/PRS/publication/2003/ofr2003-30/
improved the manuscript. BP-American provided
toc0.htmlS.
the donated core material. We also would like to Duda, R.O., Hart, P.E., Stork, D.G., 2001. Pattern Classifica-
thank Geoplus Corporation for providing access to tion, second ed. Wiley, New York (654pp).
Petra. Dunham, R.J., 1962. Classification of carbonate rocks according
to depositional texture. In: Ham, W.E. (Ed.), Classification of
Carbonate Rocks, vol. 1. American Association of Petroleum
Geologists Memoir, pp. 108–121.
Goebel, E.D., 1968. Mississippian rocks of western Kansas.
References American Association of Petroleum Geologists Bulletin 52,
1732–1778.
Abegg, F.E., 1991. Sedimentology and lithostratigraphy of the Handford, C.R., 1988. Review of carbonate sand-belt deposition
Upper Mississippian Ste. Genevieve and St. Louis Lime- of ooid grainstones and application to Mississippian reser-
stones, southwestern Kansas. In: Integrated Studies of voir, Damme Field, southwestern Kansas. American Associa-
Petroleum Reservoirs in the Mid-Continent, Midcontinent tion of Petroleum Geologists Bulletin 72, 1184–1199.
American Association of Petroleum Geologists Section Meet- Handford, C.R., Francka, B.J., 1991. Mississippian carbonate-
ing, pp. 1–33. siliciclastic eolianites in southwestern Kansas. In: Lamando,
Abegg, F.E., 1994. Lithostratigraphy of the Hugoton and Stevens A.J., Harris, P.M. (Eds.), Mixed Carbonate Siliciclastic
members of the St. Louis Limestone and the Ste. Genevieve Sequence. Society of Sedimentary Geology (SEPM) Core
Limestone (Upper Mississippian), southwestern Kansas. In: Workshop, No. 15, pp. 205–244.
Baars, D.L. (Compiler), Revision of Stratigraphic Nomen- Lane, H.R., Dekeyser, T.L., 1980. Paleogeography of the late
clature in Kansas. Kansas Geological Survey Bulletin, No. Early Mississippian (Tournaisian) in the central and south-
230, pp. 39–66. western United States. In: Fouch, T.D., Magathan, E.R.
 ARTICLE IN PRESS
964 L. Qi, T.R. Carr / Computers & Geosciences 32 (2006) 947–964

(Eds.), Paleozoic Paleogeography of the West-Central United Qi, L.S., Carr, T.R., 2005b. Lithofacies cross-sections of the St.
States, Rocky Mountain Paleogeography Symposium 1. Louis Limestone, Big Bow and Sand Arroyo Creek fields,
Rocky Mountain Section-SEPM, Denver, CO, pp. 149–162. Southwest Kansas. Kansas Geological Survey Open-File
Lee, S.H., Kharghoria, A., Datta-Gupta, A., 2000. Electrofacies Report, No. 2005-14. /http://www.kgs.ku.edu/PRS/publication/
characterization and permeability predictions in complex 2005/OFR05_14/index.htmlS.
reservoirs. Society of Petroleum Engineers, Reservoir Evalua-
Oja, E., 1989. Neural networks, principal components, and
tion and Engineering, 237–248.
subspaces. International Journal of Neural Systems 1 (1),
Ligtenberg, L.H., Wansink, A.G., 2001. Neural network predic-
tion of permeability in the El Graia Formation, Ashtart 61–68.
Oilfield, offshore Tunisia. Journal of Petroleum Geology 24 Saggaf, M.M., Nebrija, E.L., 2003a. A fuzzy logic approach for
(4), 389–404. the estimation of facies from wire-line logs. American
Mathisen, T., Lee, S.H., Datta-Gupta, A., 2003. Improved Association of Petroleum Geologists Bulletin 87 (7),
permeability estimates in carbonate reservoirs using electro- 1233–1240.
facies characterization: a case study of the North Robertson Saggaf, M.M., Nebrija, E.L., 2003b. Estimation of missing logs
Unit, West Texas. Society of Petroleum Engineers, Reservoir by regularized neural networks. American Association of
Evaluation and Engineering, 176–184. Petroleum Geologists Bulletin 87 (8), 1377–1389.
Morrison, E., Maroney, M., Noah, K., O’Dell, R., 2002. Trappe, H., Hellmich, C., 2000. Using neural networks to predict
Depositional model and exploration potential in the St. Louis porosity thickness from 3D seismic. First Break 18 (9),
carbonates beneath the Hugoton Field of Kansas. Kansas
377–384.
Geological Survey Open-file Report, No. 2002-8. /http://
Venables, W.N., Ripley, B.D., 1999. Modern Applied Statistics
www.kgs.ku.edu/PRS/Poster/2002/2002-8/index.htmlS.
Qi, L.S., Carr, T.R., 2005a. Core description of the St. Louis with S-plus, third ed. Springer, New York (501pp).
Limestone, Big Bow and Sand Arroyo Creek fields, Southwest Wong, P.M., Henderson, D.J., Brooks, L.J., 1998. Permeability
Kansas. Kansas Geological Survey Open-File Report, determination using neural networks in the Ravva Field,
No. 2005-16. /http://www.kgs.ku.edu/PRS/publication/ Offshore India. Society of Petroleum Engineers, Reservoir
2005/OFR05_16/index.htmlS. Evaluation and Engineering, 99–104.