Neural Networks1
R5E R6E
Type Log
25 30
T T
-15
-15 36 35 36
,3 5
,300 35
-15
2 33 2
0
,37
N N
8
study area
Figure 2—Structure map on top of Smackover Formation–Big Escambia Creek field, Escambia County, southern
Alabama. Note the type well near the western limit of the field.
The percentage of dolomite varies from 100% to patterns, retrieve data associatively, filter noise
locally less than 10%. In some places where the from experimental data, complete missing informa-
process of dolomitization did not go to completion, tion, and estimate sampled functions (when we do
it was strongly fabric selective. Nondolomitized not know the mathematical form of the functions).
strata in Big Escambia Creek field are impermeable. In short, ANNs can be programmed to mimic
human characteristics, such as learning, generaliza-
tion, and interpretation, where traditional comput-
NEURAL-NETWORK APPROACH ing algorithms and statistical methods have been
inadequate (Hertz et al., 1991).
A brief introduction to neural networks appeared
in Rogers et al. (1992), in which an ANN was used
to determine rock types from well logs. In the fol- Backpropagation Neural Networks
lowing sections, we give additional information rel-
evant to this paper. One of the more popular architectures of ANNs
is the backpropagation neural network (BPNN).
BPNNs have proven to be successful in a variety of
Neural Networks applications and are ver y f lexible and easy to
implement. We used the BPNN because of these
The human brain is composed of approximate- attributes and because of its suitability to predic-
ly 10 billion interconnected cells, called neurons, tion problems. BPNNs are suitable for prediction
whose interaction is responsible for the character- problems where the input-output functions need
istics attributed to intelligence. The brain is a bio- to be learned from experience. Thus, they are
logical neural network. ANNs are created within a model-free estimators.
serial or parallel computer in an attempt to simu- BPNNs are good examples of supervised learn-
late the interactions among neurons in biological ing paradigms. In supervised learning, the network
neural networks. Like brains, ANNs consist of learns from a training set consisting of input-output
numerous “neurons” or processing units and pairs of vectors. An input vector is applied to the
interconnecting synapses that we can program for network, and the output vector is used as a “teach-
computation. We can program ANNs to recognize er” to show the correct (or desired) output.
Rogers et al. 1789
Figure 4—Architecture of
neural network used in this
paper.
given depth with its corresponding permeability lookup table could be created from a large number
value. If a one-to-one correspondence existed of examples. Because this is not the case, a more
between porosity and permeability, a simple innovative approach is required. Instead of using
just one porosity value, a set of adjacent porosity
values are used to predict a single permeability
value. For this work, 25 porosity values, with 12
porosity values above and below the depth to be
predicted, as well as the porosity value at the
depth were used to predict the permeability
(Figure 5). Spatial information (geographic coordi-
nates and depth) was used as part of the input to
enhance the network’s ability to make predictions.
BPNNs can learn the significance of each input to
the predicted value so that inputs that do not influ-
ence the predicted value will not interfere with
the network’s performance. This is accomplished
by adjusting the weights of connections between
pairs of nodes during the training process, as
explained in the previous section.
The training sets described in this paper were
composed of training set porosity-permeability pairs
in the form illustrated in Figure 5. Figure 4 shows the
inputs and outputs as the BPNN sees them.
CASE STUDIES
Figure 5—Schematic diagram showing how spatial
information, porosity, and permeability data were used Six cored wells from Big Escambia Creek field in
in training the network. southern Alabama provided observed porosity and
Rogers et al. 1791
Figure 6—Porosity-
permeability data for all
wells used in this study.
permeability values in 1-ft intervals (Figure 6). These as if it were a test well. At some point, the mean
wells were drilled in a part of the field that is litho- square error, which had been falling, began to rise.
logically homogeneous laterally (cf. Bradford, 1984; The network weights that provided the best results
Figure 7). Using the cored data in various combina- (lowest mean square error) on the calibration well
tions, three scenarios were run to test the BPNN’s were used for the actual prediction of the test wells.
predictive capability. These scenarios were based on Minor reversals in the trend of the mean square error
geological considerations. In each scenario, a subset were ignored to avoid choosing a local minimum as
of the six wells was used for a training set (either the solution. The calibration well was never includ-
two, three, or four wells), one well was used to cali- ed in the training sets or the testing sets so that a
brate the network, and the remaining wells were true test of the network’s performance could be
predicted. The training set used all of the observed measured. Note that the test wells are never used in
porosity and permeability values available, the X and the training set or as a calibration well because this
Y coordinates of the well’s position in the field, and would influence the results of the network training
the depth (Z) of the observed permeability value. and might cause the network to memorize rather
The BPNN was trained for several thousand itera- than generalize. Both permeability (from cores) and
tions of the training set as follows. During the itera- porosity (either from cores or logs) are necessary in
tive training process, the current network weights the training process, but only porosity data are
periodically were checked using the calibration well needed to make predictions.
1792 Predicting Permeability
Figure 7—Comparison of depth vs. porosity and permeability in wells 1704 and 1705, indicating a strong similarity
between the Smackover sections in these two wells. Dashed line indicates correlation of strata between these two
wells.
In the first scenario, two wells (permits 1877 the network in this scenario was degraded by one
and 1928) were used for the training, and perme- or more of at least two possible factors. The area
ability values were predicted in well 1930 using covered by the wells used in this scenario was
the porosity values at the same depths. All three much greater than that for scenario one. Lateral
wells lie approximately along strike from one reservoir heterogeneity over these greater dis-
another, and this arrangement minimizes lateral tances may have adversely influenced the network.
lithologic variation. Well 1802 was the calibration In addition, a boundary between two reservoir
well. Figure 8 is the location map of the wells and compartments may lie within the larger prediction
Figure 9 gives the results of ANN prediction. The area of scenario two. As mentioned in the section
predicted values from a linear regression analysis on geologic setting, the Smackover Formation in
are also shown. ANN yielded a much more accu- Big Escambia Creek field is characterized by rela-
rate prediction than the regression analysis (Figure 9; tively high values of megascopic reservoir hetero-
Table 1). geneity, which could be expressed in part as geo-
In the second scenario, three wells were used graphically distinct reservoir compartments.
for training (wells 1802, 1877, and 1930) and the The last scenario used four training wells (1802,
calibration well was 1928. In this scenario, we pre- 1877, 1928, and 1930), also used in the previous
dicted permeability in two different wells (1704 two cases, and predicted permeability in one well
and 1705), which lie along strike with and between (1704; permeability also was predicted in this well
the training wells. We expected this scenario to in scenario two). Well 1705 was used for calibra-
perform at least as well as scenario one because the tion. Figure 8 shows the wells used and Figure 12
distribution of training wells was designed to cap- gives the results, indicating that the network per-
ture information about petrophysical variation in formed no better than the regression analysis. The
the region surrounding the predicted wells. The same factors affecting the results of scenario two
results are shown in Figures 10 and 11, respective- influenced this scenario because they differed only
ly. The BPNN predicted permeability slightly more by the addition of a fourth training well, which had
accurately than the regression model, but the dif- small effect on the distribution of data used to train
ference is not significant (Table 1). Performance of the network.
Rogers et al. 1793
Porosity and permeability data for Big Escambia “window.” However, in practical terms this short-
Creek field are available on diskette from AAPG. coming may not be severe for two reasons. First, a
1-ft-thick high-permeability streak may not consti-
tute a reservoir interval. If the high-permeability
DISCUSSION streak is several feet thick, then the method does
a much better job of prediction (e.g., thicker per-
All of the case studies have a few features in meable zones in Figures 9 and 10). Second,
common that illustrate both the strengths and although it is true that thin high-permeability
weaknesses of the neural-network approach. The zones strongly inf luence the success of water-
neural network yielded more accurate predictions flood operations, the neural network does predict
than did linear regression (Figure 9; Table 1). Both the stratigraphic location and thickness of such
methods, however, are unsuccessful at predicting zones. Intervals characterized by locally high- and
very high permeability values that are surrounded low-permeability values in the “real” data exhibit
by substantially lower permeability values. The the same relative relationships to their neighbors
neural network does a poor job in these situations in the predicted data (Figures 10–12). Another
because of the averaging effect of the 25-point implication of this result is the observation that an
20
15
10
0
15453 –
–
–
–
15456 –
–
–
–
15459 –
–
–
–
15462 –
–
–
–
15465 –
–
–
–
15468 –
–
–
–
–
15471 –
–
–
–
15474 –
–
–
–
15477 –
–
–
–
15480 –
–
–
–
15483 –
–
–
–
15486 –
–
–
–
15489 –
–
–
–
15492 –
–
–
–
15495 –
–
–
–
15498 –
–
–
–
15501 –
–
–
–
15504 –
–
–
–
15507 –
–
–
–
15510 –
–
–
–
–
15513 –
–
–
–
15517 –
–
–
–
15520 –
–
–
–
15523 –
–
–
–
15526 –
–
Depth (ft)
1794 Predicting Permeability
Table 1. Mean Squared Error (in md) for BPNN and for Lithologic information was not used to predict
Linear Regression for the Three Scenarios* permeability for four reasons. First, and most
important, this project was specifically designed to
Scenario test a network that employed only spatial and
2 2 porosity information to predict permeability
Method 1 (Well 1704) (Well 1705) 3 because our goal was to develop a tool that could
BPNN 20.540 0.05400 0.08316 0.123035 be applied to any field, no matter how scanty the
Regression 101.978 0.07185 0.10127 0.110998 available data. (However, the architecture of the
network used makes it very simple to add more
Regression/
BPNN 4.964849 1.330555 1.217773 0.902166
kinds of input data if they are available.) Second,
lithologic units in the Smackover Formation, identi-
*BPNN = Backpropagation neural network. fied on the basis of mineralogy, rock fabric, pore
types, and sedimentary structures, do differ consis-
tently in porosity and per meability values
ANN is not used straight off the shelf. An ANN (Kopaska-Merkel et al., 1994). However, individual
ought to be designed for the problem at hand. In a lithofacies tend to exhibit a wide range in perme-
permeability-prediction study for evaluating a ability values, coupled with low correlation coeffi-
proposed or actual waterf lood operation, for cients for porosity and permeability within a single
instance, a narrow window should be used, or lithofacies. Thus, lithologic information is a crude
other characteristics incorporated into the model, predictor of permeability. Third, the effect of X–Y
such as recognition of the shape of the permeabil- variation in rock type on porosity and permeability
ity distribution vs. depth where permeability was minimized in this study by using a relatively lat-
changes dramatically over a short stratigraphic erally homogeneous part of the field (Bradford,
interval. This could enhance prediction of thin 1984). Fourth, porous and permeable strata in BEC
very high permeability zones. field are dominated by intercrystalline and mixed
2.5
Permeability (md)
1.5
0.5
0
15248 –
–
–
–
15252 –
–
–
–
15256 –
–
–
–
15260 –
–
–
–
15277 –
–
–
–
15281 –
–
–
–
15285 –
–
–
–
15289 –
–
–
–
15293 –
–
–
–
15297 –
–
–
15301 –
–
–
–
15305 –
–
–
–
15309 –
–
–
–
15313 –
–
–
–
15317 –
–
–
–
15321 –
–
–
–
15325 –
–
–
–
15329 –
–
–
–
15333 –
–
–
–
15337 –
–
–
–
15341 –
–
–
–
15345 –
–
–
–
15349 –
–
–
–
15353 –
–
–
–
15357 –
–
–
–
–
Depth (ft)
Rogers et al. 1795
2.5
Permeability (md)
1.5
0.5
0
15136 –
–
–
15140 –
–
–
–
15144 –
–
–
15148 –
–
–
–
15152 –
–
–
–
15156 –
–
–
15160 –
–
–
–
15164 –
–
–
15168 –
–
–
–
15172 –
–
–
15176 –
–
–
–
15180 –
–
–
–
15184 –
–
–
15188 –
–
–
–
15192 –
–
–
15196 –
–
–
–
15200 –
–
–
–
15204 –
–
–
15210 –
–
–
–
15214 –
–
–
15218 –
–
–
–
15222 –
–
–
15226 –
–
–
–
15230 –
–
–
–
15234 –
–
–
15238 –
–
–
–
15242 –
–
–
15246 –
–
–
–
Depth (ft)
pore systems (Kopaska-Merkel and Mann, 1991). of crystalline dolostone, which exhibit subtle differ-
This kind of pore system is controlled by deposi- ences in cr ystal size, shape, and distribution.
tional fabric only in the grossest sense (e.g., lime Hence, petrophysical properties, especially perme-
mudstone is commonly not dolomitized and there- ability, are relatively independent of rock type. Also,
fore is neither porous nor permeable). It is com- intercrystalline reservoirs exhibit consistently small
monly impractical to discriminate among subtypes aspect ratios (pore to throat size ratio), consistent
2.5
Permeability (md)
1.5
0.5
0
15246 –
–
–
–
15251 –
–
–
–
15256 –
–
–
–
15261 –
–
–
15279 –
–
–
–
15284 –
–
–
–
15289 –
–
–
–
15294 –
–
–
15299 –
–
–
–
15304 –
–
–
–
15309 –
–
–
–
15314 –
–
–
15319 –
–
–
–
15324 –
–
–
–
15329 –
–
–
–
15334 –
–
–
15339 –
–
–
–
15344 –
–
–
–
15349 –
–
–
–
15354 –
–
–
–
15359 –
–
–
15364 –
–
–
–
15369 –
–
–
–
15374 –
–
–
–
15379 –
–
–
–
15384 –
–
–
15389 –
–
Depth (ft)
1796 Predicting Permeability
pore and throat shapes, and are relatively homoge- Addison-Wesley, 433 p.
neous on a microscopic scale (Kopaska-Merkel and Hertz, J., A. Krogh, and R. G. Palmer, 1991, Introduction to the
theory of neural computation: Redwood City, California,
Mann, 1992). Consequently, porosity is the most Addison Wesley, 327 p.
useful predictor of permeability in intercrystalline Kopaska-Merkel, D. C., 1993, Capillary-pressure characteristics of
reservoirs. Smackover reservoirs in Alabama: Geological Survey of
Alabama Circular 170, 38 p.
Kopaska-Merkel, D. C., and D. R. Hall, 1993, Reservoir characteri-
zation of the Smackover Formation in southwest Alabama:
CONCLUSIONS Geological Survey of Alabama Bulletin 153, 111 p.
Kopaska-Merkel, D. C., and S. D. Mann, 1991, Pore facies of
We have demonstrated by way of three different Smackover carbonate reservoirs in southwest Alabama: Gulf
scenarios the utility and superiority of neural net- Coast Association of Geological Societies Transactions, v. 41,
p. 374–382.
works in predicting permeability values from Kopaska-Merkel, D. C., and S. D. Mann, 1992, Regional varia-
porosity values. A properly trained network would tion in microscopic and megascopic reservoir heterogeneity
predict permeability values in noncored wells from in the Smackover Formation, southwest Alabama: Gulf Coast
log-derived porosity values more accurately than Association of Geological Societies Transactions, v. 42,
p. 189–212.
would regression analysis. The ANN approach is Kopaska-Merkel, D. C., S. D. Mann, and J. W. Schmoker, 1994,
model free; that is, no functional form (such as lin- Controls on reservoir development in a shelf carbonate: Upper
ear, polynomial, or multiple) between permeability Jurassic Smackover Formation of Alabama: AAPG Bulletin,
and porosity values need be assumed; however, the v. 78, p. 938–959.
neural-network approach is but one of several pre- Nelson, P. H., 1994, Permeability-porosity relationships in sedi-
mentary rocks: The Log Analyst, v. 35, p. 38–62.
dictive tools available in the geologist’s toolbox. Osborne, D. A., 1992, Neural networks provide more accurate
Also, more field studies should be conducted to fur- reservoir permeability: Oil & Gas Journal, (Sept. 28),
ther explore the scope and power of this new p. 80–87.
approach. Rogers, S. J., J. H. Fang, C. L. Karr, and D. A. Stanley, 1992,
Determination of lithology from well logs using a neural net-
The network used in this study was specifically work: AAPG Bulletin, v. 76, p. 731–739.
designed to use only spatial and porosity informa- Vinet, M. J., 1984, Geochemistry and origin of Smackover and
tion to predict permeability. These kinds of informa- Buckner dolomites (Upper Jurassic), Jay Field area,
tion are available for nearly any hydrocarbon field Alabama–Florida, in W. P. S. Ventress, D. G. Bebout, B. F.
and are inexpensive to collect; however, the neural- Perkins, and C. H. Moore, eds., The Jurassic of the Gulf Rim:
Proceedings of the Third Annual Research Conference, Gulf
network architecture readily permits use of other Coast Section, SEPM, p. 365–374.
kinds of information to improve the performance of Wasserman, P. D., 1989, Neural computing: theory and practice:
the network, if such information is available. New York, Van Nostrand Reinhold, 230 p.
Wasserman, P. D., 1993, Advanced methods in neural computing:
New York, Van Nostrand Reinhold, 255 p.
REFERENCES CITED Weiner, J. M., J. A. Rogers, J. R. Rogers, and R. F. Moll, 1991,
Predicting carbonate permeabilities from wireline logs using a
Block, S., 1991, Empirical prediction of porosity and permeability back-propagation neural network, CM1.1: Society of
in sandstones: AAPG Bulletin, v. 75, p. 1145–1160. Exploration Geophysicists, Expanded Abstracts with
Bradford, C. A., 1984, Transgressive-regressive carbonate facies Biographies, Technical Program, v. 1, p. 285–289.
of the Smackover Formation, Escambia County, Alabama, in Wendt, W. A., S. Sakurai, and P. H. Nelson, 1986, Permeability pre-
W. P. S. Ventress, D. G. Bebout, B. F. Perkins, and C. H. diction from well logs using multiple regression, in L.W. Lake,
Moore, eds., The Jurassic of the Gulf Rim: Proceedings of the ed., Reservoir characterization: Orlando, Florida, Academic
Third Annual Research Conference, Gulf Coast Section, Press, p. 181–221.
SEPM, p. 27–39. Zurada, J. M., 1992, Introduction to artificial neural systems: New
Hecht-Nielsen, R., 1990, Neurocomputing: Reading, Massachusetts, York, West Publishing Co., 683 p.
Rogers et al. 1797
S. J. Rogers D. C. Kopaska-Merkel
Samuel J. Rogers is an application engineer for the David Kopaska-Merkel is ground water section head
Intergraph Corporation in Huntsville, Alabama. He at the Geological Survey of Alabama. He earned a Ph.D.
received his M.A. degree in computer science from the in geology from the University of Kansas in 1983 and has
University of Alabama. He is currently working on a book worked as a petroleum geolgist, carbonate sedimentolo-
adressing object-oriented neural-network architectures. gist, and hydrogeologist for Shell Oil Company, the
Northeastern Science Foundation, and the Geological
H. C. Chen Survey of Alabama.
Hiu-Chuan Chen is a professor of computer science at J. H. Fang
the University of Alabama. She received a Ph.D. from the
State University of New York at Buffalo in 1972, and was Jen-Ho Fang is a professor of geology at the University
a visiting scholar in the Department of Computer of Alabama. He received his B.S. degree from the
Science at Stanford University from 1979 to 1980. Her National Taiwan University, and his Ph.D. from
research interests include algorithm design, AI tech- Pennsylvania State University in 1961. He is a fellow of
niques, expert systems, neural networks, and fuzzy the Mineralogical Society of America and the Geological
logic. She has conducted research for JPL, the Society of America. He was once a mineralogist, and a
Department of Energy, Marine Terminal Computer mineral, “Fangite” (TI3ASS4), was named after him. He
Systems, and several oil companies. served as president of the Mathematical Geologists of the
United States.