CSUG/SPE 137822

Well Log Cluster Analysis: An Innovative Tool for Unconventional Exploration

T. Euzen, E. Delamaide, SPE, T. Feuchtwanger, IFP Technologies; K.D. Kingsmith, K.D. Kingsmith Resource
Consultants

Copyright 2010, Society of Petroleum Engineers

This paper was prepared for presentation at the Canadian Unconventional Resources & International Petroleum Conference held in Calgary, Alberta, Canada, 19–21 October 2010.

This paper was selected for presentation by a CSUG/SPE program committee following review of information contained in an abstract submitted by the author(s). Contents of the paper have not been
reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect any position of the Society of Petroleum Engineers, its officers, or
members. Electronic reproduction, distribution, or storage of any part of this paper without the written consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is
restricted to an abstract of not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
Well Log Cluster Analysis: an Innovative Tool for Unconventional Exploration
Tristan Euzen, Eric Delamaide, Tom Feuchtwanger, Kim Kingsmith.

Exploring for economically viable accumulations of hydrocarbons, the task of explorationists is becoming considerably more
complex in mature basins because of the unconventional nature of most remaining resources. Analyzing a large volume of data is
required in order to develop a comprehensive understanding of reservoir distributions and/or their production performance
characteristics. Furthermore, the unconventional nature of reservoirs makes it challenging to interpret well log data and indentify
potential gas pay using conventional data analysis approaches.
Well log cluster analysis is an innovative approach that provides the explorationist with an efficient tool for analyzing, screening
and filtering a large volume of well log data, in order to identify and map potential hydrocarbon accumulations. The method
involves seeking high density areas (clusters) in the multivariate space of well log data, in order to define classes of similar log
responses. These classes called electrofacies can be calibrated using various sources of data such as core description or production
data and can then be used to process very quickly hundreds of wells.
Cluster analysis and electrofacies classification have been applied to develop unconventional gas prospects in the Upper Mannville
incised valley fills of the WCSB. The reservoirs dominantly consist of arkoses with complex mineralogy, and are characterized by
both conventional and unconventional production from low resistivity pays. A cluster analysis was first performed on 5 producing
wells used as training samples, and an electrofacies classification was then performed on 395 wells in an area of 28 townships. A
quality control of the results was done by checking the electrofacies interpretation in front of producing intervals from 41 wells.
The preliminary results are very promising, with 89% of the producing intervals used in the blind test properly predicted by the
electrofacies classification. Finally, potential gas bearing intervals identified by the electrofacies classification have been mapped
in order to define prospective areas. This approach has many potential applications in the field of unconventional hydrocarbons,
especially where conventional log analysis doesn’t work properly and where a large volume of well log data is available.

Introduction:
As the exploration maturity of the Western Canada Sedimentary Basin increases, finding and extracting the untapped oil and gas
resources becomes more challenging. This is due to the fact that a significant number of the high quality conventional reservoirs
has been discovered and extensively drilled over the last 30 years. Lower quality reservoirs, on the other hand, have been
overlooked because of the higher risks and lower rewards associated with them. However, due to the recent advances in drilling
and completion technologies, some of these reservoirs can now be produced economically. New opportunities can be found by
revisiting the huge amount of well data publically available and identifying reservoirs suitable for a specific well design and
completion technology. However, the time and cost involved in analyzing a large well log database in order to identify potential
gas pay can be prohibitive. Furthermore, the unconventional nature of many reservoirs, often characterized by complex mineralogy
and low permeability, makes it difficult to use conventional log analysis methods. In this paper, we propose an alternative and
cost-effective approach which tackles both challenges of analyzing large amount of well log data and identifying hydrocarbon in
2 CSUG/SPE 137822

unconventional reservoirs. This method is based on statistical analysis of well log data. Identification of high density clusters in the
multivariate log space enables electrofacies to be defined and associated with specific lithology and fluid content. Once properly
calibrated using training samples, electrofacies can then be automatically assigned to hundreds of wells instantly. This method has
been successfully applied to a log dataset of 395 wells in the Upper Mannville incised valley sandstone reservoirs in a 28
townships area of Central Alberta.

Theory and method

Cluster analysis:
Well log data are a set of depth values, each of which is associated with a set of log values corresponding to different physical
properties measured along the well path. In the multi-dimensional space of logs, each depth measurement can be viewed as a point
or “sample”, whose coordinates are the respective value of each log. In the case of a set of 2 or 3 logs, we can easily visualize a
sample as a point in a two or three-dimensional cross-plot. The principle remains the same when more than three variables are
involved, although it becomes more difficult to have a graphical representation of the data points.
The purpose of cluster analysis is to look for similarities/dissimilarities between data points in order to group them into classes.
In the multi-dimensional space of logs, the distance between data points is a measurement of their dissimilarities. Samples with
similar log responses will tend to form clusters, separated by areas with a lower density of points (Fig. 1). In order to identify the
clusters, we can use a density function that measures the number of samples located in a moving interval of a given length around
each data point (Cailly, 1999). If only one variable is used, the density function is a curve that can be approximated by smoothing a
histogram. In the case of two variables, the density function is a surface with low density troughs and high density hills.
Representing the density function with more than 3 variables is possible by using the Kittler algorithm (Kittler, 1976). It can be
described as a point exploring the entire dataset, following the steepest density gradient when climbing toward maxima, and the
most gentle density gradient when moving down toward minima. The resulting graphical representation of this mode mapping
process is a curve showing density peaks corresponding to the clusters (Fig. 2). The mode mapping provides a description of the
structure of the data without any a priori interpretation (Fournier, 1997). The modes (high density clusters) correspond to the most
representative samples of each potential class.

Electrofacies classification:
An electrofacies is a class defined by a unique set of log responses, which also corresponds to a cluster in the multivariate
space of log data. For instance, shale would typically form a population of data point characterized by a high gamma ray, a high
neutron porosity and a very low or no SP defletion. The log signature of a given electrofacies reflects the chemical and physical
properties of the rocks and fluids contained in the volume investigated by the logging tools. A geological facies, on the other hand,
is generally based on the observation of lithology, texture and sedimentary structures and their interpretation in terms of
depositional or diagenetic processes. The relation between electrofacies and geological facies is not straightforward because they
are based on different types of data and defined at different scales. Traditionally, electrofacies are calibrated on core data to insure
their consistency with the geological interpretation, and then used as training samples to predict the geological facies from logs
where core is not available (Lim et al, 1997; Hohn et al, 1997; Chautru et al, 2005; Crampin, 2008; Handwerger et al, 2008; Garner
et al, 2009). The training samples can be selected by using a supervised or an unsupervised approach. In the supervised approach,
the training samples are defined based on the geological interpretation alone. In the unsupervised approach, the training samples
are identified on the density peaks from the cluster analysis, and associated a posteriori with the geological interpretation. This
method insures that the electrofacies are consistent with the structure of the data and are thus statistically meaningful. In practice, a
combination of these two approaches enables the user to optimize the integration of information from both statistical analysis of
log data and geological interpretation.
Once the training samples are defined, electrofacies are automatically assigned by the software to all other samples of the
study. For each sample, the probability of belonging to a class is computed using the Bayes formula and the sample is then
allocated to its most probable belonging class (Fournier, 1998; Cailly, 1999). The result of the assignment depends on the
intraclass probability laws. In this study, we used a parametric method (multigaussian laws) with a linear hypothesis (constant
covariance), because it was found to give a more consistent and stable result.

Application and results

Electrofacies are generally used to capture the variability of the lithology from a well log dataset, in order to better predict
reservoir heterogeneity. In such an approach, the selection of the log data and the definition of electrofacies are optimized for
characterizing rock types and minimizing the impact of fluid content on the classification (Crampin, 2008; Lim et al, 1997).
Electrofacies can then be used as an input for constructing geocellular reservoir models (Ravenne et al, 2000; Chautru et al, 2005;
Garner et al, 2009).
CSUG/SPE 137822 3

In this study, we propose a different application of electrofacies which rather focuses on the identification of fluids (water
versus gas in this case), and where no attempt is made to capture subtle variations of lithofacies. The electrofacies are calibrated
using producing intervals and a qualitative interpretation of well logs, and the lithofacies model is kept as simple as possible. The
objective of this approach is to screen and filter a large quantity of well log data, in order to automatically detect and map potential
gas bearing sandstone intervals. The classification output can then be used to identify gas accumulation and to define prospective
areas on which to focus for more detailed geological mapping and interpretation.

Case study:
The study covers an area of 28 townships in Central Alberta, North-West from Edmonton. In this area, the Upper Mannville
series consists of an overall prograding succession ranging from offshore to alluvial deposits (Deschamps et al, 2008). This
progradational trend was punctuated by higher frequency relative sea-level falls (forced regressions) associated with the formation
of incised valleys (Cant, 1996). The fluvial to estuarine sandstones that fill these incised valleys represent the most prolific gas
production zones of the Upper Mannville in the area. However, this play remains relatively immature in regard to exploration of
gas charged reservoir because of the following reasons:
- Most of the wells drilled in the area were targeting deeper reservoirs in the Lower Mannville or below the pre-cretaceous
unconformity.
- Incised valley fills form discontinuous sand bodies and subtle gas traps involving stratigraphic and structural components.
- The complex mineralogy of the sandstone reservoirs has a strong impact on the resistivity and porosity logs and makes it
challenging to identify gas pay (Boyd et al, 1995).
About 50 bcf of gas have been produced to date from the study area in the Upper Mannville, half of it coming from incised
valley fills and the other half from thin widespread sandstones at the very top of the Mannville and coastal deposits from the lower
part of the Upper Mannville. A good vertical well can produce more than 3 bcf of cumulative gas from a Upper Mannville incised
valley fill reservoir.
The reservoir quality of the incised valley fills is highly variable and largely controlled by the amount of carbonate cements
and clays (Euzen et al, 2010). This variability has an impact on the productivity of wells, and hydraulic fracturing may be required
to achieve economic production.

Database construction:
A set of 395 wells were selected based on the availability of digital logs in the study area (LAS files). Logs were selected based
on their sensitivity to lithology and to fluids, as well as on their widespread distribution. These logs include Gamma Ray (GR),
Normalized Spontaneous Potential (SPN), Deep Resistivity (DR), Density (RHO) and Neutron Porosity (NPHI). The Sonic and the
PEF logs were discarded because only respectively 64% and 69% of the wells had these curves available in the study area.
The well logs were imported into a multi-disciplinary 1D data processing and editing tool. Statistical analysis of log data and
electrofacies classification is one of the functionalities of this software.
The first step of the database construction consisted of a data quality control and normalization of logs when required. Cross-
plots of average log values over the studied interval (Upper Mannville) were used to indentify outliers and to discard them if their
anomalous log values were found to be related to poor data quality. Gamma Ray was normalized using the average log value of the
Joli Fou Formation, a consistent and homogeneous shale layer over the study area. SP was normalized between 0 and -100,
because absolute values depend primarily on the difference of salinity between the mud filtrate and the formation water. The
normalization was done by using the minimum (shale baseline) and the maximum deflection values encountered in the Mannville
at each well. The SP log was discarded when shale baseline was found to be drifting over the study interval.

Calibration of electrofacies:
The objective of this electrofacies classification was to discriminate between three classes: wet sandstone, tight gas sandstone
and conventional gas sandstone. Ideally, the lithology would have been calibrated on core. However, in this case, as no core was
available, the reservoir and fluid assessments was based on qualitative interpretation of well logs and production performance of
producing intervals. Wet sandstone was interpreted where a well defined gas-water contact was visible on the resistivity log, in
conjunction with an early water breakthrough from a production zone located close to the contact (Fig. 3a). Wet sandstone
intervals typically show deep resistivity below 3.5 Ohm.m and a fair to strong SP deflection. Conventional gas sandstone was
interpreted in intervals producing gas with high initial rates, without well stimulation. Conventional gas producing intervals
generally have deep resistivity above 5 Ohm.m, but they seldom show a neutron-density crossover (Fig. 3b). This is due to the
presence of clay and some heavy minerals (siderite, pyrite) that affect the log response. Tight gas sandstone was interpreted in
intervals with limited or no SP deflection, and producing with very low gas rates, unless stimulated. Tight gas sandstone typically
shows a progressive increase of the deep resistivity throughout the reservoir, and a larger separation between the density and
4 CSUG/SPE 137822

neutron porosity logs than the conventional gas sandstone (Fig. 3c). Three other classes were defined on the basis of the well log
interpretation: shale, coal and compact layers.
A critical step in the electrofacies calibration is to insure that the classes defined by the geological interpretation, correspond to
distinct clusters in the multivariate space of the selected logs. A set of 5 producing wells with log responses typical of the three
sandstone classes described above, were selected in order to define the training samples for the classification. This was done by
performing a mode mapping from the logs of these 5 producing wells. The samples of the density peaks were identified on the well
logs and it was found that each peak corresponded to one of the classes defined above (see Fig. 2). This means that the set of
selected wells and logs provided an appropriate training sample for discriminating between shale, wet sandstone, tight gas
sandstone and gas sandstone. On the other hand, the two other classes - coal and compact layer - didn’t correspond to well-defined
peaks on the density curve. This is because these facies form thin layers, resulting in a larger dispersion of the data points in the
multivariate space due to border effects. A supervised approach was used for these two classes, by selecting the training samples
directly on the well logs.

Electrofacies assignment and quality control:

The electrofacies were automatically assigned by the software to the 395 wells in the study area. The quality control of the
results was done by performing a blind test on 41wells (with 44 producing intervals), which were not used in the training sample
set. The objective of this blind test was to verify if the presence of gas was correctly predicted by the electrofacies classification in
the producing intervals. Tight gas or gas sandstone was predicted in 39 of the 44 producing intervals, which represents a success
ratio of 89%. Five producing intervals were improperly classified in the wet sandstone class instead of gas sandstone. An analysis
of the log values of the training wells classification suggests that the most discriminant logs for differentiating gas sandstone from
wet sandstone are the deep resistivity and the normalized SP (Fig. 4). The mismatched samples are in the high range of SPN
deflection between -49 and -75, compared to the gas sandstone class of the training wells. This discrepancy was corrected by
adding a cut-off rule switching samples with deep resitivity higher than 5 Ohm.m from wet sandstone to gas sandstone class. Even
the mismatched interval with an average deep resistivity of 4 ohm.m (see Fig. 4), had its upper part classified as a gas sandstone
class after the cut-off, which raised the success ratio of the blind test to 100%.
The distinction between tight gas sandstone and gas sandstone is more ambiguous because (1) there is a continuum in terms of
rock properties between these two classes and (2) there is a large range of water saturation in tight gas sandstone due to increasing
capillary forces with decreasing permeability. This type of ambiguity is inherent in any classification when dealing with
continuous phenomena. Fig. 5 shows two producing intervals from the blind test that were classified respectively in gas sandstone
and tight gas sandstone classes. The gas sandstone interval (Fig.5a) had an initial production rate (IP) of 800 mcf/d at an intitial
reservoir pressure of 7300 kpa, without stimulation. The dominantly tight gas interval (Fig. 5b) had an IP of 850 mcf/d at an initial
reservoir pressure of 9500 kpa, after hydraulic fracturing. This interval has suppressed SPN deflection and higher separation
between neutron and density porosities than the conventional gas interval. The result of the electrofacies classification is consistent
with both log interpretation and production data.

Practical application of electrofacies in exploration.

In this application, electrofacies classification is used as a mapping tool and to identify potential gas accumulation. For each
stratigraphic sequence, the cumulated thicknesses of wet, gas and tight gas sandstone electrofacies were mapped to identify areas
with higher potential of gas accumulation. Integrating this classification result with incised valley trend maps provided a good
basis for selecting prospective areas on which to focus for more detailed work. This detailed work includes a quick look log
interpretation of potential gas bearing intervals, a mapping of fluid distribution and an interpretation of potential trapping
mechanisms. In intervals where visual log interpretation and mapping confirms a high potential for gas pay, additional information
are provided by cuttings analysis and drilling and completion reports.
Electrofacies classification can also provide useful information on the relative proportion of gas charged versus wet and non-
reservoir rocks, in a given area. If we assume that these proportions are statistically representative of an area when derived from a
large number of wells (here 395 wells for a 28 township area), it can be used for calculating resource in place.

Discussion
The originality of the method described in this paper resides in the use of electrofacies as an exploration tool for identifying gas
accumulations. As no core was available, the training samples were calibrated on log and production data. However, production
performance depends on many factors other than reservoir quality and fluid content, including but not limited to reservoir pressure,
well design, completion and potential damage. A careful integration of production and completion data, pressure data when
available with qualitative log analysis is required for an optimal choice of the training samples and calibration of the electrofacies.
The choice of the training samples has a drastic impact on the electrofacies classification. We have seen that a few producing
intervals used in our blind test were mismatched because of their strong SPN deflection compared to the training sample of the gas
CSUG/SPE 137822 5

sandstone class. In this case we have used a cut-off on the deep resistivity to correct this discrepancy, because of the otherwise
good result of the blind test (89% of success). The representativity of the training samples can also be challenged by the non-
stationarity of variables at regional scale. For instance, mechanical compaction induces a progressive increase of density with
depth. In our study, we chose the training samples in order to cover the whole depth range of the area, of about 500 m and the
increasing density with depth had no noticeable impact on the classification.
The selection of logs also has a major impact on the electrofacies classification. In this study, the use of deep resistivity and
normalized SP was crucial for identifying fluids and tight sandstone.
The simplistic lithology model used in this electrofacies classification has been designed to focus on the fluid distribution at
sub-regional scale. To capture subtle heterogeneity at reservoir scale, a conventional use of electrofacies classification based on
rock types and calibrated on cuttings or cores is required.

Summary and Conclusion:

A method using well log cluster analysis and electrofacies classification for exploration purpose has been presented. This
method aims at identifying areas with higher potential of hydrocarbons accumulation. A blind test performed on 41 producing
wells has demonstrated that the electrofacies classification consistently predicts gas producing intervals. The classification was run
on 395 wells in a 28 township area and the results were used to identify potential gas accumulation and prospective areas.
This approach is particularly well adapted for exploration in mature basins, where large amounts of well data are available and
where reservoir complexity makes conventional log analysis less effective.

Acknowledgement
The authors would like to thank Kara McKay for her hard work in gathering the data, building the database and clearing the
path. We are also grateful to Beicip-Franlab for their technical support on Easytrace.

References
Boyd, A., Darling, H., Tabanou, J., Davis, B., Lyon, B., Flaum, C., Klein, J., Sneider, R.M., Sibbit, A., and Singer, J. 1995. The lowdown on
low-resistivity pay. Schlumberger Oilfield Review, 7, ( 3): 4-18.
Cailly, F. 1999. Mise en oeuvre de l'application électrofaciès d'Easytrace, IFP, internal report 45722.
Cant, D.J. 1996. Sedimentological and sequence stratigraphic organization of a foreland clastic wedge, Mannville Group, Western Canada Basin.
Journal of Sedimentary Research, 66, p. 1137-1147.
Chautru, J.M., Arab, H. Fujita, N., Plucherie, E., Poli, E., Shebl, H., El-Henshiri, M.Y. and Shikama, T. 2005. An integrated RRT prediction
method combining Thin Sections, Sequence Stratigraphy and logs. Paper SPE 93563 presented at the SPE Middle East Oil and Gas Show
and Conference, Kingdom of Bahrain, 12-15 March. DOI: 10.2118/93563-MS.
Crampin, T. 2008. Well lof facies classification for improved regional exploration. Exploration Geophysics, 39, p. 115-123.
Deschamps, R., Euzen, T., Delamaide E., Feuchtwanger, T. and Pruden A. 2008. The Upper Mannville Incised Valleys of Central Alberta: An
Example of Subtle Gas Traps. extended abstract, CSPG CSEG CWLS Joint Annual Convention. Calgary, 12-14 May.
Euzen, T., Power, M, Sliwinski, J., Lenormand, R., Deschamps, R. and Durand, O. 2010. Advanced reservoir characterization: key to unlocking
gas resources of the Upper Mannville incised valley systems (Lower Cretaceous, West-central Alberta). abstract, CSPG CSEG CWLS Joint
Annual Convention, Calgary, 10-14 May.
Fournier, F., 1997: Méthode statistique de classements d'évènements liés aux propriétés physiques d'un milieu complexe tel que le sous-sol.
Brevet EN 97/12.182 N°2.768.818 (France 22/09/97), brevet US n°6.052651 (USA, 21/09/98).
Garner, D.L., Woo, A. and Broughton, P. 2009. Application of 1D electro-facies modeling. abstract, CSPG CSEG CWLS Joint Annual
Convention, Calgary, 4-9 May.
Handwerger, D., Vaughn G., Greaves, K., Milner, M. and Yang, Y.K., 2008. Improved characterization of heterogeneous reservoirs using cluster
analysis of log data. abstract, CSPG CSEG CWLS Joint Annual Convention. Calgary, 12-14 May.
Hohn, M.E., McDowell, R.R, Matchen, D.L and Vargo, A.G. 1997. Heterogeneity of fluvial-deltaic reservoirs in the Appalachian Basin; a case
study from the Lower Mississippian oil field in central West Virginia. AAPG Bulletin, 81 (6). 918-936.
Kittler, J. 1976. A locally sensitive method for cluster analysis. Pattern Recognition, 8, 23-33.
Lim, Jong-Se, Kang, J.M. and Kim, J. 1987. Multivariate Statistical Analysis for Automatic Electrofacies Determination from Well Log
Measurements. Paper SPE 38028 presented at the SPE Asia Pacific Oil and Gas Conference and Exhibition, Kuala Lumpur, Malaysia, 14-
16 April. DOI: 10.2118/38028-MS.
Ravenne, C., Cole, J. and Coury, Y. 2000. Characterization of reservoir heterogeneities and super permeability thief zones in major oilfield in the
Middle East. Proc., sixteenth World Petroleum Congress, Calgary, 11-15 June.
6 CSUG/SPE 137822

Figure 1: 3-D representation of clusters in multivariate space of log

Figure 2: Interpretation of the density peaks from the mode mapping curve.
CSUG/SPE 137822 7

Figure 3: Examples of intervals used as training samples

Figure 4: Position of the sandstone electrofacies of the training wells in SP versus deep resistivity cross-plot.
8 CSUG/SPE 137822

Figure 5: Comparison of the electrofacies classification of (a) conventional and (b) tight gas producing intervals used in the blind test.