Professional Documents
Culture Documents
Summary
Phi-k transforms are used widely to predict permeability. Some of
the difficulties of this exercise are well identified, such as the
homogeneity of the population (rock typing), the matching of cores
and logs (especially depth matching), and the problem of permeability upscaling. Not so well-known, however, are the pitfalls
of a statistical and geostatistical nature that may create significant biasesalways in the same directionan underestimation
of permeability.
The passage from Phi to k is performed in three steps: (1) in
cored wells, an exponential regression equation is established between core porosity and core permeability k; (2) in uncored wells,
log porosity is used instead as input to predict permeability; and
(3) the same equation is sometimes used again to populate the cells
of a dynamic reservoir model in 3D, where input porosity values
are obtained by interpolation.
The core-scale regression equation generally underestimates
permeability by at least a factor of 2. The origin of the bias lies in
the reverse transformation from logarithmic to arithmetic scale. To
avoid this pitfall, a new permeability estimator is proposed, based on
the quantile curves of the Phi-k crossplot. This estimator is data driven
and does not assume a priori any particular functional relationship
between Phi and k, such as an exponential-regression function.
One of the simplest diagnostic tools to check the agreement
between log and core porosity is a crossplot of one against the
other. In the absence of bias, the points are expected to be distributed along the yx line. In reality, they either are or they are not,
according to which variable is plotted along the x-axis. This apparent paradox is elucidated by bivariate regression theory and
related to the difference of investigated volume between core and
log data.
Direct input of upscaled cell porosity into an exponential corescale permeability transform amounts to forcing geometric permeability averaging, which may again lead to serious underestimation
of the true upscaled permeability when heterogeneity is significant.
Introduction
Porosity-permeability correlations are often used to predict permeability and to populate the cells of a dynamic reservoir model. The
passage from Phi to k typically involves three steps:
1. In cored wells, a regression equation, or transform, is established between core porosity and core permeability, or more exactly, between core porosity and the logarithm of permeability.
2. In uncored wells, log-derived porosity is used as input to this
equation to predict permeability.
3. The same equation is sometimes used again to distribute
permeability in 3D at the scale of the cells of a reservoir model,
where input porosity values are now obtained by interpolation
because most cells are not traversed by a well.
Each step has its own complexities and pitfalls. This paper will
mention just a few.
In cored wells, it is important to ensure that the core measurements are representative and reliable and will not bias the Phi-k
law (k and K denote point and block permeabilities, respectively).
Plugs with microfissures enhancing permeability should be removed. Likewise, oven drying may cause clay collapse and per-
meability increases by a factor of 5 to 30 (Soeder 1986). [Scanning-electron-microscope (SEM) images show how clay fibers can
be smashed against the pore walls (Pallat et al. 1984)]. On the
other hand, very-high permeability data may be missing because of
poor core recovery in unconsolidated sands. All these biases
should be removed from the analysis. Of course, the porositypermeability correlation should be established only in reservoir
sections, and is generally improved by working rock type by rock
type (i.e., rocks that have a similar porous network resulting from
a common geological history). As k ranges over several orders of
magnitude, porosity-permeability data are usually plotted in semilog scale, which, as we will see, is not without consequences.
In uncored wells, a Phi log is usually available and is substituted for Phi core in the porosity-permeability equation. This is
valid if, for the rock type considered, the two measurements are
consistent, which cannot be taken for granted. The porosity measured in the laboratory on core samples depends on the process of
core cleaning and drying, but is closest to total porosity. Log
analysis provides both total and effective porosity (i.e., excluding
clay-bound water in shaly formations). When comparing logderived porosity with core porosity, one must make sure to use
total porosity. Another source of discrepancy is the difference
between the volume measured by a core plug and the volume
investigated by logs.
Finally, when comparing core and log data, be it porosity or
permeability, correct depth matching is critical because minimal
differences in depths may modify the correspondence between log
and core measurements drastically and lead to erroneous conclusions.
The final step is to populate the cells of a dynamic reservoir
model with petrophysical data. Because most of the cells (say,
99.8% of them) are not traversed by a well, sophisticated geostatistical techniques are used to interpolate porosity from well
data, guided by geological trends and combined, if possible, with
seismic attributes. Porosity is a key parameter controlling volume
in place and is generally validated by different disciplines, whereas
permeability belongs to reservoir engineers. A correlation established at the scale of a 11-in. cylindrical core plug may not be
used for a 10010010-m grid cell. Direct use of the core-scale
Phi-k transform leads to a permeability field that can be significantly pessimistic. It amounts to geometric averaging, which
generally is not the correct upscaling method in 3D. In the
presence of strong permeability contrasts, the bias between the
correct upscaled permeability and the geometric mean can be quite
large. The same is true for the arithmetic mean, which is even
more sensitive to permeability dispersion. In this paper, much is
made of the need to get the arithmetic mean correct, not because
it is necessarily the correct averaging technique, but to avoid estimation bias. Permeability estimation and upscaling are two different problems.
In the above, we have discussed core only as the source of
permeability data. Wireline formation testing (WFT) also provides
dynamic information in the form of mobility data. An example will
show that the methods presented here also apply. The dynamic
data that reservoir engineers seek, however, are those of well
tests because they are more representative of the reservoir scale.
Phi-k transforms also remain useful in this case to redistribute the
global permeability value provided by the test vertically over the
tested interval.
Phi-k Transform
Bias in the Standard Transform. A Phi log k crossplot based on
261 plug measurements from a reference well is shown in Fig. 1.
The linear regression explains 41 % of the variance of log k, which
609
Fig. 4Pseudoflow profiles: the actual, that predicted by exponential regression, and that predicted by Swansons formula.
December 2007 SPE Reservoir Evaluation & Engineering
The Swanson curve in Fig. 4 shows the improvement in the predicted pseudoflow profile.
It is interesting to superimpose the straight-line fit of Fig. 1
(exponential regression function) on the quantile curves of Fig. 6.
We observe that it is close to the X50 curve. As linear-regression
estimates the mean of log k given Phi (conditional distribution), we
conclude that the mean and the median of log k are close, which
would suggest that log k given Phi has a symmetric distribution,
such as a normal one (i.e., k log normal). But if this were true, the
X50 curve should be halfway between X10 and X90, which is not the
caseX50 being closer to X90, especially for larger values of Phi.
We also observe that the X90 curve levels off, indicating a kind of
capping of permeability. As noted earlier, the log normal model is
not appropriate for the present data setthis is not a problem
Swanson = 41.47.
A Simple Geostatistical Model. We now propose a simple geostatistical model explaining the regression equations. It is based on
the notion of support, which is the rock volume over which a
physical quantity is measured or defined (Chils and Delfiner
1999). It is difficult to specify exactly the volume investigated by
porosity logs; moreover, it depends on the logging tool and on the
formation properties themselves. Surely this volume is much larger
than the volume of a plug (say 1,000 times). A larger support
means a smaller dispersion, which explains why Phi-log has a
lower standard deviation than Phi-core (Table 1). We may regard
the log measurement as a spatial average of plug measurements, or
equivalently, we may regard the measured plug as a random
sample from the population of all the plugs that can be taken
within the volume seen by the logs. The mean value of Phi-core
over all these possible locations is equal to Phi-log.
E -core | -log) = -log. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (8)
This theoretical model (slope1, intercept0) is consistent with
the experimental regression parameters (slope1.023, intercept 0.175). Furthermore, Eq. 8 implies the following additive model:
1
|V|
kx dx
, . . . . . . . . . . . . . . . . . . . . . . . . . . . (12)
Fig. 13Phi-core/Phi-log is uncorrelated with -log, and positively correlated with -core.
614
Fig. 14Homogeneous interval, (a) estimate from upscaled Phi and (b) arithmetic permeability averaging.
Fig. 15Heterogeneous interval, (a) estimate from upscaled Phi and (b) arithmetic permeability averaging.
Fig. 16Porosity and permeability trend, (a) estimate from upscaled Phi and (b) arithmetic permeability averaging.
December 2007 SPE Reservoir Evaluation & Engineering
615
616
2
i
Acknowledgments
The author is indebted to Martin Santacoloma and Bernard Lebon
for providing examples and insight, to Bruno Lalanne for reviewing the manuscript, and to Grard Massonnat for fruitful discussions on upscaling. Thanks go to the management of Total for
permission to publish this paper.
References
Brown, G.C., Hurst, A., and Swanson, R.I. 2000. Swansons 304030
Rule. The American Association of Petroleum Geologists Bulletin 84
(12): 18831891.
Chils, J.P. and Delfiner, P. 1999. Geostatistics: Modeling Spatial Uncertainty, New York City: John Wiley & Sons.
Knecht, L., Mathis, B., Leduc, J.P., Vandenabeele, T., and Di Cuia, R.
2004. Electrofacies and Permeability Modeling in Carbonate Reservoirs using Image Texture Analysis and Clustering Tools. Petrophysics
45 (1): 2737.
Matheron, G. 1967. Elments Pour une Thorie des Milieux Poreux. Paris:
Masson.
Noetinger, B. and Haas, A. 1996. Permeability Averaging for Well Tests in
3D Stochastic Reservoir Models. Paper SPE 36653 presented at the
SPE Annual Technical Conference and Exhibition, Denver, 69 October. DOI: 10.2118/36653-MS.
Noetinger, B. and Zargar, G. 2004. Multiscale Description and Upscaling
of Fluid Flow in Subsurface Reservoirs. Oil & Gas Science and TechnologyRev. IFP 59 (2): 119140.
Pallat, N., Wilson, J., and McHardy, W.J. 1984. The Relationship Between
Permeability and the Morphology of Diagenetic Illite in Reservoir
Rocks. JPT 36 (12): 22252227. SPE-12978-PA. DOI: 10.2118/12978PA.
Pearson, E.S. and Tukey, J.W. 1965. Approximate Means and Standard
Deviations Based on Distances Between Percentage Points of Frequency Curves. Biometrika 52 (34): 533546.
Renard, P. and de Marsily, G. 1997. Calculating Equivalent Permeability:
a Review. Advances in Water Resources 20 (56): 253278.
Soeder, D.J. 1986. Laboratory Drying Procedures and the Permeability of
Tight Sandstone Core. SPEFE 1 (1): 16. SPE-11622-PA. DOI:
10.2118/11622-PA.
Worthington, P.F. 2004. The Effect of Scale on the Petrophysical Estimation of Intergranular Permeability. Petrophysics 45 (1): 5972.
Pierre Delfiner is scientific advisor to the director of geosciences of Total in Paris. E-mail: pierre.delfiner@total.com. He
specializes in decision and risk analysis, probabilistic reserves
estimation, and geostatistical modeling. Delfiners recent work
includes prospect evaluation methodology, appraisal of
blocks with multiple objectives, aggregation of gas resources
for LNG projects, and assessments of the value of information
of 4D seismic for reservoir monitoring. He served as a Review
Chair for SPE Reservoir Evaluation & Engineering. Previously,
Delfiner was with Schlumberger Wireline, in charge of processing and interpretation software development, notably formation imaging. Before that, he was research associate in Professor G. Matherons team at the Center for Geostatistics of the
Ecole des Mines de Paris, developing Kriging and stochastic
modeling methods for the petroleum industry. Delfiner holds an
engineering degree from the Ecole des Mines and a PhD degree in statistics from Princeton University.
617