Summary
Two techniques of preprocessing data from core plugs have been investigated
to enhance the quality of synthetic permeability estimation from conventional
logs by use of artificial neural networks (ANNs). A first technique consists of
“cleaning” the core-plug data set by removing the measurements deemed
log-incompatible (i.e., those from plugs corresponding to log measurements
affected by shoulder-bed effect or layers with thickness below the vertical log
resolution). The second technique relies on building high-resolution digital
models of cored intervals by use of a process-oriented-modeling (POM)
approach—the core model is populated with permeability values from core plugs
and then upscaled to a log-equivalent support volume.
Synthetic permeability curves estimated with these techniques have been
compared to synthetic permeability curves estimated without core-data
preprocessing and to permeability estimated directly from core plugs and
properly calibrated permeability curves from a nuclear magnetic resonance (NMR)
log tool in a turbidite reservoir, the ground truth value being represented by
actual dynamic data. Results highlight that core-to-log scale effects play a
major role in the permeability estimation from conventional logs and show that
the proposed preprocessing techniques can be effective in improving
permeability prediction, because they significantly reduce cross-scaling
problems related to the differences in support volumes.
Strengths and weaknesses of the two preprocessing approaches also have been
compared. The first technique is faster, but its application is strongly
constrained by the statistical and geological representativeness of the
selected data set. This is because some lithologies go underrepresented so as
to question the use of estimation tools like ANNs. Conversely, the POM
preprocessing technique is more time-consuming and needs detailed core
descriptions, but has the great advantage of supplying—starting from core data
only—a reliable permeability curve that retains its validity at the log
scale.
Introduction
Permeability prediction in hydrocarbon reservoirs is probably the most
challenging issue that geologists, petrophysicists, and reservoir engineers
have to deal with. In particular, the availability of permeability curves in a
large number of wells is one of the most desired targets in a
reservoir-characterization study. In recent years, logging techniques such as
NMR have been developed that allow permeability curves to be generated along
reservoir intervals. Nevertheless, the availability of NMR logs is not the
rule: In the majority of the wells, especially those from older fields, the
only permeability measurement available comes from plugs sparsely sampled from
bottomhole cores. On the other hand, bottomhole cores are usually available
only in a few reservoir intervals and/or wells, whereas conventional log
recordings (natural gamma ray, density, and neutron) are available from nearly
every well. Attempts to correlate core permeability to porosity and/or other
conventional logs using mathematical/statistical tools date back to the early
1960s. Since then, regression analysis has been the most widely used approach
for permeability prediction: This approach assumes that the permeability vs.
porosity—or, alternatively, vs. conventional logs—functional relationships can
be known in advance. As a matter of fact, functional relationships are
unknown.
© 2007. Society of Petroleum Engineers
View full textPDF
(
2,257 KB
)
History
- Original manuscript received:
7 June 2006
- Meeting paper published:
24 September 2006
- Revised manuscript received:
28 May 2007
- Manuscript approved:
3 June 2007
- Version of record:
20 October 2007
View Cart
