Page 82 - Machine Learning for Subsurface Characterization

68 Machine learning for subsurface characterization

intervals include siltstone, sandstone, dolostone, and dolo-mudstone intervals.
Different minerals as present in these intervals, such as quartz, K-feldspar, pla-
gioclase feldspars, illite, dolomite, calcite, kaolinite, and pyrite.
The US and LS formations are black shale formations deposited during the
Late Devonian period to Early Mississippian period. The MS formation
displays a range of grain size sorting from poorly sorted, argillaceous siltstone
to moderately well sorted fine-grained sandstone and is more complex than
US and LS formations. MS can be categorized into upper middle shale (UMS)
and lower middle shale (LMS) based on their distinct grain sizes, depositional
textures and diagenetic calcite cement contents. UMS contains better reservoir
quality with well sorted, fine-grained sandstone, whereas LMS contains more
bioturbated, silt-dominated, shallow-marine deposits. CR 1–4 intervals are pri-
marily dolostone with alternating porous dolosiltite facies. The interlaminated
CR dolostone is interbedded with clay-rich, conglomeratic dolo-mudstone,
which marks stratigraphic intervals that partition CR into four distinct sequences
from top to bottom. CR1 is the principal oil-bearing interval of CR. CR2 is also
oil-bearingbutonlylocallychargedwithoil,mainlyinthecenterofthebasin.Itis
rare to find oil in CR3, and the remaining CR4 is nonreservoir dolo-mudstone.

2.2 Petrophysical basis for the proposed data-driven log synthesis
NMR logging tool is generally suited for uniform boreholes of diameter greater
than 6.5 inches run at a speed of around 2000 feet/hour. NMR responses of
subsurface formation due to the inherent physics have poor signal-to-noise ratio.
Highly trained logging engineers and good well conditions are required to
ensure high-quality NMR log acquisition in the subsurface borehole environ-
ment. After the data acquisition, the NMR logs need to be processed using robust
inversion methods to obtain the T 2 distribution. Due to the financial and oper-
ational challenges involved in running the NMR logging tool, oil and gas com-
panies do not deploy NMR logging tool in every well. One alternative is to train
data-driven models to process conventional logs for predicting the entire NMR
T 2 distribution spanning 0.3 ms to 3000 ms. An accurate prediction of NMR T 2
distribution will assist geoscientists and engineers to quantify the pore size dis-
tribution, permeability, and bound fluid saturation in hydrocarbon-bearing res-
ervoirs, thereby improving project economics and reservoir characterization
capabilities. The use of data-driven methods to predict the entire NMR T 2 dis-
tribution without core data is a challenging and novel task.
Generation of NMR T 2 distribution log using conventional “easy-to-acquire”
logs is feasible because the various combinations of conventional logs are sen-
sitive to fluid saturations, pore size distribution, and/or mineralogy.
For example, resistivity measurements and the mud-filtrate invasion effect on
resistivity logs are influenced by the pore sizes, pore throat sizes, tortuosity,
and permeability, which also influence the NMR T 2 distribution. Neutron log
is sensitive to hydrogen index of fluids in the formation that correlates with
porosity of the formation, which also influences the NMR T 2 distribution. GR