Page 122 - Machine Learning for Subsurface Characterization

98 Machine learning for subsurface characterization

FIG. 3.C1 Feature importance for the first ANN model. The index numbers on the y-axis indicate
the various logs and flags, namely, #1–#5, Flags 1–5; #6, GR; #7, DPHZ; #8, NOPR; #9, PEFZ; #10,
RHOZ; #11, VCL; #12, AT10; #13, AT90; #14, DTCO; #15, DTSM; #16, VPVS; #17, total Sw;
#18, illite; #19, chlorite; #20, bound water; #21, quartz; #22, K-Feldspar; #23, calcite; #24, dolo-
mite; #25, anhydrite; #26, unflushed water; and #27, unflushed oil.

importance for the desired synthesis task. Sonic travel logs, which are sensitive
to effective porosity and rock consolidation, are the most important conven-
tional logs. Logs sensitive to porosity and clay content are also important for
the NMR synthesis. Inversion-derived logs exhibit higher feature importance
in comparison with the conventional logs. Among the inversion-derived logs,
the chlorite, unflushed oil, K-feldspar, unflushed water, and total water satura-
tion are the most important features, in order from high to low. AT90, VPVS,
anhydrite, PEFZ, VCL, dolomite, and illite have low ranking, indicating these
logs are not important for the proposed ANN-based predictive task. Deep sens-
ing logs, such as AT90, and logs that have correlation with other important logs
tend to have lower feature ranking.

Appendix D Estimations of specific reservoir parameters from
NMR T 2 distributions
The first parameter ϕ N is calculated by integrating the NMR T 2 distribution; in
other words, ϕ N is the summation of the 64 discrete amplitudes for the corre-
sponding 64 T 2 bins at each depth. The T 2 distribution is discretized into 64