Page 87 - Machine Learning for Subsurface Characterization

72 Machine learning for subsurface characterization

relatively similar statistical characteristics. Conventional logs acquired in CR1,
CR2, CR3, and CR4 also exhibit similar statistical characteristics. Statistical
parameters for most of the input logs in MS are closer to those of CR intervals
rather than those in US and LS intervals. Median values of VPVS, PEFZ, and
RHOZ are similar throughout the seven intervals. Median values of GR, AT10,
AT90, mineral compositions, bound water content, and fluid saturations vary
significantly across the seven intervals (Fig. 3.A1).
Coefficient of variation of a feature (log) is the ratio of standard deviation over
mean (S d /μ), which is used to measure the dispersion of the feature (Fig. 3.A2).
AT10 and AT90 show large dispersion in bottommost CR4 interval. DPHZ shows
abnormally high dispersion for the CR3 interval. RHOZ, feldspar, and dolomite
showabnormallyhighdispersionfor USintervalascompared withother intervals.
Skewness is the measurement of asymmetry of data about its mean. A few con-
ventional logs, such as AT10 and AT90, have large values of coefficient of var-
iation and skewness in all intervals. Most of the logs in LS, CR3, and CR4 exhibit
higher skewness compared with other intervals. There is a large variability in
dispersion of logs for the seven intervals under investigation (Fig. 3.A3).

2.4 Categorization of depths using flags

After selecting the conventional logs and data preprocessing, we generate five
synthetic categorical features (referred as Flags) for each depth point. These cat-
egorical features (flags) will be used as synthetic discrete-valued logs because
their implementation improves the performance of the ANN models in synthe-
sizing the NMR T 2 distribution. Flag-1 is an integer ranging from 1 to 7 iden-
tifying the seven distinct intervals intersected by the well, namely US, MS, LS,
CR1, CR2, CR3, and CR4. Fig. 3.B1 provides qualitative schematic descrip-
tions of the Flags 2–5. These flags are computed based on certain characteristics
of the T 2 distribution, and the underlying pore size distribution. Flag-2 is either 0
or 1 identifying unimodal and bimodal pore size distribution, respectively, at a
certain depth. Flag-3 is an indicator of pore sizes in a bimodal system, such that
its value is 1, 0, or 1 identifying the abundance of small pores, comparable
volumes of small and large pores, and abundance of large pores, respectively.
Similar to Flag-3, Flag-4 indicates the relative abundance of pores of certain
pore size in a bimodal system, such that Flag-4 is assigned a value of 1 when
certain pore size (either small pores or large pores) is negligible or else it is
assigned to be 0. Depths for which Flag-4 is assigned a value of 1 can be
regarded as unimodal distributions. Flag-5 defines the deviation/spread of pore
sizes around the two dominant pore sizes of a bimodal distribution, such that 1
indicates that the spreads around the two peaks are wide and 0 indicates either a
unimodal distribution or a narrow spread around the two dominant pore sizes. In
brief, Flag-1 classifies intervals based on lithology, Flag-2 identifies number of
peaks in the pore size distribution, Flag-3 identifies the dominant pore sizes in
bimodal pore systems, Flag-4 checks if certain pore sizes can be neglected, and