Page 142 - Machine Learning for Subsurface Characterization

118 Machine learning for subsurface characterization

Next, we determine the smallest set of features/inputs required for maintain-
ing the desirable log-synthesis performance using the SNN model. The 15 con-
ventional logs (features) that are fed into the SNN model may not be available
during the real-world deployment of the model. For such scenarios, it is useful
to figure out the smallest set of logs that can be fed into the SNN model without
significant drop in accuracy of log synthesis. This is determined by deleting fea-
tures one by one starting with the least important feature, as identified in
Fig. 4.4. We conclude that at least 11 most important features should be retained
to maintain an accuracy drop less than 10% in the log synthesis as compared to
when all the 15 logs/features are used to develop the SNN model. This set of 11
log inputs is obtained by removing RLA2, RHOZ, PEFZ, and RLA1 (Fig. 4.5).
As shown by the leftmost bar in Fig. 4.5, a set of features containing only the six
most important features (DTC, DTS, NPOR, lithology flag, RLA3, and VCL
logs) results in 17% accuracy drop in the log synthesis as compared to when
all the 15 logs/features are used to develop the SNN model. Notably, when
DTC, DTS, NPOR, and six resistivity logs of different depths of investigation
are retained and other low-importance logs (RHOZ, PEFZ, GR, DPHZ, VCL,
and lithology) are removed, we observe accuracy drop less than 10% as com-
pared to when all the 15 logs/features are used to develop the SNN model
(Fig. 4.6).
To study the sensitivity of the SNN model to noise in training/testing data,
20% Gaussian noise is added one at a time to each feature (inputs and conven-
tional logs), to the six resistivity log inputs together (i.e., 3.33% noise is added
to each resistivity log), and to the eight DD log outputs together (i.e., 2.5% noise
is added to each DD log). The sensitivity of the SNN model to noise in feature,
from the highest to lowest sensitivity, is as follows: resistivity, DTS, GR,
RHOZ, NPOR, VCL, DPHZ, DTC, and PEFZ (Fig. 4.7). Twenty percent of
overall noise in the six resistivity logs results in maximum reduction in the
log-synthesis performance by 4.5%. Log-synthesis performance is also highly

FIG. 4.5 Comparison of reduction in the performance of DD log synthesis by deleting the con-
ventional logs one by one based on the importance of the conventional log as described in Fig. 4.4.