Page 73 - Machine Learning for Subsurface Characterization

Characterization of fracture-induced geomechanical alterations Chapter 2 59

FIG. 2.10 Comparison of the effect of using features derived from both prefracture and
postfracture waveforms. Noninvasive visualization of geomechanical alterations in the
postfracture Tennessee sandstone sample in the axial plane obtained by K-means clustering of
(left) 180 STFT-based features derived from only postfracture shear waveforms and (right) 360
STFT-based features extracted from both pre- and postfracture shear waveforms. For both the
cases, STFT-derived features were dimensionally reduced to 120 PCA-derived features, which
account for 98% of variance. Hotter colors indicate larger geomechanical alteration.

both pre- and postfracture features has 360 features. This study was done for
visualization in the axial plane for which both prefracture and postfracture
waveforms were measured. There are no prefracture waveforms for the
frontal plane.
In both situations, the clustering produces qualitatively similar alteration
indices. The maximum alteration zone is the middle of sample that coincides
with the acoustic emission. Surrounding regions show relatively lesser
alteration. Visualizations obtained by processing features from both pre- and
postfracture waveforms indicate a larger region of low alteration. It appears
using prefracture features in this case is equivalent to padding the feature set
with constant values. This fact is corroborated by the observation that the
same number of PCA-derived components (120) explained 98% variance in
both cases. This implies that addition of prefracture data did not introduce
any additional variance to the feature set. Overall, using prefracture data
may aggravate the curse of dimensionality and reduce the importance of
postfracture features for the clustering method.

7 Physical basis of the fracture-induced geomechanical
alteration index

According to the displacement discontinuity theory, a lower amplitude of the
first arrival and a delayed arrival of the shear waveforms indicate a reduction
in the rock stiffness due to fracturing [5]. This principle is used to assign a
geomechanical alteration index to the cluster labels/IDs generated by the