Page 298 - Machine Learning for Subsurface Characterization

260 Machine learning for subsurface characterization

FIG. 9.13 Methodology for developing data-driven model for the noninvasive characterization of
static mechanical discontinuities in material: (A) one realization of 2D numerical model of material
containing discontinuities generated using DFN model and the location of one source/transmitter
and the 28 receivers/sensors, (B) FMM simulation of compressional wavefront propagation through
material shown in (A), (C) the arrival times computed at each sensor for 10,000 realizations, and
(D) nine data-driven classifiers are trained and tested on the dataset to learn to relate the
28-dimensional feature vector to the user-assigned label of a realization.

Our hypothesis is that the data-driven models can be developed and deployed for
noninvasive static fracture characterization under constrained sonic-measurement
scenario. To that end, we only focus on the arrival times of compressional wave-
front (due to single source) at 28 receiver/sensor locations. To develop models
under data-constrained scenario, we do not use shear wave, full waveforms, wave
reflections and phase changes, multiple sources/transmitters at various locations,
and hundreds of sensors/receivers. When data-driven models perform well in a
desired task under constrained data scenario, there is high likelihood that the
data-driven approach will perform significantly better when exposed to varied
measurements and larger datasets.
Fig. 9.12 presents one realization of 2D numerical model of material con-
taining discontinuities and the location of 1 source/transmitter and the 28
receivers/sensors. Numerical models (realizations) of material containing static