Page 256 - Machine Learning for Subsurface Characterization

220 Machine learning for subsurface characterization

LASSO least absolute shrinkage and selection operator
LSTM long short-term memory
OLS ordinary least squares
SVC support vector classifier
SVR support vector regression
VAE variational autoencoder
w coefficient vector
X input log matrix
Y output log matrix
y i original log response at depth i

Subscripts
i formation (i)

1 Introduction
Economical reservoir development relies on accurate geological and
petrophysical characterizations. One characterization technique is based on
core samples acquired from the subsurface. Core samples provide a direct way
to analyze petrophysical and geomechanical properties of subsurface
formations. Acquiring core samples is expensive and is restricted by
operational constraints. Another characterization technique is based on well
logs that measure formation responses corresponding to various geophysical
phenomena. Different well logging tools utilizing different physical principles
are deployed in the wellbore environment for measuring the subsurface
formation responses. Certain well logs, such as density, natural radiation, and
resistivity logs, are “easy-to-acquire” measurements because of the relatively
simple tool design, tool physics requirements, and operational protocols. On
the other hand, measurements of NMR logs and imaging logs tend to be
expensive and prohibitive due to the tool size, complex tool physics, intricate
operational procedures, and slow logging speed. Such logs can be categorized
as “hard-to-acquire” logs. Well logs can be used in the raw form, such as
resistivity or density, or can be inverted/processed to obtain estimates of
certain desired physical properties, such as fluid saturations and mineral
composition. Formation properties like permeability and pore size distribution
can be inverted from the raw NMR logs.
NMR logging tool excites the hydrogen nuclei in the in situ subsurface fluids
by applying an external magnetic field, and the relaxation of the hydrogen
nuclei, upon the removal of the external magnetic field, generates a
relaxation signal that is inverted to obtain the NMR T2 distribution
comprising 64 T2 amplitudes corresponding to 64 T2 bins. NMR T2
distribution is the relaxation time distribution of relaxing hydrogen nuclei in
the formation fluid at a specific formation depth. NMR T2 distribution can
be processed to estimate the pore size distribution and fluid mobility of the