Page 214 - Machine Learning for Subsurface Characterization

184 Machine learning for subsurface characterization

1 Introduction
1.1 Log-based subsurface characterization

Subsurface characterization involves estimation, computation, and
measurement of the physical properties of the subsurface geological
formations. Surface-based deep sensing measurements, borehole-based near-
wellbore measurements (logs), and laboratory measurements of geological
core samples extracted from wellbores are interpreted using empirical,
numerical, and mechanistic models to quantify the physical properties of the
subsurface formations. Subsurface measurements (referred as logs), acquired
using downhole logging tools, sense the near-wellbore subsurface formation
volume by inducing/monitoring various physical/chemical processes.
Subsequently, relevant tool physics modeling and geophysical interpretation
models are used to process the logs for purposes of subsurface
characterization. For example, multifrequency electromagnetic logs are
processed using stochastic inversion for fluid saturation estimation [1],
various petrophysical models are used to process dielectric dispersion logs to
characterize hydrocarbon pore volume and salinity in shales [2],
electromagnetic short pulse borehole imaging method is used to characterize
cracks and rugosity [3], poroelastic inversion of sonic velocity logs improves
permeability characterization [4], and triaxial electromagnetic induction
measurement facilitates the estimation of dip and anisotropy of the
formation [5].
Use of logging tools, geophysical models, and inversion- and machine
learning-based data interpretation techniques for purposes of subsurface
characterization has been evolving with the advancements in sensor physics
and computational methods. For example, Wong et al. [6] classified well log
data into different lithofacies followed by the estimation of porosity and
permeability using genetic neural networks. Similarly, lithology
determination from well logs was performed by Chang et al. [7] in
Ordovician rock units in northern Kansas using fuzzy associative memory
neural network. Xu et al. [8] listed the recent advances in machine learning
applications on well logs for purposes of improved subsurface
characterization. He and Misra [9] used various architectures of shallow
neural networks to synthesize dielectric dispersion logs in shales. In another
application of neural networks and shallow learning methods, Li et al. [10]
generated compressional and shear wave travel times in shale oil reservoir
for improved geomechanical characterization. Other than simple machine
learning methods, there have been very limited public demonstrations of
development of deep learning methods and their applications for formation
evaluation and well log analysis.