Page 156 - Machine Learning for Subsurface Characterization

Robust geomechanical characterization Chapter 5 131

be used to improve subsurface characterization. This chapter has two
objectives: (1) develop a workflow to synthesize both DTC and DTS logs
from “easy-to-acquire” conventional well logs and (2) apply clustering
techniques to the “easy-to-acquire” conventional well logs to simultaneously
determine the reliability of the synthetic DTS and DTC logs. Novelty of this
study is to use unsupervised clustering methods to generate an indicator of
the reliability/accuracy of the DTC and DTS logs synthesized by the
shallow-learning regression methods when deployed in new wells.
Easy-to-acquire well logs can be processed using statistical and machine
learning methods to synthesize the advanced well logs that are not acquired
in most of the wells in a field. Researchers have explored the possibility of
synthesizing certain “hard-to-acquire” well logs under data constraint [1–3].
Several studies have tried to implement machine learning techniques to
determine sonic log from other well logs. One study used ANNs, adaptive
neurofuzzy inference system, and support vector machines to synthesize both
compressional and shear sonic travel time by processing GR, bulk density,
and neutron porosity [1]. In another study, shear wave velocity (reciprocal of
DTS) was predicted using fuzzy logic and ANNs [4, 5]. Apart from machine
learning algorithms, DTS and DTC have been predicted using empirical
equations [6, 7], empirical correlations [8], or self-consistent models [9].
Our study aims to evaluate the performances of six regression models,
developed using supervised learning, for simultaneous synthesis of DTC and
DTS logs by processing conventional “easy-to-acquire” logs. The shallow-
learning models implemented in this study are ordinary least squares (OLS),
partial least squares (PLS), least absolute shrinkage and selection operator
(LASSO), ElasticNet (combination of LASSO and ridge regression),
multivariate adaptive regression splines (MARS), and artificial neural
networks (ANN). The first four models are linear regression models, and the
last two models are nonlinear regression models. The 13 “easy-to-acquire”
conventional logs together with DTC and DTS logs acquired in two wells are
used for training and testing the machine learning methods.
A novel aspect of our study is the implementation of clustering techniques to
determine the reliability of the sonic logs synthesized using the regression models.
Our study shows that the K-means clustering method can process “easy-to-
acquire” logs to group together certain depths such that the group/cluster
assigned to any given depth is correlated to “a certain range of accuracy”
(referred as reliability) of the shallow-learning model in synthesizing the DTC
and DTS logs at that depth. We compared five clustering techniques to find the
one that can be best used to determine the reliability of the sonic-log synthesis
at any given depth. The five clustering algorithms tested in this study are K-
means, Gaussian mixture model, hierarchical clustering, density-based spatial
clustering of application with noise (DBSCAN), and self-organizing map
(SOM). Dimensionality reduction technique is used to facilitate the
visualization of the characteristics of each clustering technique.
In this chapter, 6 shallow-learning models and 5 clustering algorithms process
13 “easy-to-acquire” conventional logs to synthesize the compressional and shear