Page 257 - Machine Learning for Subsurface Characterization

Shallow and deep machine learning models Chapter 8 221

near-wellbore formation volume. The pore size distribution information
contained in the NMR logs is critical to evaluate the pore-scale phase
behavior of formation fluid, and no other well logs can provide this valuable
information. However, due to financial or operational considerations, NMR
logging tool cannot be deployed in all the wells drilled in a reservoir. To
overcome the limited access to the “hard-to-acquire” NMR logs, we
processed easy-to-acquire conventional logs using the shallow-learning and
deep learning methods to synthesize the in situ NMR T2 distributions.
At any depth, various well logs exhibit strong to intermediate correlations
because they are the responses from the same fluid-filled porous material. Well
logs measure various responses of a formation by utilizing various physical
excitations/fields/phenomena, such as electrical, acoustic, electromagnetic,
chemical, and thermal processes. Although different physical phenomena are
utilized, the responses measured by different well logging tools at the same
formation depth are from the same material, and each well log records certain
similar aspects of the formation properties. Consequently, there are several
complex relationships, dependencies, and distinct patterns between the “easy-
to-acquire” logs and the “hard-to-acquire” logs. Machine learning can find
hidden relationships and patterns in large data. Machine learning methods can
synthesize “hard-to-acquire” well logs by learning the relationships between the
“hard-to-acquire” well logs and the “easy-to-acquire” well logs.
Shallow-learning models have been widely applied in oil and gas industry
for prediction, classification, and regression tasks. For example, regression-type
models trained using supervised learning have been applied in pseudo-log
generation and petrophysical property prediction to improve reservoir
characterization. Among all shallow models, artificial neural network (ANN)
is the most widely used model due to its capability to account for nonlinear
trends; nonetheless, ANN predictions are hard to generalize, interpret and
explain. ANN has been successfully applied in pseudo-log generation [1–3]
and petrophysical property prediction [4–6]. Support vector regression has
been applied in the prediction of unconfined compressive strength from other
petrophysical properties [7]. LASSO and k-nearest neighbor regression
models have been applied in drilling and reservoir engineering tasks [8, 9].
Compared to shallow learning models, deep learning models are suited for
learning hierarchical features from large-sized high-dimensional dataset
without explicit feature extraction prior to the model training [10].Deep
learning models have been applied to log synthesis [11, 12] and petrophysical
characterization [13]. Various architectures for deep learning can be designed
specific to the dataset and learning task. For example, convolutional neural
networks are suitable for spatial data and have been applied to image analysis;
while recurrent neural networks are suitable for sequential data and have been
applied to production data analysis [14]. This chapter compares the
performances of shallow-learning and deep learning models trained to
synthesize NMR T2 distribution by processing 12 “easy-to-acquire”
conventional logs.