Page 88 - Machine Learning for Subsurface Characterization

Shallow neural networks and classification methods Chapter 3 73

Flag-5 captures difference in the deviation of pore size distributions about a
dominant pore size. These flags help improve the prediction performance as
they provide additional relevant information to the predictive models about
each depth.
An expert geologist provided us formation tops and dominant lithology
along the well length. We used this information to assign integer values to
Flag-1 for each depth. Values for Flags 2–5 were manually assigned to each
depthbyexamining theNMR T 2 distribution for that depth. After manually
creating the flags, which are categorical features, we trained a k-nearest neigh-
bor (KNN) classifier (supervised learning) to correctly predict the categorical
values of the five flags by processing a specific combination of conventional
logs and inversion-derived logs. KNNtrainingshouldbedoneonlyon the
training dataset used for training the ANN models for NMR T 2 synthesis. This
avoids contaminating the performances of the ANN models when synthesiz-
ing the NMR T 2 distributions for the testing dataset. Once the KNN is trained,
new conventional and inversion-derived logs from new wells without the
NMR T 2 distributions can be processed to generate the Flags 2–5. The
KNN predictions of Flags 2–5 along with Flag 1 obtained from the Geologist
are fed into the ANN models along with the conventional and inversion-
derived logs for purposes of NMR T 2 synthesis in the new wells. Our hypoth-
esis is that the Flags 1–5 act as additional informative features that help
improve the performances of the ANN models. Table 3.B1 lists the substantial
improvement in predictive performance of the first ANN model when the flags
were generated and then used as additional features when training and testing
the ANN model. Fig. 3.C1 shows that Flags 1–4 are very important features for
the proposed NMR T 2 synthesis task.
When creating the training dataset, Flags 2–5 can be easily generated by
examining the available T 2 distributions acquired along the well length and then
assigning specific values to the Flags for each training depth depending on the
characteristics of the T 2 distribution at that depth. However, for testing data or
any new data, T 2 distribution needs to be predicted and is unavailable; conse-
quently, we cannot manually assign specific values to these flags. So, for the
testing data and new data, we first train a KNN classifier on the training dataset
with manually assigned flags to relate the available conventional logs to the
flags. Following that, the trained KNN classifier is used to process the available
conventional logs in the testing dataset and new dataset to generate the flags.
Flags 2–5 need to be predicted prior to the primary objective of generating
the T 2 distribution. The goal of KNN classifier is to first learn to relate the avail-
able “easy-to-acquire” conventional and inversion-derived logs to the Flags 2–5
in the training dataset, in the presence of NMR T 2 distribution. The trained KNN
is then used to process the easy-to-acquire conventional logs to generate the
Flags 2–5, in the absence of NMR T 2 distribution. Twenty-two conventional
and inversion-derived logs are used as features to predict the Flags 2–5one
by one with four separate KNN classifiers.