Page 101 - Machine Learning for Subsurface Characterization

84 Machine learning for subsurface characterization

3.3 Testing the first ANN model
There are only 62 testing depths, which are comparatively fewer than the 354
depths for training the ANN model. The prediction performance on the testing
dataset (also referred as the generalization performance) is similar to that on the
training dataset. Fig. 3.6 presents the prediction performance (in terms of
NRMSE) of the first ANN model on the testing dataset (Fig. 3.6). The median
2
R and median NRMSE of predictions on the testing dataset are 0.8549 and
0.1218, respectively. This testing performance is remarkable given the hostile
subsurface borehole conditions when acquiring the logs, which result in low
signal-to-noise ratio, and the limited size of the dataset available to build the
model, which gives rise to overfitting and poor generalization. Fig. 3.B2 shows
the histograms of NRMSE for training and testing datasets without the imple-
mentation of the five categorical Flags (1–5) as additional features. Comparison
of Fig. 3.B2 with Fig. 3.6 highlights the necessity of Flags as categorical fea-
tures to achieve good generalization performance.
Notably, Fig. 3.C1 lists all the features in terms of their importance to the
data-driven task of T 2 synthesis. Feature importance was performed to find
the most important features out of the 27 features, which include 10 conven-
tional logs, 12 inversion-derived logs, and 5 categorical features. Importance
of a feature for a machine-learning task depends on the statistical properties
of the feature and on the relationship of the feature with other features, targets,
and the machine-learning algorithm used to develop the data-driven model.
Feature importance indicates the significance of a feature for developing a
robust data-driven model. Feature importance helps us understand the inherent
decision making process of a data-driven model and helps in evaluating the con-
sistency of a data-driven model by making the model easy to interpret.

3.4 Training the second ANN model
The second ANN model involves a two-step training process: (1) parameteriz-
ing the T 2 distribution by fitting a bimodal Gaussian distribution and (2) training
the ANN model to predict the six parameters governing the bimodal Gaussian
distribution fit to the T 2 distribution. By following the two-step training process,
a trained ANN model can generate the six parameters of the bimodal Gaussian
distribution. Prediction performance of the second model is affected by the
errors in fitting the T 2 distribution with a bimodal Gaussian distribution (listed
in Table 3.1). Fig. 3.8 presents the prediction performance of the second ANN
model for 25 randomly selected depths from the training dataset.
2
The median R and median NRMSE of predictions of the second ANN model
on the training dataset are 0.7634 and 0.1571, respectively, as compared with
0.8574 and 0.1201, respectively, for the first ANN model. Consequently, the pre-
diction performance on the training dataset (also referred as the memorization
performance) of the first ANN model is superior to that of the second model,
but the computational time of the first ANN model is 30% more than that of
the second model. Histograms of NRMSE of predictions for the 354 depths