Page 247 - Machine Learning for Subsurface Characterization

212 Machine learning for subsurface characterization

decode a latent vector into an NMR T2 distribution. By sampling from the latent
space, we can visualize VAEc’s understanding (also referred as abstraction) of a
typical NMR T2 distribution. In Fig. 7.13, the learned T2 distribution changes
gradually in the 3D latent space with slight successive variations in the number
of peaks, smoothness of the distribution, the positions of the peaks, and heights
and widths of the peaks. The VAEc learning ensures that the gradual changing
characteristic of the learned NMR T2 (as plotted in Fig. 7.13) is in accordance
with the characteristic of NMR T2 in the formation. NMR T2 in adjacent
formations should be similar to each other with gradual changing
characteristics. After the first stage of training the VAEc-NN that focuses on
the training VAEc, the second stage trains the four-layered neural network
that learns to relate the learned features in NMR data to the 10 inversion-
derived logs.
The trained VAEc-NN is tested on 100 depths of the testing dataset. The
testing performance of the VAEc-NN is shown in Fig. 7.14. The average R 2
of NMR T2 synthesis for the 100 testing depths is 0.75. In Fig. 7.14, the
randomly selected NMR T2 testing data show a variety of shapes. About
70% of 100 samples are single-peak NMR T2, and 30% are double-peak.
For single-peak NMR T2, the predicted NMR T2 and true NMR T2 almost
overlay with each other. For double-peak NMR T2, the accuracy is lower
than single-peak data. Although the VAEc-NN model can predict the peak
position of double-peak NMR T2 with high accuracy, the model has low
prediction accuracy for the height and shape of double peaks.
2
Histogram of R of NMR synthesis for the 100 testing depths is shown in
2
Fig. 7.15. R achieved for 60 out of the 100 depths is above 0.7. In the 300-ft
thick shale formation studied in this chapter, close to one-third of the depths
have two-peaked T2 distribution. Variation in the shape of T2 distribution
indicates changes in formation characteristics and changes in the relationships
between T2 and formation mineral content and fluid saturation. In the absence
of large volume of suitable training data, the deep neural networks cannot
accurately determine the statistical relationships between T2 distributions and
the inversion-derived logs.

8 Application of the LSTM network model
LSTM network is trained and tested with the same dataset as the one used for the
VAEc-NN model, described in the previous section. The average accuracy of
2
the LSTM-based synthesis of NMR T2 in terms of R is 0.78. LSTM-based
synthesis of T2 distributions are shown in Fig. 7.16. Similar to other models,
LSTM better synthesizes the single-peaked T2 distribution as compared with
2
the double-peaked T2. Histogram of R in Fig. 7.17 indicates that LSTM
2
synthesizes at R > 0.9 for 33 out of the 100 depths, which higher than the
performance of VAEc-NN model (Fig. 7.15). LSTM model performance for
single-peaked T2 is better than that of the VAEc-NN model, whereas an