Page 270 - Machine Learning for Subsurface Characterization

Shallow and deep machine learning models Chapter 8 233

FIG. 8.6 Two-step training process schematic of the GAN assisted neural network.

4.3 Variational autoencoder with convolutional layer assisted
neural network
VAEc-NN model incorporates convolutional layers within VAE architecture
for the first step of training. Convolutional layers extract spatial/topological
features in the NMR T2. The details of the model are shown in Fig. 8.7.

4.4 Encoder-decoder long short-term memory network
The encoder-decoder long short-term memory (LSTM) network performs
sequence-to-sequence (many-to-many) mapping similar to language
translation. In this study, the sequence of “easy-to-acquire” logs at a specific
depth is processed by the LSTM network to generate the sequence of NMR
T2 distribution at that depth. The encoder LSTM learns to process the
sequence of “easy-to-acquire” logs to compute an internal state
representative of the variations in the input sequence. The decoder network
learns to process the encoded internal state to obtain the sequence of 64 T2
amplitudes, constituting the NMR T2 distribution. The encoder and decoder
networks are designed by implementing one or more LSTM layers. As the
LSTM recurrently processes the sequential information, it generates an
internal state that summarizes previous information in the sequence. LSTM
model is good at solving the long-term dependency in the input and output