Page 271 - Machine Learning for Subsurface Characterization

234 Machine learning for subsurface characterization

FIG. 8.7 Two-step training process schematic of the VAEc assisted neural network.

sequences. LSTM model treats the NMR T2 distribution as a sequential data as
a function of T2 time. The encoder and decoder network in this model are
different from those in VAE. The encoder here encodes the input logs into a
context vector. The encoded vector is repeated 64 times and recurrently
decoded to 64-dimensional NMR T2 distribution. The training of the
implemented encoder-decoder LSTM model is a one-step process (Fig. 8.8).
They are capable of learning the complex dynamics within the ordering of
input/output sequences.

4.5 Comparisons of the accuracy and computational time
of the deep learning models
Table 8.3 shows the overall coefficient of determination, R2, that quantifies the
fit between original and synthesized NMR T2 distributions for the entire 300 ft
of the shale formation when using four deep learning models. Table 8.4 lists the
computation times required to train the models. R2 and computational time are
evaluated for both inverted logs and raw logs. Median R2 for NMR T2 synthesis
using the deep learning models is relatively similar. Deep models perform
slightly better when processing the inversion-derived logs. LSTM model
has the lowest performance and one-order higher computational time. The
inversion-derived logs act as features extracted from well logs.