Page 253 - Machine Learning for Subsurface Characterization

216 Machine learning for subsurface characterization

2
FIG. 7.17 Histogram of R for the T2 distributions synthesized for the 100 discrete depths in the
testing dataset when using LSTM network.

processing the measured NMR T2 in the training dataset. In the second stage a
simple neural network is connected with the decoder/generator trained in the
first stage to synthesize the NMR T2 by processing the 10 inversion-derived
logs. On the other hand, the LSTM model treats the 10 inversion-derived
logs as an input sequence and the 64-dimensional NMR T2 as an output
sequence. LSTM performs sequence-to-sequence modeling by representing
the input sequence as an intermediate/context vector using the LSTM
encoder network, which is then decoded by the LSTM decoder network to
sequentially generate the entire NMR T2 distribution, one bin at a time.
LSTM and VAEc-NN models have a similar number of parameters, and
LSTM requires only one training stage; however, LSTM requires one order
of magnitude higher training time due to the complexity in sequentially
encoding and decoding the internal state. The accuracy values of NMR T2
2
synthesis for the VAEc-NN and LSTM models in terms of R are 0.75 and
0.78, respectively, and the training times are 43 and 580 s, respectively. This
study opens up the possibility of applying deep neural networks to enhance
reservoir characterization and improve project economics by characterizing
the fluid-filled pore size distribution of the subsurface geological formations.

References
[1] Han Y, Misra S, Wang H, Toumelin E. Hydrocarbon saturation in a Lower-Paleozoic organic-
rich shale gas formation based on Markov-chain Monte Carlo stochastic inversion of
broadband electromagnetic dispersion logs. Fuel 2019;243:645–58.
[2] Tathed P, Han Y, Misra S. Hydrocarbon saturation in Bakken Petroleum System based on joint
inversion of resistivity and dielectric dispersion logs. Fuel 2018;233:45–55.
[3] Guo C, Liu RC. A borehole imaging method using electromagnetic short pulse in oil-based
mud. IEEE Geosci Remote Sens Lett 2010;7(4):856–60.