Page 278 - Machine Learning for Subsurface Characterization

Shallow and deep machine learning models Chapter 8 239

0.75 to 0.8. However, the training time of the deep learning models is 2–3orders
of magnitude more as compared with the shallow models. Among the four model,
LSTM model does not perform as good as the other three models, probably,
because the input “easy-to-acquire” logs are inherently not sequences.
Two types of the input logs (feature set) are used for training the shallow-
and deep learning models. The first type of input logs is the raw form of the
“easy-to-acquire” logs, while the second type of input logs is the inversion-
derived fluid saturation and mineral composition logs obtained by inverting
the raw form of the “easy-to-acquire” logs. Best performing shallow-learning
models, SVR and ANN, perform much better on the inverted logs; however,
deep learning models exhibit relatively performance when processing
inversion-derived logs and raw logs. LSTM performance drops significantly
when processing the raw logs. In the raw form, the logs have strong
correlations/co-linearity because the raw logs are influenced by similar
physical properties, fluid saturations, and mineral compositions of the
formation. When data inversion is performed on the raw logs to obtain the
mineral composition and fluid saturation logs, the inversion-derived logs are
less correlated and better suited for the robust training of the deep learning
models and nonlinear shallow-learning models, like SVR and ANN.

6 Discussions

In the oil and gas industry, several well logs are acquired in the borehole
thousands of feet below the ground for purposes of accurate reservoir
characterization. We applied shallow- and deep learning models to synthesize
the NMR T2 distributions from inversion-derived and raw well logs.
Permeability, residual saturations, and pore size distribution can be estimated
from the synthesized NMR T2 distribution. These petrophysical estimations
are critical for reservoir characterization. This study demonstrates the potential
of shallow- and deep learning models to synthesize NMR T2 distribution.
Limitations of this study are as follows:
(1) Dataset available for this study was limited in size;
(2) Due to limited data size, we could not perform a robust cross validation of
deep learning models to prevent overfitting;
(3) NMR T2 distributions in the available dataset were primarily single-peak
spectra, and there were limited depths with two-peak spectra;
consequently, our models tend to have poor performance in depths with
two-peak NMR spectra; and
(4) The log-synthesis performance in different wells could not be assessed for
purposes of blind testing due to data unavailability.
NMR logs, specifically NMR T2 distributions, are hard to acquire due to
financial and operational constrains. Under such data constraints, deep
learning models can be used to reliably synthesize the NMR T2 distributions
to facilitate improved reservoir characterization.