Page 279 - Machine Learning for Subsurface Characterization

240 Machine learning for subsurface characterization

7 Conclusions
Six shallow- and four deep learning models were used to process the easy-to-
acquire well logs for synthesizing the NMR T2 distributions for a 300-ft interval
of a shale formation. Both the raw form of “easy-to-acquire” well logs and the
inversion-derived formation mineral and fluid composition logs (obtained by
processing the “easy-to-acquire” logs) were used for the synthesis of the NMR
T2 distribution logs. Log-synthesis performances of the deep learning models
quantified in terms of R2 score range from 0.75 to 0.8, whereas the
performances of shallow-learning models range from 0.6 to 0.75 in terms of
R2. Two-step training of deep neural networks based on variational
autoencoder, generative adversarial network, and variational autoencoder with
convolutional layers resulted in robust deep learning models that exhibit
physically consistent reconstruction. Deep learning models and nonlinear
shallow-learning models, like support vector regressor and artificial neural
network, perform better NMR T2 synthesis by processing the inversion-
derived formation mineral and fluid composition logs. Inversion-derived logs
can be considered as specially engineered features extracted from the raw
logs; consequently, the inversion-derived logs are less correlated and have
more independent, relevant information that boosts the performance of the
deep learning and nonlinear shallow-learning models.

References
[1] Alzate GA, Arbelaez-Londono A, Naranjo Agudelo AJ, Zabala Romero RD, Rosero
Bolanos MA, Rodriguez Escalante DL, et al. Generating synthetic well logs by artificial
neural networks (ANN) using MISO-ARMAX model in cupiagua field. In: SPE Latin
America and caribbean petroleum engineering conference, Maracaibo, Venezuela, May 21,
2014. SPE: Society of Petroleum Engineers; 2014.
[2] Alloush RM, Elkatatny SM, Mahmoud MA, Moussa TM, Ali AZ, Abdulraheem A. Estimation
of geomechanical failure parameters from well logs using artificial intelligence techniques.
In: SPE Kuwait oil & gas show and conference, Kuwait City, Kuwait, October 15, 2017.
SPE: Society of Petroleum Engineers; 2017.
[3] Li H, Misra S, He J. Neural network modeling of in situ fluid-filled pore size distributions in
subsurface shale reservoirs under data constraints. Neural Comput Appl 2019;1–13.
[4] Akande KO, Olatunji SO, Owolabi TO, Abdulraheem A. Feature selection-based ANN for
improved characterization of carbonate reservoir. In: SPE Saudi Arabia section annual
technical symposium and exhibition, Al-Khobar, Saudi Arabia, June 4, 2015. SPE: Society
of Petroleum Engineers; 2015.
[5] Moghadasi L, Ranaee E, Inzoli F, Guadagnini A. Petrophysical well log analysis through
intelligent methods. In: SPE Bergen one day seminar, Bergen, Norway, April 5, 2017. SPE:
Society of Petroleum Engineers; 2017.
[6] Olayiwola T. Application of artificial neural network to estimate permeability from nuclear
magnetic resonance log. In: SPE annual technical conference and exhibition, San Antonio,
TX, October 9, 2017. SPE: Society of Petroleum Engineers; 2017.
[7] Negara A, Ali S, AlDhamen A, Kesserwan H, Jin G. Unconfined compressive strength
prediction from petrophysical properties and elemental spectroscopy using support-vector