Page 81 - Machine Learning for Subsurface Characterization

Shallow neural networks and classification methods Chapter 3 67

formation using ANNs to process the density, neutron, resistivity, and photoelec-
tric logs. Recently, Mahmoud et al. [3] predicted TOC for Barnett shale by using
ANN to process the resistivity, gamma ray, sonic transit time, and bulk density
logs. The model was then applied to estimate TOC for Devonian shale. ANN has
been used by several researchers to predict NMR-T 2 -derived parameters. Salazar
and Romero [4] predicted NMR porosity and permeability in a carbonate reser-
voir using ANNs that processed gamma ray, resistivity, and neutron logs. Moha-
ghegh et al. [5] synthesized NMR-derived free fluid, irreducible water, and
effective porosity using ANN that processed SP, gamma ray, caliper, and resis-
tivity logs. Later, Elshafei and Hamada [6] predicted permeability using bulk gas
model and ANN model. The predicted permeability agreed with permeability
measurements on core samples. Labani et al. [7] estimated free fluid-filled poros-
ity and permeability using a committee machine with intelligent systems (CMIS)
in the South Pars gas field. CMIS combines the results of fuzzy logic, neuro-
fuzzy, and neural network algorithms for overall estimation of NMR log param-
eters from conventional log data. Recently, in the Asmari formation, Golsanami
et al. [8] predicted porosities of eight T 2 bins and T 2 logarithmic mean (T 2,LM )of
NMR T 2 distribution using intelligent models. Notably, coauthors of this book
were the first to develop ANN-based predictive models for synthesizing the entire
NMR T 2 distribution partitioned into 64 bins.
In relation to the second predictive model implemented in this chapter, Genty
etal. [9] fitted NMR T 2 distributionacquired in a carbonate reservoir withmultiple
Gaussian (or normal) distributions that were parametrized using three parameters
(α, μ, σ) for each distribution. In their case, T 2 distributions required three Gauss-
ian components and nine corresponding parameters for purposes of the fitting.
Genty et al. [9] utilized these fitted parameters to identify genetic pore types in
a carbonate reservoir. Di [10] implemented a method similar to that proposed
by Genty et al. [9] to identify different lithofacies in a tight oil reservoir based
on the parameters estimated by fitting a Gaussian to the T 2 distribution. We
use a similar approach to compute six parameters that describe the NMR T 2 dis-
tribution in the shale formation. These six parameters are used for training and
testing the second ANN-based predictive model for NMR T 2 prediction.

2 Methodology
2.1 Hydrocarbon-bearing shale system
The shale system under investigation contains conventional and unconventional
elements. Conventional intervals consist of middle shale (MS) and conventional
reservoir (CR) layers, whereas the source rock intervals include lower shale (LS)
and upper shale (US). From the top to bottom, the shale system consists the fol-
lowingintervals: uppershale(US),middleshale (MS), lower shale(LS), andfour
conventional reservoirs (CR 1–4).Theseintervalsare distinctlydifferentanddis-
play highly heterogeneous distributions of reservoir properties. Oil and gas pro-
duced in the US and LS got accumulated in the MS and CR intervals. CR 1–4