Page 83 - Machine Learning for Subsurface Characterization

Shallow neural networks and classification methods Chapter 3 69

is sensitive to clay volume that determines the clay-bound water and capillary-
bound water, which can be sensed by the NMR T 2 distribution. Density log is
sensitive to mineral densities and fluid densities in the formation. Mineral den-
sity is controlled by mineralogy that controls the lithology and pore size distri-
bution of the formation. Mineralogy can lead to secondary effects on the NMR
T 2 distribution because a certain type of lithology gives rise to a certain type of
pore size distribution; for example, sandstone formation with predominantly
quartz mineralogy will have a distinct pore size distribution compared with car-
bonate formation with predominantly calcite mineralogy, which will be distinct
from shale formation with predominantly clay mineralogy. NMR T 2 distribution
is also affected by surface relaxation, which depends on the mineralogy of the
formation. Finally, both the NMR log and the conventional logs are influenced
by the vertical distribution of fluid saturations along the formation depth, which
is controlled by capillary pressure as governed by the pore size distribution.
In summary, several physical properties that govern NMR T 2 distribution
also influence the conventional logs, namely, neutron, density, resistivity,
sonic, and GR logs. The petrophysical basis for choosing these conventional
logs for predicting the NMR T 2 distribution is that each of these conventional
logs are sensitive to pore size distribution, fluid saturation, mineralogy, and
capillary pressure. By using artificial neural networks to process the log data,
we aim at identifying and extracting the complex and latent statistical relation-
ships that cannot be quantified using any other mechanistic models. The goal
of this study is not to develop a precise mechanistic model to describe the phys-
ical relationships between the conventional logs and NMR T 2 distribution.
Rather, we aim to train neural network models to learn the hidden relationships
between the conventional logs and NMR T 2 distribution.

2.3 Data preparation and statistical information
Data preparation and data preprocessing are essential steps prior to the data-
driven model development (Fig. 3.1). The first ANN model predicts T 2 distribu-
tion discretized into 64 T 2 time bins, whereas the second ANN model predicts
the six statistical parameters that parameterize the T 2 distribution as a sum of two
Gaussian distributions. Before implementing the second model, we need to com-
pute the six statistical parameters that define the bimodal distribution.
For developing the data-driven model, we used data from XX670 ft to
XX990 ft of a shale system comprising seven intervals of a shale reservoir
system. Logs acquired in the well include 12 conventional logs and 10
inversion-derived logs (Fig. 3.2). The conventional logs include gamma ray
(GR) log sensitive to volumetric shale concentration, induction resistivity logs
measured at 10-in. (AT10) and 90-in. (AT90) depths of investigation sensitive
to the volumes of connate hydrocarbon and brine, and neutron (NPOR) and den-
sity porosity (DPHZ) logs that are influenced by the formation porosity. Other
conventional logs include photoelectric factor (PEFZ) log indicating the forma-
tion lithology, VCL log measuring the volume of clay, and RHOZ log sensitive