Page 129 - Machine Learning for Subsurface Characterization

106 Machine learning for subsurface characterization

there are multiple targets (outputs), a general practitioner either uses one ANN to
simultaneously predict all the targets or uses one ANN model to individually pre-
dict each target by processing all the features (none of the targets) as inputs to the
ANN model. In both the cases, the relationships/dependencies between the avail-
able targets are not fully used when training the ANN model. In other words, a tra-
ditionalANNmodelisnottrainedtoprocessfewoftheavailabletargetsalongwith
allthefeaturesasinputstotheANNmodelforpredictingtherestofthetargets.The
noveltyofthisstudyistheSNNarchitecturethatutilizestheinherentdependencies
amongthetargetsbyusingfewtargetsastheinputs/featurestosynthesizetherestof
the targets. Consequently, the SNN model learns to synthesize each target by pro-
cessing 15 conventional logs along with few other targets as inputs.

2 Method
2.1 Data preparation
The log-synthesis workflow is trained and tested on logs acquired along a 2200-
ft depth interval in a well intersecting an organic-rich shale formation compris-
ing six different lithologies. In total, 23 logs were measured during different
logging runs; therefore, depth corrections were performed on all the logs prior
to any further processing. The 15 easy-to-acquire conventional logs include
gamma ray (GR), density porosity (DPHZ), neutron porosity (NPOR), bulk den-
sity (RHOZ), volume of clay (VCL), photoelectric factor (PEFZ), delta-T com-
pressional sonic (DTC), delta-T shear sonic (DTS), lithology indicator, and
galvanic resistivity at six depths of investigation (RLA0, RLA1, RLA2,
RLA3, RLA4, and RLA5). The eight dielectric dispersion (DD) logs consist
of four conductivity-dispersion logs (Track 7) and four permittivity-dispersion
logs (Track 8) measured at four distinct frequencies in the range of 10 MHz to
1 GHz. The 15 easy-to-acquire conventional logs are processed by the stacked
neural network model shown in Fig. 4.1 to generate the 8 DD logs. The stacked
neural network uses the 15 conventional logs (Tracks 2–6, Fig. 4.2) as the fea-
tures (inputs/attributes) and the 8 DD logs (Tracks 7 and 8, Fig. 4.2) as the tar-
gets (outputs) to train and test the log-synthesis method. Lithology indicator is a
synthetic integer-valued log generated by assigning an integer value in the range
of 1 to 6, representing the mineralogy/lithology of the intersected formation.

2.2 Methodology for the dielectric dispersion (DD) log synthesis

The stacked neural network (SNN) model developed in this study was trained
and tested on 15 “easy-to-acquire” conventional logs and 8 dielectric dispersion
(DD) logs. These logs were acquired in one well intersecting a 2200-ft depth
interval of an organic-rich shale formation (Fig. 4.2). The eight DD logs consist
of four conductivity-dispersion logs and four permittivity-dispersion logs mea-
sured at four distinct frequencies in the range of 10 MHz to 1 GHz. The stacked
neural network uses the 15 conventional logs as the features (inputs) and