American Oil and Gas Reporter - November 2017 - 46

SpecialReport: Oil & Gas Computing
FIGURE 5
Machine Learning Training with Ensemble of Naive Neural Networks

seismic acquisition, capturing their respective structural complexities and noise patterns.
Resolving Shale Lithofacies
Shale plays are highly variable in the spatial distribution of
lithofacies defined by their respective mineralogy, organic content
and brittleness. The shale formation is a component of a broader
interconnected series of overlying and underlying formations
(often carbonates) that define the full resource play and petroleum
system. Facies changes can occur over relatively short distances.
An accurate determinations of facies distributions will invariably
lead to a more realistic measure of reservoir behavior.
Predicting and modeling lithofacies heterogeneities allows
geoscientists and drilling engineers to better understand the depositional processes that produced these heterogeneities and to
optimize well placement and well plans for long laterals, staying
in the sweet zones defined by the most productive lithofacies.
Modeling these lithofacies heterogeneities, however, is problematic as useable well data may be sparsely sampled and seismic
data may lack the resolution to resolve them. Seismic inversion
methods make use of both well data and seismic data to predict
elastic properties, but suffer from the problem of "non-uniqueness"
and overlap in cluster (classification) space. Unsupervised seismic
classification procedures also are limited by the fact that lithology
is not linearly correlated with seismic data and do not carry a
measure of uncertainty with their classifications.
To resolve these ambiguities, machine learning methods are
being introduced that elevate seismic data from an exploration
tool to a field development tool. The method also elevates the
classification scheme from an unsupervised to a supervised ap46 THE AMERICAN OIL & GAS REPORTER

proach, while carrying probabilities for each facies and the
most probable facies.
The strength of the method is the ability of the system to
integrate data of different types (core, wireline and seismic data)
and resolutions (Figure 4). The workflow begins by constructing
rock type models from pairs of well log data where well log
measurements that best describe the facies are selected. In this
process, another machine learning method referred to as multiresolution graphic clustering (MRGC) is used. This method defines
clusters of different resolutions and can differentiate between homogeneous sands and finely laminated sands, for example.
The next step is to generate a probabilistic rock type model
for the full area with the guidance of seismic data. Toward this
end, an ensemble of naïve neural networks, each with a different
learning strategy, are run simultaneously as an associative combination to predict the rock types (Figure 5). Using multiple
neural networks prevents biasing of any one of the neural
network architectures.
To train the neural networks, both seismic data and lithofacies
data interpreted at each well location are introduced. This constitutes the hard training data. The seismic data training set can
carry the full dimensionality of prestack data, if desired. In the
second phase of the training, soft data (seismic data away from
the well bore) are introduced and the neural networks "vote" on
their inclusion into the training set. The introduction of soft
data avoids overlearning with a limited dataset.
The final step is to propagate the neural network properties
into full probabilistic facies models, where the most probable
facies or the probability of individual facies can be analyzed.
Figure 6 shows the most probable facies of the Lower Eagle

American Oil and Gas Reporter - November 2017

Table of Contents for the Digital Edition of American Oil and Gas Reporter - November 2017