American Oil and Gas Reporter - November 2023 - 43

SpecialReport: Comprehensive Computing
Machine Learning-Based Method
Provides Economic Option
To Seismic Inversions
By Alvaro Chaveste,
Rocky Roden
and Tom Smith
HOUSTON-Seismic inversion, an essential
part of reservoir characterization,
has been the accepted methodology for
predicting lithofacies between wells for
decades. Inversion-based reservoir characterization
computes rock properties-
including both compressional (P) and
shear (S) wave impedance-and density
through seismic inversion. The desired
reservoir properties (lithology, porosity
and fluids) are obtained through petrophysical
inversion that relates rock properties
to reservoir properties through rock
physics models.
A new method based on machine
learning shows promise in producing
the same or even enhanced results over
traditional advanced seismic inversion
methods, which can take several months
and require costly fees in software and
geoscience services. Fortunately, the alternative
ML approach to reservoir characterization
does not require data to
match physical models. Instead, it relies
on an ML process that takes any number
of seismic attributes and clusters them
in multidimensional space, eliminating
the dimensionality roadblock. The new
methodology is based on applying selforganized
maps (SOM), a form of ML,
and correlating the results ( " neurons " )
to seismic lithofacies from petrophysical
logs.
The neurons generated by SOM have
some of the characteristics of inverted
data; that is, they map intervals, not interfaces,
and show the effect without
knowledge of the seismic wavelet. Indeed,
when the seismic wavelet is more
broadband, the ML results are more detailed.
Additionally, each sample is classified
as a neuron, which provides interpretable
data below the wavelet's resolution
limit.
This SOM-lithofacies correlation approach
to reservoir characterization can
significantly impact project economics,
enabling lithofacies prediction between
wells in weeks and at a fraction of the
cost of sophisticated seismic inversion
methods. The methodology initially was
developed using synthetic seismic data
and evaluated on data from the Niobrara
formation in the Denver-Julesburg Basin
in Colorado. The new technology now
is being expanded and applied to both
conventional and unconventional geologic
settings.
Conventional Inversion
In its most basic form, seismic inversion
transforms seismic reflection data into
lithology, and ultimately, rock and fluid
properties. The objective of seismic inversion
is to convert reflectivity data (an
interface property) to acoustic impedance
(a layer property). Today's most advanced
inversion approaches attempt to discriminate
specifically between lithology, porosity
and fluid effects. Inversions can be
grouped into pre-stack versus post-stack,
deterministic versus geostatistical, or relative
versus absolute.
The most common inversions today
are deterministic, with pre-stack simultaneous
inversion employed routinely.
Pre-stack simultaneous inversion is a prestack
inversion method that uses multiple
offset or angle stacks along with well
control. This method solves for S impedance,
P impedance and density, which
are particularly key for discriminating
lithology. All models, partial stacks and
wavelets are input into a single inversion
algorithm and solved simultaneously,
compensating for offset-dependent phase,
bandwidth, tuning and normal moveout
stretch effects.
The accuracy of advanced inversions
is directly related to the number of wells
in the area of interest and the quality of
their well logs. Well logs establish the
lithologies, and ultimately, the reservoir
properties that define the models for the
inversion process. These advanced inversion
workflows typically take months
of computer processing and seismic
processor review.
There are important technical challenges
associated with seismic inversion.
There is an issue of nonuniqueness because
the seismic amplitudes (interface property)
are converted to layer properties (impedance
= velocity x density). Defining
lithologies from inversion-derived impedances
can be mathematically difficult.
The inversion process requires the determination
of the wavelet and the low-frequency
trend in the seismic data, which
are not trivial. Inversion also requires
relatively high-quality seismic data because
the process is sensitive to nonlinearities,
uncertainties and assumptions that may
not be valid.
Moreover, most inversions are inappropriate
for quantitative interpretation
for thin beds (below seismic tuning). This
is because the time separation of the seismic
reflections from the top and base of
a reservoir does not change below tuning;
only the amplitude varies. Essentially,
all seismic inversions employed in the
industry today assume all the information
related to lithology and reservoir properties
is incorporated in the amplitude data
NOVEMBER 2023 43
American Oil and Gas Reporter - November 2023

Table of Contents for the Digital Edition of American Oil and Gas Reporter - November 2023
