Insights using machine learning in predicting faults and horizons: a case study onshore Texas
Dan Ferdinand Fernandez, Mustafa Karer, Richard Hearn, Ryan King, Sunil Manikani, Gavin Menzel-Jones
Digital and Integration, Schlumberger
Summary
The application of machine learning (ML) technology to
seismic interpretation has greatly improved both workflow
efficiencies and product quality. Advancements in speed,
precision, and accuracy leads to valuable insights in the
seismic data and ensure a better understanding of our
subsurface geology and drilling targets. In this paper, we
apply machine learning technology in predicting faults and
horizons in a structurally and geologically complex onshore
Texas dataset. By employing ML technology through
convolutional neural networks (CNNs) trained on real data
we predict multiple layers of faults from small-scale to
regional displacements within the study area. In addition,
using separate CNNs with sparse local labels, we predict
and extract two high-resolution stratigraphic horizons of
the middle Frio and top Wilcox events. These horizons
consistently track on the proper signal amplitude and are
extracted in a fraction of time compared to manual effort.
With the assistance of ML, we can automate the fault and
horizon predictions which dramatically reduce the picking
time to enable interpreters to focus on local complex areas,
assist in generating more accurate horizons, accelerate the
process toward exploration, understand the
reservoir compartmentalization and provide valuable
information to de-risk drilling and improve well placement
decision making.
Introduction
Seismic interpretation is a crucial activity in energy
exploration and production where the seismic interpreter
constructs three-dimensional representations of the
subsurface of seismic surveys. To accurately model the
subsurface, a skilled interpreter depends heavily on the
understanding of their data and their expertise in mapping
out and identifying the structural and stratigraphic elements
related to well information. Only when the important
horizons and faults are completely delineated can
modelling commence for the process of data evaluation,
exploration, risk assessment, and well planning to ensue.
The generation of precise structural and stratigraphic
elements is a laborious and time-consuming endeavor.
Typically, it extends between multiple days to weeks to
complete the extraction of a single horizon. Similar
timelines are expected when extracting swarms of faults
within the same survey.
For horizon interpretation, an interpreter embarks on a
coarse grid, building the framework of the stratigraphic
horizon extent within the seismic survey. Standard grid
interpretation requires providing meticulous control points
on both inline and crossline. Interpreters ensure no mis-ties
in both directions. When working with a team, merge
points must be mis-tie free along the polygon border and
the horizon should consistently follow the same correlated
event across structural discordances, slope changes, poor
seismic image, and natural geologic disruptions. Once the
team has completed the process on a coarse grid, the
process is repeated to tighter grids. Finally, when the
tightest grid is completed, surfacing follows, and line by
line evaluation completes the manual intensive
interpretation task.
Similarly, for fault delineation and extraction, various
papers discuss the use of semblance (Marfurt et al., 1998),
variance (van Bemmel and Pepper, 2000), curvature
(Roberts, 2001), edge detection and enhancement
(Pederson et al, 2002), and fault likelihood (Hale, 2013).
Others (Randen et al., 2001; Bounaim et al., 2013; Hale,
2013; Wu and Hale 2016; Etchebes et al., 2019) employed
advanced structural seismic attributes with better
automation for fault extraction that provided some
efficiency gains. However, widespread adoption and
implementation suffered due to the required proficiency in
tuning parameters for optimum results.
In this study, we demonstrate that by applying machine
learning technology in predicting faults and horizons in a
structurally complex onshore Texas dataset, we simplify
the seismic interpretation process. With the assistance of
ML and minimal upfront labels, the process for experts to
commence on data evaluation and gain valuable insights for
drilling and exploration is significantly improved and
expedited.
Details of study area
Seismic Dataset
This study consists of a land survey acquired using
vibroseis and dynamite covering about ~1280 square kms.
It consists of a 3D pre-stack depth migrated (PSDM)
volume located within northern onshore Gulf Coast basin in
South Texas. The main producing hydrocarbon formations
within the 3D seismic survey are the Frio Formation, and
the Anahuac Formation. The Frio Formation consists of
sand-rich fluvio-deltaic systems, while the Anahuac
Formation are extensive transgressive marine shale
overlying the Frio Formation. The Frio Formation extend
as deltaic and slope sandstones in Louisiana and Texas and
carbonate rocks in the eastern Gulf of Mexico (Swanson et
10.1190/image2022-3749510.1
Page 1322
Second International Meeting for Applied Geoscience & Energy
© 2022 Society of Exploration Geophysicists and the American Association of Petroleum Geologists
Downloaded 09/12/22 to 192.23.26.186. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms
DOI:10.1190/image2022-3749510.1