American Oil and Gas Reporter - January 2021 - 87

15%-29% of the lateral length on the
heel side.
From these observations, it was concluded that critical failures were resulting
from a condition that was not easily observable; a condition composed of several
parameters likely related to wellbore construction, formation geomechanics, and
stimulation treatment design. With this
expensive, complex and multivariate problem negatively impacting project economics, the operator sought to develop a
systemic approach to reduce risks associated with casing deformation. Because
interpreted seismic features had intermittent correlation to completion difficulties,
the team chose a geo-informed, datacentered modeling approach incorporating
machine learning to delineate possible
causes and solutions.
Machine Learning Approach
Machine learning requires two categories of features: predictive data features
and supervising data features. The predictive data features for the models used
in this study came from wellbore surveys,
drill bit geomechanics and seismic data
(Table 2). The geomechanical data was
acquired by a sensor containing an accelerometer and gyro to capture data in
three directions (lateral, axial and tan-

gential) at high frequencies. Once a well
was drilled, the acceleration data was
processed for both isotropic and anisotropic
rock properties. Because these measurements were collected at very high resolution, both the average value of the
measurement and the standard deviation
of the measurement along a stage were
included in the models.
It initially appeared that the most
direct machine learning approach for predicting stages at risk of casing deformation
was to mark each stage as " casing deformation " or " normal " and set this binary
column as the supervising data feature.
However, with were only six recorded

instances of casing deformation, and restricting the dataset to stages with all the
data features of interest left only four
recorded casing deformation instances.
Trying to directly model this would have
led to statistically insignificant results.
Instead, two completion parameters
consistent across all the casing deformation
stages were used as proxies: a final fracture
gradient in excess of 1.0 psi/foot, and a
sand concentration lower than 0.89 ppg.
With these two proxies identified and
enough data points available to give statistically valid results, the machine learning
workflow was separated into two phases:
predictive modeling and probabilistic re-

TABLE 2
Predictive Data Features used in the
Machine Learning project and their sources
Predictive Data Feature
Shmin
Shmin gradient std. deviation
Young's modulus avg.
Young's modulus std. deviation
Layering avg.
TVD
Distance from heel
Instantaneous tortuosity index
Cumulative tortuosity index
Similarity
Most negative curvature

Source
Drill bit geomechanics
Drill bit geomechanics
Drill bit geomechanics
Drill bit geomechanics
Drill bit geomechanics
Wellbore surveys
Wellbore surveys
Wellbore surveys
Wellbore surveys
Seismic model
Seismic model

Continuous
duty mea
ans
full rod lo
oad,
all the tim
me.
SPM®
QEM 30 0 0

Oil & Gas
www.global.w
.
eir

Image provided by Kohl Koppinger of RevEnergy
Copyright © 2020 S.P.M. Flow Control, Inc.
All rights reser ved.

JANUARY 2021 87

http://www.global.weir

American Oil and Gas Reporter - January 2021

Table of Contents for the Digital Edition of American Oil and Gas Reporter - January 2021