Cuttings Analysis Using Machine Learning Reduces Reservoir Uncertainty - CGG
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
JANUARY 2021 The “Better Business” Publication Serving the Exploration / Drilling / Production Industry Cuttings Analysis Using Machine Learning Reduces Reservoir Uncertainty By Ross Taylor, Chi Vinh Ly, Aiguo Bian and Stephen Secrest HOUSTON–Cuttings are a valuable and cost-effective source of subsurface data from both vertical and lateral wells that is commonly overlooked in oil and gas exploration and development drilling. Advancements in analytical technologies have enhanced the data extraction from cuttings to provide key subsurface insights, in lieu of core or sidewall core data. However, there have always been challenges and uncertainties with sample depth allocations and upscaling/downscaling the data to log resolution for petrophysical analysis. This causes uncertainties throughout the subsurface characterization process, affecting results at all scales. A new methodology integrates high-resolution scanning electron microscopy (SEM) analysis with machine learning techniques to allow for semi-autonomous lithotyping of in- dividual cuttings-particles within each cuttings bag (typically representing 10-30 foot intervals) and their direct correlation with measured well logs. This new workflow greatly reduces cuttings uncertainties through the spe- ciation of bulk mineralogy back into the original rock types from which the cuttings originate, helping upscale results to petrophysical models and beyond. The result is an improved understanding of rock characteristics, lithology and geologic heterogeneity along the wellbore to optimize completion design. Moreover, by recalibrating representative cores, the technology provides a cost-effective alternative to core analysis. Reproduced for CGG with permission from The American Oil & Gas Reporter www.aogr.com
Tech Trends 2021
The approach takes advantage of bed boundaries to delineate rock types how is it possible to assess reservoir het-
SEM-based mineralogical analytical based on mineralogy and/or porosity dis- erogeneity in the absence of core and/or
tools, which combine the imaging capa- tribution. FMI (formation micro imager) logs?
bility of an SEM with energy dispersive Quantitative analysis of cuttings, the
Lithotype-Based Analysis
spectrometry (EDS) to provide both tex- closest thing you can get to core, can pro-
tural and elemental data. The value of Figure 2 shows the resolution at vide a solution. The core-to-cuttings
using SEM-EDS over traditional tech- which the reservoir can be appraised, lithotype calibration workflow is shown
niques, such as X-ray diffraction (XRD) and the pitfalls of thin-bedded reser- in Figure 3, where core lithotypes are
and X-ray fluorescence (XRF), is that the voirs. The benefit of characterizing cut- used as seed testing for cuttings litho-
original rock texture of each cuttings par- tings data as lithotype proportions as types classification. A lithotype model is
ticle is retained and collected. Figure 1(A) opposed to mineral volumes is that a di- created by training a machine learning al-
shows an example of an output grey-scale rect comparison can be made with the gorithm to automatically assign litho-
photomicrograph generated by the SEM, lithofacies assigned in the core interval types to cuttings based on similarity to
while Figure 1(B) shows the equivalent to delineate net reservoir. the training set. The benefit of this tech-
mineral map with each color representing The CT density curve data, which ac- nique is that consistent lithotyping can
a different mineral species detected by quires data at 0.01-foot increments aver- then be conducted on adjacent wells,
EDS analysis of the same area. aging over 0.1-foot intervals, shows good which do not have conventional core, and
The calibration of lithotypes from cut- sensitivity to the porous sandstone beds. where formation characterization other-
tings against lithotypes from core data, Core gamma ray (GR) shows slightly wise would have been purely reliant on
using both the mineralogical and textural poorer resolution (0.1-foot increments petrophysical log interpretation.
data, provides the crucial calibration and averaged over 0.5-foot intervals) with
Machine Learning
upscaling data for this workflow. Core some of the cleaner sandstone beds delin-
description allows for qualitative assign- eated, but relatively poor definition be- The workflow was applied to train a
ment of lithofacies over the cored inter- tween dense siltstone beds (similar GR machine learning algorithm on a calibra-
val, harnessing quantitative continuous response) and mudstones. The log GR tion well and its subsequent application
data, such as core gamma ray, XRF shows almost no variation throughout the to a test well from an asset development
(hand-held) and CT density as well as illustrated short stratigraphic intervals. area in the Midland Basin in West Texas.
discrete data from plugs, such as petrog- Essentially the productive and nonpro- The machine learning lithotyping algo-
raphy/mineralogy and core analysis. ductive beds combine to form a solitary rithm was trained using the lithotypes de-
Continuous data allows for definition of response through the 10-foot interval. So, fined in the core and cuttings samples as
FIGURE 1
Backscattered Electron Image of Cuttings Samples from SEM-EDS (A) and Map of Detected Minerals (B)
(A) (B)
Quartz
K-Feldspar
Albite
Illite/Muscovite
Montmorillonite
Chlorite
Calcite
Dolomite
Ankerite
Pyrite
Apatite
Anhydrite
Rutile
Barite
Void
Others
FIGURE 2
Lithotype Analysis of Thin-Bedded Reservoirs
Log Scale
10 ft Stratigraphy Comparison y
3 ft Stratigraphy
De
nsit What does an average quartz content indicate in a 10 ft interval with
GR re GR CT
Log Co Co
re sand, calcareous sand, mudstone and limestone?
Appraisal of reservoirs requires Unravelling this puzzle requires delineation of which lithotypes are
the analysis of various contributing to this quartz content.
datasets
captured at a range of scales. Quantitative mineralogical analysis of individual particles within cut-
tings is the only method to hold the logs to account.
Conventional log suites struggle CUTTINGS CUTTINGS
to resolve thin beds (
Tech Trends 2021
ponents adjusted to optimize the classifi- Figure 4C is a confusion matrix gener- inhibiting cementation, are intercalated
cation model’s performance. A key re- ated from these datasets, showing a ma- with dense siltstones and mudstones,
quirement for training a machine learning jority of true positive outcomes. which are baffles to hydrocarbon flow.
algorithm is a large volume of data. This The reservoir zone is heterogeneous with
Case Study Well
is one of the major advantages of using beds typically less than one foot in thick-
cuttings, as each cuttings particle repre- In this application, a vertical well lo- ness and composed of very fine-grained
sents a potential training or testing data cated in Midland County, Tx., was se- sands, stacked into amalgamated sand
point. With each sample containing hun- lected because of its intersection with the bodies, which conventional log suites
dreds–if not thousands–of cuttings, train- Dean/Spraberry transition, an interval of cannot distinguish.
ing data is no longer a limiting factor in established hydrocarbon production The Lower Spraberry formation is
the analysis. across the Midland Basin and subject to composed of mudstones and carbonates,
For this case study, the scaled mineral current lateral well drilling campaigns. deposited as gravity flows on a carbon-
dataset and textural data were divided Well success is driven by the ability to ate-influenced slope. Bed thickness is
into a training versus testing dataset at a create an effective fracture network in hugely variable, with the carbonate beds
ratio of 1-to-2. The outcome from the order to maximize oil recovery. pinching and swelling within a shale-
training dataset was compared with the The Dean formation is composed of a dominated interval. Accurately determin-
testing dataset, and it was found that the series of clastic, submarine fan lobes ing the composition and distribution of
precision of the match peaked at ~81%. within the central Midland Basin. The brittle (carbonate) beds has important im-
Figures 4A and 4B show a comparison formation is a series of heterogeneous plications for landing zone placement
between the final machine-driven lithofa- pay zones. The productive beds of very and in estimating fracture height growth,
cies prediction and the “truthed” facies. fine-grained sandstones, with clay coats which ultimately controls hydrocarbon
production.
FIGURE 4 For the purposes of the study, 13 main
Comparison of True Facies (A) versus Predicted Facies (B) and lithofacies were defined using mineral-
Confusion Matrix (C) Between the Datasets ogy criteria belonging to carbonate-,
clay- and siliciclastic-dominated groups.
True Predicted Petrographic control points (mineralogy)
Facies Facies
were used as inputs for descriptive statis-
tics to determine the range of key mineral
values and total organic carbon content
Confusion Matrix within lithofacies.
Predicted Facies
PLT ALT DL AL LW LPG MSM CM SM C AS CS CSAND ASAND SAND
Reservoir Appraisal
PLT
ALT As shown in Figure 5, the Dean con-
DL tains porous sandstones at the base of the
AL interval, which decrease in frequency and
LW sum thickness toward the contact with the
LPG overlying Spraberry. These very fine-
True Facies
MSM
grained sandstones have porosity ranges
CM
of 8%-10%, where calcite and clay con-
SM
tent are both independently less than 10%
C
(i.e., quartz greater than 90%). An up-
AS
ward increase in calcite- or dolomite-ce-
CS
CSAND
mented sandstones and siltstones
adversely impacts reservoir quality.
ASAND
SAND
Interval 1 (highlighted in Figure 5) in
the Dean formation is composed of a di-
(C) verse range of sandstones, siltstones and
(A) (B) mudstones with minor carbonates
(dolomitized siltstones). The reservoir fa-
cies in this instance is the clean sand- tings lithotypes data in the test well nonrepresentative and rely on log data
stones, which form 24.0% of the cored shows a greater proportion of sandstone alone.
lithofacies. Clean sandstone forms 20.3% reservoir facies, and a greater proportion This new lithotype dataset, therefore,
of the cuttings proportion over this inter- of brittle beds in the overlying Spraberry allows wells to be ranked and risked on
val, which is a very close match to the (owing to the greater frequency and/or the basis of cuttings. It can be supported
core proportion. Average quartz volume magnitude of carbonate gravity flows in by log data, but can be used independ-
in this interval (48.1%) is distributed in this interval). This is in keeping with the ently as a dataset to appraise reservoir
various proportions across all clastic average cleaner GR observed in both the quality distribution in terms of the pro-
lithologies. A back calculation from an Dean and the Spraberry formations in portion of reservoir facies or seal facies
average volume theoretically could be the test well compared with the calibra- and provide ground-truthing to log-based
possible, but would be reliant on assump- tion well. reservoir/pay cut off values (e.g., GR val-
tions and circular logic if restricted to log However, it must be noted that in the ues of less than 50 API and neutron
data (i.e., reliant on VClay logs) as op- test well, the intervals with the greatest porosity values of greater than 0.15).
posed to primary data collection from proportion of reservoir facies (8,990-9,010
rock material. feet), ranging from 30%-40%, all have Landing Zone Appraisal
An adjacent uncored well was in- quartz volumes from 50%-60%. A hasty Interval 2 (highlighted in Figure 5) in
cluded to test the accuracy of the litho- interpretation of average mineral data the Spraberry interval is a mix of dolo-
typing parameters assigned from the could lead to a decision to rule out pay stone, limestone and mudstone in the
cored well to an uncored well. The cut- zones, or to distrust the cuttings data as cored interval. Carbonate lithofacies form
FIGURE 5
Proportion of Lithotypes per 10-foot Interval versus Normalized Sub-Area Percent of Minerals
1. SANDSTONE - SILTSTONE-MUDSTONE 2. CARBONATE-SHALE STRAT COLUMN
WHITE LIGHT
WHITE LIGHT
LITHOLOGY
LITHOLOGY
Core Depth
Core Depth
UV Images
UV Images
CORE CORE
GR RHOB g/cc XRD GR RHOB g/cc XRD
CARBONATE-SHALE
SPRABERRY
FORMATION
CORE LITHOFACIES
Dolomitic limestone, Grainstone
Limestone, argillaceous, Wackestone
Limestone, Packstone, grain dominated
Limestone, Packstone, mud dominated
Mudstone, calcareous
Mudstone, siliciclastic
Siltstone, siliciclastic
SANDSTONE-SHALE
DEANFORMATION
MINERAL VOL%
Sum (Quartz%)
Sum (K-Feldspar%)
Sum (Albite%)
Sum (Illite/Muscovite%)
Sum (Montmorillonite%)
Sum (Chlorite%)
Sum (Amiphibole%)
Sum (Calcaerous Clay%)
Sum (Calcite%)
5.5%
10.5% Sum (Dolomite%)
Sum (Ankerite%)
24.7%
MINERAL %
Sum (Pyrite%)
MINERAL %
6.7% Sum (Apatite%)
IMAGE
IMAGE
48.1% 32.7% Sum (Anhydrite%)
Sum (Rutile%)
13.5% Sum (Barite%)
13.5%
12.7% 13.3% RoqLith LITHOTYPE
Sum (Pyrite Lithotype)
Sum (Apatite Lithotype)
Sum (Dolomitic Limestone)
11.0% 7.3% Sum (Argillaceous Limestone)
5.1% 8.2% Sum (Limestone - wackepackstone)
20.3% 23.0% 18.0% 15.0% Sum (Limestone Packstone-Grainstone)
LITHOTYPE %
LITHOTYPE %
6.3% 9.5% Sum (Mixed - Siltstone)
LITHOFACIES
Sum (Mixed - Claystone)
LITHOFACIES
13.7% 8.0% Sum (Mixed - Wackestone)
11.0% Sum (Mudstone - Calcareous)
14.7% 5.4% Sum (Mudstone - Sandy)
8.0%
40.0% 29.0% Sum (Claystone)
5.1% Sum (Siltstone - Argillaceous)
24.0% Sum (Siltstone - Calcereous)
7.9% 21.0%
22.4% 15.0% 5.7% Sum (Sandstone - Calcereous)
Sum (Sandstone - Argillaceous)
Sum (Sandstone - Clean)
Tech Trends 2021 31% of the 10-foot core interval. Carbon- lineating lithotypes from cuttings inter- ric calculation by over- or underestimat- ate-dominated lithotypes form 24.9% of vals. The end-user has the capacity to ing reserves within a field. A good core- the ditch-cutting interval, which again is evaluate the distribution of minerals to-cuttings calibration utilizing this a very close match with the core. It is im- within each lithotype, allowing for a dy- lithotyping approach enables an operator portant to note that in both of these ex- namic evaluation of reservoir parameters to run multiple wells within a field and ample intervals, the dominant mineral (e.g., clay/carbonate volumes or image build their reservoir maps on the basis of does not correspond to a comparable pro- porosity) when new rock types are en- a consistent and comprehensive rock-cal- portion of that end member. countered in a well. This significantly in- ibrated petrophysical model. For instance, Interval 2 has an average creases the accuracy of the reservoir Ultimately, the ready availability of carbonate volume of 50.1%, but most of description in comparison with alternative cuttings from all wells, combined with the the carbonate is portioned into “lime- techniques, which are confined to simply application of this workflow, allows for stone” lithotypes that form only 24.9% of (homogenized) bulk mineralogy values. consistent objective lithological descrip- the interval. These observations can have Second, incorporating lithotype distri- tion of subsurface formations. The lower important implications for the geome- bution in zones where wireline log data cost of cuttings analysis, compared with chanical behavior of the formation in underappreciates the formation hetero- log collection, also allows for the cost-ef- terms of fracture propagation and how geneity (as a result of thin bed effects, for fective generation of a detailed lithologi- the thickness, composition and lateral example) can lead to a more accurate ap- cal database from cuttings. This database continuity of these brittle carbonate beds praisal of net reservoir within any given then becomes an effective tool in mapping vary along lateral well stages. well or interval. Reservoir models built and interpreting the spatial reservoir qual- As illustrated in the Midland Basin ap- from petrophysical logs alone could lead ity of a target formation and allows for plication, there are strong benefits of de- to upscaling of errors, affecting volumet- more informed landing zone decisions.❒
ROSS
TAYLOR
Ross Taylor is a sedimentologist in CGG’s Reservoir Americas group, which offers RoqLith™ ad-
vanced cuttings analysis as part of its RoqScan™ quantitative and automated rock properties and
mineralogical analytical solutions. He worked for Corex for two years before joining CGG five years
ago, leading international exploration projects. Since joining the Houston team, Taylor has gathered
extensive experience with core and wireline-based data across the Lower-48, including in the Hay-
nesville, Wolfcamp, Dean, Barnett, Frontier and Niobrara. He holds a B.S. and Ph.D. from the Uni-
versity of Aberdeen.
CHI
VINH LY
Chi Vinh Ly is a regional technical manager for the CGG Reservoir Americas Geoscience group,
where he is in charge of the RoqScan analytical group. He has been working in the minerals, mining
and energy sector since 2004 for CSIRO and ALS Minerals, focusing on analytical techniques, such
as XRD, XRF, ICP-MS and IXP-OES, with a focus on automated mineralogical analysis. He holds
an undergraduate degree in applied physics from Curtin University and a Ph.D. from the Hiroshima
University.
AIGUO
BIAN
Aiguo Bian is a geoscience technician providing geotechnical support to CGG’s geology team
and the RoqScan group. He joined CGG in 2014 as a RoqScan technician performing day-to-day
analysis on instrumentation in the laboratory and in the field. Bian began his career as a staff geol-
ogist for Uranium Energy Company, advising the production and development team for potential
new resources. He completed his undergraduate studies in geoscience and a Ph.D. in science at
Texas A&M University.
STEPHEN
SECREST
Stephen Secrest is the geoscience manager for PetroLegacy Energy II in Austin, Tx. He primarily
oversees all well planning/steering, business development and exploration initiatives while also man-
aging a regional multi-operator geochemistry study. Before joining PetroLegacy, he had served for
seven years as a district geologist at Samson Resources. Secrest holds a B.S. and an M.S. in
geology/earth science from Baylor University.
You can also read
