Conference PaperPDF Available

From a Core to an Oil Field: Machine Learning Applied to Sedimentological Models

Abstract

Accurate sedimentological models in hydrocarbon reservoirs are fundamental to reduce exploration and production risks. The most accurate sedimentological models are built from detailed description and interpretation (facies analysis) performed on cores. Most mature reservoirs contain a large number of producing wells but the core database contains limited information. As a consequence, precise and accurate sedimentological studies performed on cores are often difficult to apply to the entire field model due to scale problems between well logs and seismic analysis. One possible solution is the definition and calibration of electrofacies from core studies. However, the interpretation of electrofacies maps are often complex and poorly accurate. Using state of the art machine learning techniques it is possible to apply detailed sedimentological studies to a large well log database inside a desired stratigraphic intervals. Algorithms are trained to recognize sedimentary facies using as input a set of lithologic well logs in the available core intervals. Results consist on a new set of logs that contain the inferred facies distribution for the rest of the analyzed reservoir wells. The new predictions allow a fast and accurate mapping of facies, facies associations, and depositional elements in the study area. If facies and depositional elements are populated with results of conventional analysis, it is also possible to generate detailed maps showing changes in porosity and permeability along the entire field. This information contributes to a substantial risk reduction in predicting reservoir quality in undrilled areas. The procedure was successfully applied to different oil fields in Argentina, Mexico and Russia.
From Cores to an Entire Field:
Improving Reservoir Sedimentological
Models Using Machine Learning
Javier Iparraguirre1,2, Carlos Zavala1,3, Mariano Arcuri1,3
(1) GCS Argentina SRL.
(2) Universidad Tecnológica Nacional, Bahía Blanca, Argentina.
(3) Universidad Nacional del Sur, Bahía Blanca, Argentina.
Introduction
Accurate sedimentological models in reservoirs are fundamental to
reduce exploration and production risks.
The most accurate sedimentological models are built from detailed
description and interpretation (facies analysis) performed on cores.
Most mature reservoirs contain a large number of producing wells but
a limited core database.
As a consequence, precise and accurate sedimentological studies
performed on cores are often difficult to apply to the entire field model
due to the different scales.
Electrofacies
Electrofacies analysis assume that
rock-types, facies, and
paleoenvironments can be
identified from well logs.
The classification is based on the
analysis the shape of lithological
logs.
The concept is based on the
assumption that similar rocks
provide the same pattern in log
shapes.
Types of electrofacies analysis
Two types of interpretation:
Based on cores data: log
shapes are contrasted with
cores. Best practice.
Based on logs:
interpretation is based only
on logs. Not recommended.
Cant, 1992
Why machine learning?
Machine learning brings a formal methodology to train a classifier and
infer predictions.
Quantitative metrics help to evaluate the prediction.
Machine learning is rapidly evolving and unprecedented levels of
prediction can be achieved.
Supervised facies prediction: two approaches
Traditional approach: from log to core
1. Identify of log shapes.
2. Associate shapes to rock type using
core data.
3. Manually determine electrofacies
(EF).
4. Train a classifier based on
annotated data.
5. Predict EF for the rest of the wells
(EFP).
6. Create an EF map.
Novel approach: from core to log
1. Describe the cores.
2. Manually determine lithological
facies association (FA)
3. Train a classifier based on log
data.
4. Predict (FA) for the rest of the
wells (FAP).
5. Create FA map.
Workflow
Input files
(.las)
Relevant logs
(features)
Annotations Train a
classifier
Evaluate
training
Input files
(.las)
Relevant
logs
Model
Prediction
Training
Prediction
Practical application examples:
from log to core (EF)
Traditional approach (EF): annotation process
1. Different log shapes are
identified as electrofacies.
The initial classification is
not related to rock type.
2. Electrofacies are
characterized by
sedimentary facies
recognized on cores.
3. A classifier is trained based
on the annotated data.
Normalize
logs
Electrofacies
recognition
and
mapping
Describe EF
in terms of
sedimentary
facies
EF training example
Gamma Ray
(GR) log was
used as
reference
A second log
was created
using a low-
pass filter
A random
forest classifier
was trained.
F-Score = 0.731
EF: prediction example (1 of 2)
EF2EF1 AUTOMATICMANUAL AUTOMATIC
MANUAL
EF: prediction example (2 of 2)
MANUAL
EF1 EF2 EF3 EF4
AUTOMATIC
Practical application examples:
from core to log (AF)
Novel approach (AS): annotation example (1 of 2)
1. Core description
2. Manually annotate the
facies association
3. Using the core description,
it is possible to evaluate
how the annotations relate
to rock properties
4. A classifier is trained based
on the annotated data
Novel approach (AS): annotation example (1 of 2)
There is a clear relation
between facies and
reservoir properties
AF Training Example
Gamma Ray (GR)
log was used as
reference
A second log was
created using a
low-pass filter.
A random forest
classifier was
trained.
F-Score = 0.68
AF prediction (1 of 2)
MANUAL AUTOMATIC
MANUAL
AUTOMATIC
AF prediction (1 of 2)
MANUAL AUTOMATIC
AF 7 AF 6 AF 5 AF 4 AF 3 AF 2 AF 1
Detailed reservoir maps
Machine learning allows to recognize facies associations from well logs
Facies analysis can be propagated to the entire field using the well log
database.
Facies associations are directly related to depositional systems, and
can be conveniently calibrated by net-to-gross, and results of
conventional analysis (porosity and porosity)
The last permitted to construct detailed maps showing different
reservoir properties along the study area
Examples of reservoir maps
FAM Results: Isopach - Facies Association Mode
Case study area 2: 4 km2
FAM Results: Permeability - Porosity
Case study area 2: 4 km2
FAM Results: Recognizing depositional Elements
Case study area 2: 4 km2
Conclusions
Conclusions
The use of machine learning allows a fast and accurate mapping of
facies associations and depositional elements in the study area.
If facies and depositional elements are populated with results of
conventional analysis, it is also possible to generate detailed maps
showing changes in porosity and permeability along the entire field.
The proposal contributes a substantial risk reduction in predicting
reservoir quality in undrilled areas.
The procedure was successfully applied to different oil fields in
Argentina, Mexico and Russia.
Thank you!
Questions?
j.iparraguirre@computer.org
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.