Sie sind auf Seite 1von 5

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/331984999

Facies Analysis Using Machine Learning Techniques With Logs and Core Data:
an Application to the Johan Castberg Field -Norway

Preprint · January 2018

CITATIONS READS

0 187

1 author:

Maura Serreli
Università degli Studi di Perugia
2 PUBLICATIONS   0 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

MSc Thesis View project

All content following this page was uploaded by Maura Serreli on 25 March 2019.

The user has requested enhancement of the downloaded file.


Facies Analysis Using Machine Learning Techniques With Logs and Core
Data: an Application to the Johan Castberg Field – Norway

Maura Serreli - MSc Thesis 2018, University of Perugia, Italy (mauraser@hotmail.com)

Keywords: Depositional facies, Wireline logs, Cores, Machine Learning, Predictions.

The objective of this study is to evaluate whether it is possible to create facies descriptions in
uncored/unclassified wells from indirect measurements using wireline logs and core data from
other wells. Facies classification is an important factor to understand depositional history of
the rocks and therefore to improve the hydrocarbon reservoir characterization.
An innovative approach has been illustrated in this abstract (derived from my MSc thesis) to
attempt to solve the facies classification problem. In this work, a Machine Learning supervised
algorithm (Random Forest) helped to predict depositional facies in a blind well, using core and
wireline data from nearby wells. This method can be applied in case we desire to classify
quickly facies of many wells after having already classified the data of wells in the same area
and can support the “manual” interpretation of well logs for lithofacies classification. Essential
condition for applying this method is that the targeted well has to be part of the same reservoir
to satisfy similar depositional environments.

Methodology

The workflow was developed using data of four oil/gas wells from the Johan Castberg field
(Norway), drilled between 2011 and 2014. Jurassic reservoirs Stø and Nordmela formations
have been investigated as the only two showed in all wells.
A standard practice when training
supervised algorithms is to separate
some data from the training set to
evaluate the accuracy of the classifier.
In this project, a different approach has
been chosen. Three wells over four have
already assigned depositional facies
classes based on core descriptions,
therefore used for training and
validating models. The fourth, with only
wireline data, was our target for facies
classification.
The workflow consists of four phases
(Fig.1): Fig.1: Typical machine learning workflow used:
preparing a data set, training a classifier, and
• Data preparation and exploration
evaluating the model to be applied for predictions
through statistical analysis of well log (Screenshot from Orange software).
measurements;
• Training of supervised learning model to recognize hidden patterns in the measurements;
• Verifying the accuracy of models through cross-checking with labelled wells and merging
of dataset;
• Application of models on targeted well for Facies predictions.

Each well dataset consists of 12 features: 8 numeric type wireline log measurements (features),
3 categorical target variables (Facies Environments, Subsetting and Energy classes) and the
depth measurements as meta-attribute (Fig. 2). In machine learning terminology, the set of

1
measurements at each depth interval is called INSTANCE (feature vector), each of which is
associated with a class (the facies type or label).

DEPT.M AC.US/F DEN.G/CC


GR.GAPI K.v/v NEU.v/v PEF.B/E TH.PPM U.PPM Facies EnvironmentEnergy
7220_5-1
5-1Subsetting 5-1
1337,97 94,8064 2,2411 78,2512 0,0134 0,1945 2,4241 8,4207 1,8855 Lower shoreface medium Shoreface
1338,12 95,6361 2,2327 76,3702 0,0140 0,1949 2,5487 8,0940 1,7782 Lower shoreface medium Shoreface
1338,27 96,9088 2,2273 75,7772 0,0145 0,1884 2,6344 7,8936 1,7208 Lower shoreface medium Shoreface
1338,43 96,9798 2,2267 78,0331 0,0149 0,1878 2,7427 7,9878 1,6497 Lower shoreface medium Shoreface
1338,58 96,6985 2,2355 77,4588 0,0156 0,1916 2,8190 8,3573 1,4262 Lower shoreface medium Shoreface
1338,73 96,8590 2,2432 74,9622 0,0154 0,1906 2,8570 8,0607 1,4424 Lower shoreface medium Shoreface
1338,88 97,5702 2,2572 72,8496 0,0152 0,1851 2,9183 7,6650 1,3162 Lower shoreface medium Shoreface
Fig.2: Highlighted in red an “instance”, in green a “feature”, in blue a “class or label”

Fig. 3 shows Cross plots between facies types (x-axis) and some of our features (y-axis) to see
if there is any direct correlation between them. It is not clear what direct relationships exist
between the measurements and facies types. This is where machine learning will prove useful.

Fig.3: Crossplots from 7220_5-1 generated with Facies Environment classification vs


some features. Each pane in the Figure shows the relationship between the Facies
Environments (X - axis) and the features (Y- axis). Similar results has been obtained for
7220_7-1 and 7220_8-1 datasets.

The first step was to train algorithms from


each well and test it on the other two, after
removing their labels. Knowing true facies,
predicted and actual classifications were
compared through Confusion matrix (Fig.4)
and cross plots. The second step was to
merge the three datasets to increase the
number of instances per facies and therefore
to improve the accuracy of predictions for
the targeted well. All the models have been
then applied on the unclassified well for
facies predictions, having only wireline Fig.4: Confusion matrix for 7220_5-1 dataset
logs. showing Actual vs Predicted Facies classes

Algorithms has been run with three classifications to understand if the accuracy improve with
reducing the number of classes:
2
Facies Environment (up to 12 facies), Subsetting (up to 5 facies) and Energy (3) classifications.

Results shows an impressive matching between the actual and predicted classes using merged
datasets (Fig.5) and bringing down the number of classes, although some misinterpretation has
been detected using the classification with lowest number of class.

Conclusions

The trained models generally have better recognized classes with the highest number of
instances per facies: it makes the classification problem strongly unbalanced, not easily
solvable in Orange. A solution is to change the weighting to a ‘balance weighting’: more
sophisticated machine learning algorithms would allow a better analysis using balanced
weighting to overcome the wide variation in the number of samples;

Fig.5: Facies Environment classification: actual (first bar on each panel) and predicted
classes using merged datasets (last bar). The well 7220_7-3s is the target.

Misinterpretations have been reduced through:


• Merging the datasets: more instances per facies helped to recognize better the patterns
between data. The more data we have the better the answer becomes: if 10 wells were
available with several thousand data points, the eventual answer would be much better again.
• Reducing the number of classes from 12 of Facies Environments to 5 of Subsettings,
predictions resulted more accurate with a better classification due to an improved pattern
recognition of log data from algorithms.
Although the number of classes has been further reduce to 3, results for Energy classification
didn’t show an improved accuracy on facies classification. The output should be that all wells
display a similar class distribution. What happened? The workflow started with a geological
interpretation and then reclassified this into energy. Any error at any of these steps could have
increase the overall error and hence reduce the precision. It is often thought that by reducing

3
the complexity of a problem you increase the precision. Not true always in statistics as results
from Energy facies predictions showed on targeted well.
There are a large number of wells without core data but all wells have wireline logs. The ideal
sources for facies classification are core samples of rocks extracted from wells, not always
obtained due to high costs associated. Therefore, a method is necessary to improve and speed
up the facies classification.

Acknowledgements

Data and supervision of this MSc thesis project has been kindly provided by Joe Johnston (CGG Chief
Petrophysicist), and my deep gratitude is for him. His constant and immediate support, precious
suggestions and patience for an absolute beginner have been invaluable during this project.
Open source data mining software Orange has been used during all the steps of the workflow.

References

• Al-Mudhafar [2017] - Integrating well log interpretations for lithofacies classification and
permeability modeling through advanced machine learning algorithms.
• Anna Yankina GEO - 3900 Master’s Thesis in Geology – [July 2017] - Petroleum systems of the
Barents shelf: a regional well - based study of the Mesozoic.
• Bjørn Anders Lundschien, Tore Høy & Atle Mørk - Triassic hydrocarbon potential in the Northern
Barents Sea; integrating Svalbard and stratigraphic core data.
• Brendon Hall [2016] - Facies classification using machine learning -The Leading Edge.
• Faleide [2008] - Structure and evolution of the continental margin off Norway and the Barents Sea.
• Glossary of terms: http://robotics.stanford.edu/~ronnyk/glossary.html
• http://www.npd.no
• Jan Inge Faleide, Knut Bjørlykke, and Roy H. Gabrielsen [2010] - Geology of the Norwegian
Continental Shelf, Chapter 22.
• Joe Johnston, CGG, and Aurelien Guichard, Teradata [2015] - Using Big Data Analysis Tools To
Understand Bad Hole Sections on the UK Continental Shelf.
• Mohammed Abdalla Ayoub, Abdollah Esmaeili [2016] - Application of artificial neural networks
technique for estimating permeability from well log data.
• Mohseni [2015] - Application of artificial neural networks for prediction of Sarvak Formation
lithofacies based on well log data, Marun oil field, SW Iran.
• Musawi [ 2018]- Introduction to Machine Learning.
• Paolo Bestagini, Vincenzo Lipari and Stefano Tubaro, Politecnico di Milano, Italy - A Machine
Learning Approach to Facies Classification Using Well Logs.
• Paolo Dell’Aversana [2018] - A Multimedia Cross-disciplinary Machine Learning Framework.
• Paolo Dell’Aversana [2018] - Machine Learning and Data Fusion for Litho-fluid Facies Classification
Using Well Logs).
• Paolo Dell’Aversana [2018] - Multimedia Human-Machine Learning, From Big Data to Deep
Meaning.
• Saadullah Nisar [2015] - Reservoir Quality of Lower-Middle Jurassic Sandstones within the Johan
Castberg Field in the SW Barents Sea - Petrographical and Petrophysical approach.
• Santini [2015] - Machine Learning for Language Technology.
• Tschannen [2017] - Facies classification from well logs using an inception convolutional network.
• Zhaolong Liu [2013] - Petroleum System Analysis in Skrugard Area, SW Barents Sea.

View publication stats

Das könnte Ihnen auch gefallen