There is no leaderboard for this competition. If you want to share screenshots of your current algorithm please send them to the organizers who will publish them on a gallery page.

This contest is developed in collaboration with FORCE.

The aim of the contest is to benchmark which team or company who has currently the best fault mapping algorithm and to provide a status of this evolving technology. The winning teams do not have to publish their code in detail. The submitted attribute cubes and 15% or maximum 2000 patches  of the used training data will be published anonymously. This should also allow vendors to participate in this competition.

Several software vendors offer the competition data with free licenses to their packages for people less familiar with Python. For the well log machine learning data you can use https://www.i2g.cloud/. https://www.dgbes.com/ offers a pre loaded project with full Python development environment that supports any machine learning model. This means you don't have to worry about data IO and data visualization. They also offer free licenses to all their specialized plugins. https://www.cegal.com/ offers you temporary free licenses to their Petrel python plugin for the time of the competition. Please contact support.geo@cegal.com if you would want to test a license.

Sponsors:

All questions and discussion related to the challenge should be submitted to the challenge specific discussion forum (under the "Discussion Tab").

Competition Details

1. Prizes are decided by the FORCE organizers and could include travel stipends or cash prizes. See the FORCE site for additional information.

2. We strongly encourage to only using synthetic models to train the fault classifier. The test dataset is from a different seismic survey with a slightly different frequency /noise content and from a slightly different geography.

3. Several public synthetic models and model generators can be found in the Force dataset. You can use these, but can also use your own. The seismic dataset is from the Ichthys Field and is provided courtesy of Geoscience Australia.

4. Any synthetic model or the code used to generate them does not need to be published but a representative random sample of the training data used, (15% of total) must be made public with the submission.

5. The two winning teams must publish the used neural network structure, augmentation procedures and regularization, post processing used to achieve the result in a descriptive form. The actual code does not need to be published.

6. Non winning teams are encouraged to share their code and training data in the spirit of advancing the open source community.

7. All submitted final fault attribute cubes and 15% sample or maximum 2000 patches of the training must be released and accessible under a cc-by-4.0 license anonymously. No team names will be attached to the attribute cubes and training patches unless expressly stated by the submitting team.

8. Scoring:

   1. Scoring of the fault detection will be undertaken by structural geology experts

   2. On the last day of the competition participants will be sent a small seismic cube to run their respective fault detection algorithm on. The fault attribute cube must be returned to the organizing committee within 15 hours of receiving the blind cube.

   3. The “blind cube” will be of slightly similar in seismic fidelity and geology to the training cube.

9. Winner will be announced 3 weeks after submission of the classified blind cube.

10. The final submitted attribute cube must be a segy volume with the value range from 0 to 1. Where 0 indicates no fault and 1 indicates that there is a fault. The submitters need to specify the display thresholding with the final submission. Else FORCE will choose an appropriate color and display range. For citation please use: Bormann P., Aursand P., Dilib F., Dischington P., Manral S. 2020. 2020 FORCE Machine Learning Contest. https://github.com/bolgebrygg/Force-2020-Machine-Learning-competition

11. Per FORCE's IP rules, Participants retain ownership over their submissions, but are expected to comply with the citation and release licensing detailed above.

12. To register as a team, contact support@xeek.ai and provide the email addresses of each team member. Teams must conform to the submission requirements and all team members must individually register for the challenge.

13. Teams must be limited to 4 participants or fewer.

Dataset Overview

The Ichthys Seismic dataset is provided courtesy of Geoscience Australia and is available under CC BY 4.0 license. It can be downloaded here: Ichthys seismic data or https://nextcloud.drgaff.net/index.php/s/7mXn8mSbjxdZyr8.

It contains a very well-expressed polygonal fault system in the overburden, locally intersected by larger planar faults in the eastern part of the survey and  a deeper faulted section in the Jurassic that is clearly dominated by more diffuse but human mappable fault zones.

The focus of the competition to use a synthetic model trained algorithm that accurately maps the deeper and the shallow faulting and approximates human interpretation.

1. The seismic data is available as standard segy data and can be read with the SEGYIO python library

2. The coordinate system is shown below

3. The Z axis of the seismic data is in milliseconds Two Way Time (TWT)

4. The processing and regional reports of relevance are in the same folder where you find the seismic data

Computational Resources

1. We do not offer not offer any computational resources

2. Google Colab or Amazon AWS can help, as do other cloud compute providers

Synthetic Models Description

1. You can use the FORCE cc-by-4.0 synthetic models as a starter to provide training images for your algorithm. You can also use your own synthetic models.

2. We provide a synthetic model generator and some models from Schlumberger. Click link to access the Schlumberger Models and code in C#. 

3. We also provide a realistic faulted Earth Model prepared by Equinor that approximates real seismic data. It can be found here together with a description of the data and an example notebook to load the cubes in python.

Determination of Winners

1. We will use human interpreters to assess the quality of your machine interpretations.

2. On the 15th October all teams who signed up for the ML on seismic competition will be sent a link to a blind seismic dataset that you have to run your algorithm over. You will submit your fault prediction back to us on 16th October 23.59pm CEST (Berlin Time). 

3. Your submitted data must be a standard segy seismic cube for the entire volume as well as selected in-lines an x-lines in segy format. The amplitude data must range from 0 to 1 and does not need to be binary (0= no fault; 1= certain to be a fault)

4. Your machine interpretations will be assessed by structural geologist as well as interpreters. The interpretation will be ranked from 1-6 and the average of the score will determine the winner. 

5. We understand  that there may be human bias in our assessment but we accept this as one of the axioms of life.

Previous Work Examples

1. The best published work on using synthetics to map faults on seismic is from Xinming Wu  (FaultSeg3D: Using synthetic data sets to train an end-to-end convolutional neural network for 3D seismic fault segmentation. Xinming Wu, Luming Liang, Yunzhi Shi, and Sergey Fomel)

2. The Github repro of Satyakee Sen is a great start to test different networks: https://github.com/satyakees/FaultNet

3. Unfortunately we are not aware of a good public synthetic generator for faulted seismic datasets apart from the one given here by Schlumberger.  An early code version of a synthetic generator could be this: https://github.com/dhale/ipf/tree/master/src/ipf

Fault Interpretation Strategies

Faults come in many shapes, forms and resolutions. The seismic data set chosen here exemplifies this with poorly resolved faults in the deeper layers and very well resolved faults up shallow.

Fault interpretation is also much more purpose driven than seismic reflector interpretation. I.e. for interpreting a regional structural trend one is interested in the major faults and may even connect several fault segments as opposed to detailed reservoir level fault mapping where one is interested in every little detail of the fault. For this competition we would like to see a detailed mapping of the polygonal faulting up shallow and a more regional mapping of the fault trends down deep. An adoption of the patch size both vertically and horizontally as well as treating the seismic line as a whole image first rather than only looking at sub patches could potentially help for this approach.

Unfortunately fault interpretation is not a quantitative exercise on a real seismic dataset. Therefore we decided to use domain experts for the final scoring of your submissions. The outcome of the competition therefore will be to a degree subjective and therefore the organizers decision is final and a recourse to courts is not permitted.

Final Words

While we have made and effort to provide a fair dataset and assessment of the scores we are only humans and prone to mistakes and errors. We accept no liability for our mistakes.

Whatever the outcome of the competition the organizers decision is final and a recourse to courts is not permitted. 

All Sponsors