How to create a highly consistent well database in a limited amount of time?
Back in April this spring the Organising Committee behind the upcoming FORCE Machine Learning Contest chaired by Peter Borman from ConocoPhilips, Fahad Dilib from Equinor, Peder Aursand from AkerBP, Surender Manral from Schlumberger, Peter Dischington from the Norwegian Petroleum Directorate and Matt Hall and his team from Agile Geoscience got in touch to see if we in ExploCrowd could help them with one of their challenges.
The challenge
The COVID-19 situation was raging Europe at that time, why the Organising Committee worked to find alternative solutions to a planned Hackathon that was going to take place.
So they pivoted, planned ahead and needed a high quality well database for the Norwegian North Sea with interpreted Lithology and associated Confidence for their new Machine Learning Contest.
The members of the Organising Committee had been asking around, but nothing was available from other vendors within such a short timeframe and limited budget available from initial sponsorships.
The focus of the database should be high quality, not quantity, consistency in the interpretations and they needed 100-150 wells for the ML Contest - and we would have six weeks to do the job.
This is how we solved it
Having worked with the IC well database software from Lloyds Register before, we reached out to the IC team and asked for their help to solve this challenge despite all the constraints. Derek and Ross were convinced of the positive aspects of this project and were kind to sponsor three IC licences for the duration of the building of the Well Database and make the IC team available to help us along the way.
Our idea was that the IC software would enable good consistency and speed, while maintaining quality. By having all the wells easily loaded into a database we could acquire a global perspective of the data before the work even starts. This allowed us to plan efficiently based on the real data and minimize any surprises. The global perspective of our data is also powerful while the work is being done, allowing us to make quick evaluations and take swift actions to adapt to new patterns or new instructions from the Machine Learning Community. As a team of three people working remotely, these possibilities within IC were indispensable to fine tuning our work and strive towards consistency of geological interpretation, while limiting ‘repeat-work’ to a minimum.
With licence sponsored by IC, we were confident that we could deliver the ML database within the tight time frame, and we felt that we had created yet another win-win-win situation for FORCE, Lloyds Register and three of our people from the Core Team who could be pulled out of furlough during a deep industry crisis.
Data preparations and developing the methodology
The first part of the project was used to get access to well data that were delivered by the FORCE Organising Committee, and all relevant Mud Logs, Composite Logs and Completion Logs were downloaded from the DISKOS database and made ready for hitting the ground running when a routine would be established.
Gustavo, Elzbieta and Sveinung worked hard to develop a work methodology in close collaboration with the FORCE Organising Committee and Machine Learning Community that could enable speed while delivering the high quality required.
Based on this new established routine we can deliver ca. 100 interpreted wells per month.
Delivering Interpreted Lithology Classes with assigned level of Confidence
Simplified Lithology Classes have been defined during the initial phase of the project, where some lithologies were grouped for the Machine Learning purpose as shown below. Note the new proposed key 7000032 key for Chalk lithology in the overview below:
An example of a well with interpreted Lithology and Confidence is shown below. See the Lithology Key for legend. Legend for the assigned confidence is as follows: Green- High Confidence, Yellow- Moderate Confidence and Red - Low Confidence.
Well DatAbase will be released
The plan is to release this Well Database this summer and make it public for the FORCE Machine Learning Contest - and hereafter for everyone to use in the future. Help spread the word and follow the developments via the FORCE website, hosted by the Norwegian Petroleum Directorate. More information about the FORCE 2020 Machine Learning Contest here.