Do you know that from October 15 to December 15 one of the world's largest championships in seismic data analysis "Rosneft Seismic Challenge" with a total prize fund of 1 million rubles and the final on December 21 in Moscow is being held?
It is believed that getting into the oil industry, where salaries are not inferior to the IT industry, is quite difficult from the outside. There is some truth in this, because the sphere is quite specific and does not favor people who are “off topic”. This event aims to make it easier for young and talented teams working in image recognition and machine learning to enter this underworld.
I am posting this topic in the “I am promoting” section because: a) I want to help fellow Ufa residents; b) I believe in the high qualification of habrazhiteley. And it will be great if one meets the other. At the same time, I will be a translator from technological to human for a while.
So what's the challenge?
The task is: “Isolation of seismic horizons in the amplitude cube - data segmentation using image recognition”. Championship on the Boosters.pro platform. The organizer is the corporate institute BashNIPIneft LLC, one of the leaders (oddly enough) in the development of . An illustrative example of their successful work is the development and – own industrial software for mathematical modeling and analysis of the process of creating fractures during hydraulic fracturing.
I translate the task into Russian
Despite the scary name, the task comes down to image analysis using machine learning. But, as usual, there are many nuances.
Seismic exploration is the main method for discovering oil and gas. The method is based on the excitation of elastic vibrations and subsequent registration of the response from rocks. These vibrations propagate through the thickness of the earth, refracting and reflecting on the boundaries of geological layers with different properties. Reflected waves return to the surface and are recorded. The output is the so-called seismic cube, which is cut into layers vertically and horizontally. We get some sort of sections (crosslines and interlines), which show rocks with different properties.
The task of the participants is to accurately identify and label these horizon layers on the entire seismic cube based on preliminary training at 10% of the cube. As long as it's easy, right?
And now in conventional terms:
“Correlation in seismic exploration is understood as the process of identifying and tracing reflective horizons, various seismic facies complexes (reefs, etc.) in time, depth and space, on seismograms and total time and depth seismic data.
In the process of tracking reflecting horizons, a combination of kinematic and dynamic seismic attributes is used. In their complex analysis, the correlation of the reflecting boundaries of the wave field in space is carried out by tracing the most pronounced extrema (or transition through 0) of the wave field, while mainly taking into account the sign of similarity of neighboring seismic traces.
At the same time, the property of smoothness of the change in the time of registration of the arrival of the wave is taken into account. The line connecting the characteristic features (extrema) of the same wave on different paths is commonly called the in-phase axis. Reflected waves are usually correlated by the most distinct extrema (phases). At the same time, interpreters usually adhere to the principle - from more reliable to less reliable.
First, let's trace the horizons, which in the area under study are confidently traced over a large area and have a proper geological reference. Such reflecting horizons are usually called reference or reference. They are regional markers. Their tracking and interpretation can significantly increase the understanding of all seismic material, tectonic history, and sedimentation conditions.”
Kirilov A.S., Zakrevskiy K.E., Workshop on seismic interpretation at PETREL. M.: PUBLISHING HOUSE MAI-PRINT, 2014. - 288 p.
Need more information?
On this issue, there is a large amount of reference information in Russian in almost any of the formats. Including on Youtube. For example, you can bring an excellent visual video about the automatic recognition of seismic horizons, laid out for free access by the Kazan Center for Additional Education IGiNGT KFU.
It seems to me that after it the task laid down in the challenge should become more understandable.
Ok, what should be done?
Based on the first 10% of the seismic cube already marked by a professional interpreter, it is necessary to mark the remaining slices in the test dataset along the boundaries of the given classes with the maximum metric value.
What to work with?
The initial dataset is a three-dimensional seismic data array (summary time cube of a seismic attribute). As mentioned above, a cube can be represented as 2D vertical slices: crosslines and inlines.
Each slice consists of one-dimensional vectors - traces (trace) with a length of 2562 milliseconds with a step of 2 ms. Number of crosslines: 1896. Number of inlines: 2812.
Total number of tracks > 5 million
Number of segmentation classes (i.e. breed divisions): 8.
Who is waiting for the Seismic Challenge?
The organizers are waiting for the participation of experts from the field of data analysis. The terms are limited and the challenge is suitable for those who “already know how”. Both individuals and teams of up to five people can participate in the competitive selection.
How to participate?
Participants register themselves through the website. on the site Boosters.pro. According to statistics, as of November 4, 402 teams have registered for the competition.
Dates:
15.10.19/15.12.19/XNUMX — XNUMX/XNUMX/XNUMX — contest
24.11.19/XNUMX/XNUMX — end of ability to merge teams
15.10.19/01.12.19/XNUMX - XNUMX/XNUMX/XNUMX - the first round of the contest
02.12.19 — 15.12.19 — the second round of the contest for the best 30 teams from the first round
21.12.19/10/XNUMX - full-time summing up and awarding XNUMX teams from the second round in Moscow.
The organization of the final is interesting: the expert council evaluates the final works, but does not influence the choice of winners. The distribution of finalists is determined by the results of the correspondence part of the competition based on the best segmentation quality metrics (Dice Metrics). At the same time, participants can receive an additional bonus for the best presentation of their solution in the amount of 50 rubles.
PS
I am not the organizer of this challenge, so I can hardly answer questions in the comments in detail. If the habrazhites have questions / interest, then I can invite a representative of the organizers and the guys from boosters to the comments.
Source: habr.com