Predict stratigraphic surfaces based on training on human-picked stratigraphic surfaces. Used 2000+ wells with Picks from the Mannville, including McMurray, in Alberta, Canada.
Instead of assuming there is a mathematical or pattern basis for stratigraphic surfaces that can be teased out of logs, focus on creating programatic features and operations that mimic the comparison-based observations that would have been done by a geologist.
There has been studies that attempt to do similiar things for decades. A lot of them assume a mathematical pattern to stratigraphic surfaces and either don't train specifically on human-picked tops or do so lightly. We wanted to try as close a geologic approach (as opposed to mathematical or geophysical approach) as possible. What we managed to get done by the end of the hackathon is sorta a small scale first pass.
Eventually, we want to get to the point where we've identified a large number of feature types that both have predictive value and can be tied back to geologist insight. There are a lot of observations happening visually (and therefore not consciously) when a geologist looks at a well log and correlates it. We want to focus on engineering features that mimic these observations and the multitude of scales at which they occur.
In addition to automating correlation of nearby wells based on picks that already exist, which has value, we think this will help geologist have better discussions, and more quantitative discussions, about the basis of their correlation and why correlations might differ between geologists. You can imagine a regional area with two separate teams with different approaches to picking a top. You could use this to programmatically pick tops in area B like they are picked in area A and also the inverse. The differences in pick style then becomes easier to analyze with less additional work.
Report for Athabasca Oil Sands Data McMurray/Wabiskaw Oil Sands Deposit http://ags.aer.ca/document/OFR/OFR_1994_14.PDF
Electronic data for Athabasca Oil Sands Data McMurray/Wabiskaw Oil Sands Deposit http://ags.aer.ca/publications/SPE_006.html Data is also in the repo folder: SPE_006_originalData
@dalide used the Alberta Geological Society's UWI conversion tool to find lat/longs for each of the well UWIs. These were then used to find each well's nearest neighbors as demonstrated in this notebook.
On February 11th, 2018, @JustinGosses reorganized the folder to get a lot of the notebooks out of the top-level and into sub-folders as things were getting too crowded. This might cause the directory urls to some files to be incorrect. This will be the case for any notebook from the Hackathon or 2017. Fixing this problem will just require adding a ../ or ../../ to the front of the directory in most cases.
Final Data Prep & Machine Learning for the prediction finished by end of hackathon https://github.com/JustinGOSSES/MannvilleGroup_Strat_Hackathon/blob/master/data_prep_wells_xgb.ipynb
Version of feature engineering work done during hackathon (but didn't get to include during hackathon) https://github.com/JustinGOSSES/MannvilleGroup_Strat_Hackathon/blob/master/Feature_Brainstorm_Justin_vD-Copy1.ipynb
Code development has moved to the modular_redo sub-folder. Things were made more modular to better enable short bits of work when time available. The notebooks are a bit messy but will clean up in near future.
https://github.com/JustinGOSSES/MannvilleGroup_Strat_Hackathon/tree/master/notebooks_2018/modular_redo]
The code runs faster and and mean absolute error is down from 90 to 15.03 and now 7+. Key approaches were:
4-1. First step is class-based prediction. Classes are groups based on distance from actual pick. For example, depths at the pick, depths within 0.5 meter, depths within 5 meters above, etc. 4-2. Second step is more concerned with picking between the depths predicted as being in the classes nearest to the pick. We've explored both a rule-based scoring and a regression machine-learning process for this. 4-2-1. The rule-based approach uses the class prediction and simple additive scoring of the class predictions based across different size windows. In a scenario where there are two depths with a predicted class of 100, we decide between them by adding up all the class scores across different size windows above and below each depth. The depth with the highest aggregate score wins and it declared the "predicted depth". We take this route as we assume the right depth will have more depths near it that look like the top pick and as such have higher classes predicted for depths around it while false positives will be more likely to have more lower level classes around it. 4-2-2. We're also trying regression-based machine-learning to predict the distance from each depth in question to the actual pick. The depth with the lowest predicted distance between it and actual pick is chosen as the "predicted pick". This approach hasn't given any better results than the simple rule-based aggregate scoring.
Y-axis is number of picks in each bin, and X-axis is distance predicted pick is off from human-generated pick. <img src="current_errors_TopMcMr_20181006.png" alt="image of current_errors_TopMcMr_20181006" style="float: left; margin-right: 25px;" />
Current algorithm used is XGBoost.
The plan is that once things are winnowed down to a final approach, the resulting code will be moved the StratPickSupML repository will it will be cleaned into one or more modules and demo notebooks with less clutter of failed but possibly useful if reworked approaches.
This repo isn't particularly organized and there hasn't be a lot of time spent (actually no time spent) to make jumping in and helping out easy. That being said, there's no reason you couldn't just jump in an start improving things. The original group is working on this at a low level when we have time. There are a few issues that are enhancements that would be a good place to start.