Full text loading...
-
Porosity structure prediction from conventional sonic well logs on the base of synthetic samples computed by Prodaivoda-Maslov's method
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, 18th International Conference on Geoinformatics - Theoretical and Applied Aspects, May 2019, Volume 2019, p.1 - 5
Abstract
The paper aims demonstration of the ML applicability to the problem of rock porosity structure studying by the combination of sonic and density well logs. The experimentally estimated efficiency of popular ML methods for the problem is discussed. In the test we used artificial samples of randomly generated structure with the well log parameters computed by Prodaivoda-Maslov's method of the direct problem solving.
Among the many known algorithms of ML, we selected for the study several ones which are popular and supported by standard Python libraries K-Nearest Neighbors (KNN), Logistic Regression (LRM), a feed forward artificial neural network Multilayer Perceptron (in both the classification form MLPC and the regression form MLPR), Support Vector Machine (SVM), Decision Tree (ID3) and Random Forest (Forest).
We subdivided the problem by independent sub-problems of estimation the concentration of different aspect ratio inclusions a1 cracks (10-3: disks), a2 micro cracks (10-2: disks), a3 pores (100: spheres), a4 caverns (102: streaks). To reconstruct simultaneously 4 unknown parameters we applied multi-task learning. Here represents results only for a1.
The classification algorithms performed generally worse in respect of MAE. Yet the error of about 5% was expected here because the classes were defined by 10% concentration intervals (0–10%, 10–20%, and so on). More interesting is their inability to identify right class. It is expressed by the accuracy score. The best classification algorithm MLPC leaded in both MAE and classification accuracy competitions. But its classification accuracy score is only 72. 4%.
The tests have demonstrated the ability of machine learning algorithms to estimate concentration of a known subtype inclusions on the base of sonic logs and density. The best regression algorithm, Random Forest, with its Mean Absolute Error MAE = 1. 7% in concentration provides excellent quality. Two other good reg ression algorithms demonstrate acceptable MAE < 5%.
It would be interesting to apply the ML methods to real core data. We invite for collaboration those who have access to core collections and the ability to execute more detailed analysis of the core porosity than usually.