High-order fault observation association using random forest from analog structural models
in: Proc. 2024 RING Meeting, pages 14, ASGA
Abstract
When creating a geological model from borehole data or 2D sections, the interpretation of 3D faults is often ambiguous and uncertain. This work focuses on the problem of associating partial fault observations, which has recently been formalized using a graph in which each fault observation is represented as a graph node, and graph edges carry the potential of pairwise associations. The likelihood of an association is computed using selected expert geological rules. However, fault observations are not pairwise independent, which prevents the consideration of higher-order effects such as the distribution of the throw or the length along several aligned nodes. To complement this approach, we propose a mathematical formalism for the use of high-order associations. The definition of expert rules in a multiple-point problem is challenging because of the very high dimensionality of the problem. To alleviate this, we propose to supplement expert rules by supervised machine learning using analog or incomplete interpretations. This work uses a Random Forest learner trained from a set of selected fault features computed from fault traces extracted from known 3D geological models (e.g., the length of the fault trace, the throw value, etc.). The association likelihood inference is formulated as a classification problem to determine the probability that ? fault observations belong to the same fault object based on the variables computed from the features of the k observations. To prevent overfitting, we propose to mimic a partly interpreted case: we split the 3D domain in two disjoint, spatially contiguous sectors A and B, and use sector A as training and sector B for testing. Preliminary results demonstrate the ability of Random Forest to retrieve probabilities of triplets that complete the pairwise representation.
Download / Links
BibTeX Reference
@inproceedings{fratani_high-order_RM2024,
abstract = {When creating a geological model from borehole data or 2D sections, the interpretation of 3D faults is often ambiguous and uncertain. This work focuses on the problem of associating partial fault observations, which has recently been formalized using a graph in which each fault observation is represented as a graph node, and graph edges carry the potential of pairwise associations. The likelihood of an association is computed using selected expert geological rules. However, fault observations are not pairwise independent, which prevents the consideration of higher-order effects such as the distribution of the throw or the length along several aligned nodes. To complement this approach, we propose a mathematical formalism for the use of high-order associations. The definition of expert rules in a multiple-point problem is challenging because of the very high dimensionality of the problem. To alleviate this, we propose to supplement expert rules by supervised machine learning using analog or incomplete interpretations. This work uses a Random Forest learner trained from a set of selected fault features computed from fault traces extracted from known 3D geological models (e.g., the length of the fault trace, the throw value, etc.). The association likelihood inference is formulated as a classification problem to determine the probability that ? fault observations belong to the same fault object based on the variables computed from the features of the k observations. To prevent overfitting, we propose to mimic a partly interpreted case: we split the 3D domain in two disjoint, spatially contiguous sectors A and B, and use sector A as training and sector B for testing. Preliminary results demonstrate the ability of Random Forest to retrieve probabilities of triplets that complete the pairwise representation.},
author = {Fratani, Amandine and Stoica, Radu Stefan and Caumon, Guillaume and Giraud, Jérémie},
booktitle = {Proc. 2024 RING Meeting},
language = {en},
pages = {14},
publisher = {ASGA},
title = {High-order fault observation association using random forest from analog structural models},
year = {2024}
}