Can neural networks learn how to associate structural data? {First} results.

@inproceedings{fratani_can_2022,
 abstract = { In frontier domains, fault modelling requires interpreting sparse seismic or borehole data, which is prone to structural uncertainties. Previous works on the problem of associating sparse fault evidence use graphs and expert rules to describe this task (Godefroy et al., 2019). Our work aims to evaluate the ability of machine learning to replace or complement these expert rules to determine the probability of association of sparse fault data. Precisely, we want to use the capacity of graph neural networks to work with sparse fault observations. We first consider the problem in two dimensions. We have created a database composed of fault data sampled from randomized synthetic boreholes on sections extracted from a set of 3D structural models. The probability to correlate two fault evidences in the same section is studied by comparing typical statistical analyses to random forest and to neural networks. Then, we consider multiple fault data with a graph neural network to assess the interactions between faults. },
 author = { Fratani, Amandine AND Caumon, Guillaume AND Antoine, Christophe },
 booktitle = { 2022 {RING} {Meeting} },
 pages = { 12 },
 publisher = { ASGA },
 title = { Can neural networks learn how to associate structural data? {First} results. },
 year = { 2022 }
}