Results of the FORCE 2020 fault mapping on seismic competition

Following the ground breaking work of Wu (FaultSeg3D: Using synthetic data sets to train an end-to-end convolutional neural network for 3D seismic fault segmentation) in 2019  numerous vendors have implemented Wu like fault detection algorithms into their software and are showing very nice results in webinars and trade shows.

Typically the seismic quality and the faults usually used in these example cases is very good.

We wanted to find out if these very efficient modern machine learning based fault detection algorithms perform equally as good on not so perfect data.

We provided a training seismic dataset from the Ichthys Field on the NW Shelf of Australia to train  the models along with some synthetic fault models from Schlumberger and Equinor.

The blind dataset that we provided to the contestants comes from the Adele seismic survey that is located some 15 to 20 km to the NE of the Ichthys seismic survey.

A total of 80 teams signed up for the competition but only 5 submitted a valid scored fault cube in the end. This is surprising given that we waived the necessity to submit any code and simply asked that the participants to be able to describe their approach in words such that no commercial or intellectual property was infringed.

Perhaps this seismic cube was simply too hard to map faults or the technology is still too immature? 

We will perhaps test the waters with another benchmarking exercise at some later moment to find out

The winner of the competition was determined by evaluation of the fault prediction results (from the contestants) to 19 experienced geoscientists who work normally with seismic. The evaluators ranked the fault prediction results in terms of output usefulness in seismic interpretation workflows on a scale from 1 to 10.

The Winner is Sparveon followed by Equinor and Woodside

Download the full report here!