Seismic Facies identification with RNN claims better results than CNN

The RNN seems a natural choice for seismic, team MathWorks at the SEAM AI Applied Geoscience GPU Hackathon thought the same and used RNN with Gated Recurrent Unit (GRU) layer. Based on the F1-weighted score, the team was named the winner for test dataset 1. (the link in the comments)

Facies identification is usually treated as a segmentation problem, where some modification of U-net architecture is implemented. An input could be just raw seismic, raw seismic+attributes, or in other cases, raw seimsic+relative geological time (I recently reviewed a couple of papers, here is my post https://lnkd.in/dJ87DtA). The results are reasonably good and geologically consistent but require a lot of data and time to train the network.

It is a great step to excel at a single dataset; however, it is harder to claim the general applicability that we strive for; when trained once, it does not need to be retrained for a particular seismic at hand. It seems like a far-fetched goal in Oil and Gas due to data problems (5 V's of big data).

The link for the news: https://blogs.mathworks.com/deep-learning/2021/08/03/mathworks-wins-geoscience-ai-gpu-hackathon/
LinkedIn post