Automating salt interpretation in Gulf of Mexico (GoM) is a significant challenge. Various factors such as shape, roughness, size, composition, depth, and the type of seismic data acquisition, vintage, and imaging algorithm can impact image resolution, texture, and artifacts. However, using seismic images with similar processing, vintage, and acquisition geometry makes it possible to create machine learning (ML) models explicitly tailored for regional salt interpretation.

Using salt interpretation as a motivating application, we show how several innovations in data engineering and deep learning that enables our workflow. We will discuss how MDIO, an open-source format for storing multidimensional energy data enables us to build efficient training and inference pipelines to run experiments both on on-prem datacenters and cloud.

In this talk, we show an example on building a generalized salt interpretation model in the eastern Gulf of Mexico.