Bio: Cody MacDonald holds a B.Sc. and M.Sc. in geology from Dalhousie University, Halifax, Nova Scotia, Canada. For just over a decade, he was employed at ExxonMobil, where he had roles in exploration, technology, applied geophysics, and research & development. Previous experience includes leading machine learning and deep learning projects and programs at ExxonMobil. Following interests in AI and deep learning at scale, Cody left ExxonMobil and joined SambaNova Systems in 2022, where he currently works as a customer engineer, with a heavy focus on computer vision.

Abstract: Deep learning for geophysical analysis, such as seismic interpretation has been an active field of R&D for several years (Waldeland, et al., 2018). Despite encouraging beginnings and early software development, uptake has been relatively slow. Some of this lag in adoption is due to the complex, scientific, and process driven nature of petroleum exploration. However, other factors that are frequently attributed to a lack of progress, include: a lack of available labels for supervised learning, limited easy-to-use software to create such labels, a readily available source of datasets (even within large organizations), computationally sufficient resources for AI at the scale of geoscientific datasets, and the requisite combination of geoscience, data science, machine learning engineering expertise required to productionize these methodologies. Although self-supervised approaches, such as contrastive learning or masked image modeling, may provide a means to pre-train models and help alleviate label requirements, labels of a wide variety of features will be needed to fine tune models for specific interpretation tasks. Leading industry software from SLB, Bluware, GeoTeric, Enthought, and others are beginning to provide tools for both labeling and active learning, but these packages are expensive for enthusiast use. Luckily, with Python, there are many free basic options available, from packaged interfaces such as LabelMe to customizable dashboard frameworks, such as Plotly Dash and Streamlit. This presentation aims to give an overview of tools and Python utilities that make it easier to get started with deep learning for seismic interpretation. Covered materials will include: dealing with typical seismic data formats, creating labels with free tools, and how to digest these data and labels into a machine learning framework. Computational challenges in scaling to high quality results will be touched on, but not the focus of this talk. References: Anders U. Waldeland, Are Charles Jensen, Leiv-J. Gelius, and Anne H. Schistad Solberg, (2018), "Convolutional neural networks for automated seismic interpretation," The Leading Edge 37: 529–537.