A recent webinar from Gazprom and Nvidia introduced a suite of open source software tools for machine learning geophysical applications. Gazprom’s Anna Dubovik presented the toolkit capable of ‘labeling every voxel of a seismic field’. Gazprom’s publicly-available tools include SeismicPro (noise reduction, automated first break picking), SeismicQB (horizon and fault tracking), PetroFlow (core-log matching, petrophysics) and PyDens (a differential equation solver).
Dubovik believes that upstream data is particularly amenable to machine learning. Inherent high uncertainty means that the extra knowledge is most useful, in contrast to the downstream where the value of extra information declines with more sales-oriented activity. PyTorch-based ML is now an integral part of Gazprom’s business. Models have been extensively trained on large data volumes and the company now wants to ‘broadcast these successful results to the world’.
Sergey Tsimfer took over to present SeismicPro showing an AI-based cleanup process leveraging a local homogeneity cube prior to horizon tracking. ‘Defects’ are displayed on a quality map. First break picking for static correction is tricky. Autopickers are thrown by the slightest noise. Manual picking by experts is fastidious. ML is ‘faster and more accurate’. In seismic interpretation, ‘manual/autocorrelation software is a slow process’. Gazprom proposes a neural net-based approach trained on many fields and data. With no human interaction, a 3D volume can be picked in a few hours. Seismic stratigraphic interpretation also ran, using labelled models to identify stratigraphy such as marine fans. Fault interpretation? Again, ‘manual picking is error-prone and fastidious’. ‘Experts differ and interpretation is inconsistent’. Gazprom’s neural net has been trained on its legacy field interpretations and now ‘detects faults automatically’. The ML-derived velocity model building is the ‘most sophisticated yet’.
ML is the secret to ‘major profits’ and the best returns from the upstream. But ‘there is no easy way to get into AI’. ‘Until now that is’. Gazprom has released the fruits of its AI/ML seismic efforts as an open source code base on GitHub under an Apache 2.0 license. Test data sets from Gazprom and Equinor are available with more from the community and Nvidia to come this year. Workshops are planned for the EAGE Learning Geoscience platform. Check out the blog on Medium and join the project on Git.
© Oil IT Journal - all rights reserved.