Paul de Groot (dGB Earth Sciences) and Friso Brouwer (l^3 GEO)
Abstract
Often exploration for oil and gas with seismic data starts with a quick look analysis to assess multiple areas, datasets and stratigraphic levels. This aids both the general understanding of the area and the data. It also helps prioritizing areas that are worth studying in greater detail using more time consuming and expensive workflows. In this presentation, two quick Machine Learning-based workflows are presented that aid identification of good quality reservoir elements in a fast, yet rigorous manner. The first method is a new application of an old favorite: UVQ (Unsupervised Vector Quantizer) waveform segmentation. The novelty is that instead of creating a segmented response along a mapped horizon, we now create a high-resolution 3D volume of segmented waveforms. The seismic patterns in the 3D segmentation volume are subsequently interpreted in terms of geomorphological features. To facilitate the interpretation, the segmentation is performed on a color inverted (relative) acoustic impedance volume and the interpretation is done along chrono-stratigraphic events extracted from a HorizonCube (a dense set of automatically generated horizons).
In the second workflow, we present a supervised workflow. A Convolutional Neural Network (CNN) segments a 3D seismic volume into different seismic facies classes. The CNN is trained on examples that were semi-automatically created by a human interpreter using a Thalweg tracker. A Thalweg tracker tracks 3D bodies along the path of least resistance. In practice this means that we can track sedimentary features such as channels, lobes, splays, and reef buildups. Application of the trained CNN yields a 3D segmentation volume that is further analyzed in the Wheeler domain (flattened seismic domain).
Both examples were created in OpendTect utilizing different plugins, such as Machine Learning, Seismic Colored Inversion, and HorizonCube.
About the presenters
Friso Brouwer
is geophysical consultant and the owner of I^3 GEO in Denver, USA. Friso has 20 years industry experience, involving both technical and managerial roles. His expertise covers both regular seismic interpretation and application of specialized geophysical techniques. Friso has worked offshore, onshore and unconventional areas, both in the US and internationally. His focus is on interdisciplinary work, and his areas of expertise include seismic interpretation, well log analysis for geophysics, seismic inversion and machine learning.
Paul de Groot
is co-founder of dGB Earth Sciences. Paul started his professional career as a geoscientist with Shell for whom he worked ten years in various technical and management positions. Subsequently Paul worked as a senior research geophysicist for TNO Institute of Applied Geosciences before co-founding dGB in 1995. At dGB he divides his time between driving OpendTect, dGB’s seismic interpretation system, forward and developing business opportunities in an open source business model. Paul has authored many papers covering a wide range of geophysical topics and co-authored a patent on seismic object detection. Together with Fred Aminzadeh Paul wrote a book on Soft Computing techniques in the Oil Industry. Paul holds MSc and PhD degrees in geophysics from Delft University of Technology.
download presentation:
PGK lecture 2022-05-18 Paul de Groot and Friso Brouwer – 3D seismic facies segmentation using supervised and unsupervised learning approaches
0 Comments