Back to search


Machine Learning in Geoscience; Weakly-Supervised Seismic Facies Interpretation

Alternative title: Maskinlæring i geovitenskap; weakly-supervised identifisering og klassifisering av seismiske facies

Awarded: NOK 1.8 mill.

Project Manager:

Project Number:


Application Type:

Project Period:

2022 - 2025

Funding received from:


Seismology is popularly known from the study of earthquakes. However, seismology can also entail the exploration for natural resources using man-made seismic sources. Instead of an earthquake, a large vibrating truck can hit the ground, or an air-gun can send pressure waves through the ocean, creating waves that travel down into the Earth. These waves reflect off rock layers, are recorded by receivers at the surface, and then can reveal an intuitive image of the subsurface. This is similar to how medical imaging can reveal the structure of the human body. Geophysicists in the oil industry analyze seismic data in order to locate hydrocarbon deposits, so that wells are drilled in worthwhile locations. Certain types of rock are more likely to contain hydrocarbons, which is why being able to differentiate them in seismic data is important. A classification of a rock type is called a facies. For example, a facies of a shale could be a good source rock for oil, while a facies of a sandstone with lots of space in between its grains could be a good reservoir rock. The structure of rock layers is also important as a facies characteristic, as the subsurface is rarely in simple horizontal layers. Pressure within the earth can create folds and faults that can change the structure and trap oil deposits in different ways. The process of interpreting this subsurface data can be tedious, however. Often facies must be identified, horizons separating layers must be drawn, and faults must be highlighted over large areas where the data was collected. Repetitive tasks such as these are good candidates for automation via machine learning. Machine learning algorithms have become common for tasks such as text classification, speech recognition, and -- most relevant to seismic interpretation -- image identification. Thus applying these algorithms can reduce the amount of tedious work that an exploration geophysicist needs to do, and leave more time for creative tasks.

The proposed doctoral work will focus on the application and development of new data-driven methods that take advantage of existing knowledge in the fields of image processing, signal processing, and machine learning (ML). This will involve exploring both supervised and unsupervised machine learning algorithms, and possibly combining them in a practice called weakly-supervised learning. The developed methods will aim to be generalizable, and thus independent of the seismic survey data they are applied to. In the development of these methods the candidate will use geophysical data from the Norwegian continental shelf (NCS) provided by Lundin Energy Norway. This data will include seismic surveys, well logs, and cutting images (from the Released Well Initiative), in addition to synthetically-modelled seismic data. In short, the the candidate's research will explore the following: • Use state of the art supervised, unsupervised, and weakly-supervised ML network architectures, and apply them to seismic facies identification and classification. • Improve the training optimization of ML algorithms to learn seismic stratigraphy from available subsurface data (3D seismic data, well logs, and cutting images). • Develop more optimal representations of seismic character and geometry usingsynthetically-modelled seismic stratigraphic patterns. The use of such synthetic data has two main purposes, (1) as labels for a supervised or weakly-supervised approach, and (2) as benchmark data for quantitative evaluation of new methods.

Funding scheme: