Testing the geological limits of ML-assisted fault interpretation

Abstract

The increasing role of artificial intelligence (AI), driven by its ability to mimic human intelligence, provides us with opportunities in all kinds of fields, including seismic fault interpretation. This study explores the integration of machine learning (ML) assisted techniques to assist seismic fault interpretation, testing the geological consistency and accuracy of these methods. A 3D seismic cube from the Heimdal Terrace, Viking Graben, was specifically chosen to assess the application of 3D ML-based fault detection tools using GeoTeric.

The study area, Heimdal Terrace, underwent several phases of rifting between the Permian and Early Cretaceous, which resulted in normal faults between the basement and the Lower Cretaceous Cromer Knoll Gp. These normal faults are the focus of this study. The ML fault detection tool was applied to a subset of the seismic cube. Geoteric provides several pre- trained 3D neural network (NN) models for fault detection, and these were fine-tuned by interpreting faults in selected time slices (z direction), inlines and xlines (x and y directions). The NN model applied to the seismic delivers a fault probability map. From this the faults were grouped and labelled to generate fault surfaces, and the resultant fault network was imported into TrapTester where the quality control (QC) of the faults was performed.

The seismic interpretation involved both human and ML approaches to identify key horizons and faults within the seismic cube. I interpreted key horizons in a standard way using human input and ant-tracking. Major faults were also interpreted the same way. These two highlight depth variations, graben and horst features, and mainly NE-SW trending faults. Then, the NN Birch, Larch and Meranti were used for fault interpretation. The Birch and Larch NN were fine-tuned by interpreting faults on selected time slices (z direction), while the Meranti NN was fined tuned on XL and IL sections (x and y directions). After fine-tuning, the Birch NN still struggled interpreting smaller faults and fault intersections. The Larch NN, despite capturing the faults in more detail, sometimes introduced false faults. The Meranti NN is better compared to the other two NN. However, it still had issues connecting fault segments. Overall, although NN demonstrate huge potential and deliver reasonable interpretations, they still require human input to produce a reliable result. This emphasises the importance of human expertise in seismic interpretation.

The QC of the NN fault models yields positive results, despite this there is also three main issues. The Birch NN displays errors such as pinching and crossing of fault cutoffs, resulting in unrealistic throw values. The Larch NN, despite showing improvements over Birch, still exhibit issues with local joining and crossing of cutoffs. The Meranti NN has similar challenges. In addition, it introduces holes in the fault surfaces.

The study demonstrates that while ML models achieve high accuracy and efficiency, they encounter challenges in connecting intersecting faults and interpreting wider fault zones. Post- processing, including fault editing and linkage, is essential to improve the geological model. Although fine-tuning enhances interpretation, it may introduce minor inaccuracies. The ML model effectively identifies major structures but struggles with complex intersections, as seen in QC results. Despite these challenges, NNs are invaluable for data extraction, yet human expertise remains crucial for accurate geological models.