Deep transfer learning for seismic characterization of strike-slip faults in karstified carbonates from the northern Tarim basin

Abstract

The largest pre-Mesozoic ultra-deep (> 6000 m) strike-slip fault-controlled oilfield has been discovered in the northern Tarim Basin of northwestern China, and a deeper interpretation of strike-slip faults is crucial for optimizing well trajectory and development plans. Conventional seismic methods struggle to image strike-slip faults in karstified areas. With the advancements in deep learning, researchers have begun to use it to detect seismic faults. However, challenges persist in constructing actual fault labels and obtaining a large amount of fault labels. For this contribution, we propose a method for constructing fault labels and introduce a deep transfer learning workflow using Unet to detect strike-slip faults in the northern Tarim Basin. The results demonstrate that this method effectively suppresses non-fault features such as karstification and provides clear imaging of fault geometry. Multiple NW- and NE-striking strike-slip faults were identified within the study area, which is consistent with well data and seismic interpretations. Analysis of deep transfer learning attributes revealed four styles of faults, and the degree of fault connectivity plays a significant role in hydrocarbon accumulation. The results of this work highlight the effectiveness of deep transfer learning in fault characterization and suggest its potential applicability in other regions with complex geological conditions.