Fast AI Fault Prediction Using Sparsely Interpreted Labels
Wei Xiong, Lei Li, Jingbin Cui, Xiaofang Zhang, Lihua Dai, Chen Xu · International Petroleum Technology Conference · 2025
Abstract Faults play a crucial role in the exploration and development of oil and gas resources. In recent years, deep learning algorithms for fault interpretation have shown signs of progress and are still in a rapid development stage. Since most of the research at this stage uses 3D synthetic data, ignoring the complex geological structure of the field data. This paper focus on the training data issue, using true fault sticks as labels. An interactive training data generation tool is utilized to automatically establish pairs of seismic data and fault labels, forming a 2D training data set characterized by local fault features. In the model training phase, a lightweight 2D UNETR++ architecture is trained using 2D true data to enhance fault recognition accuracy. Additionally, to tackle the problem of discontinuities results from 2D training data, an optimized U-Net model is used to improve the continuity. The proposed intelligent fault prediction method based on real labels can be applied for rapid prediction and identification of major faults, offering reliable data for automated fault interpretation and subsequent structural modeling.