Transfer Learning: A Key Approach to Fault Prediction and Extraction in Deep Learning
Muhammad Haris Khan, Andy Anderson Bery, Syed Sadaqat Ali, Saleh A. Al-Dossary · 2024
ABSTRACT: Faults are key subsurface features that significantly influence geomechanics, affecting stress fields and playing a crucial role in hydrocarbon exploration, production, and CO2 storage. The redistribution of stress around faults depends on factors like fault size, orientation, and the mechanical properties of surrounding rocks. Seismic data is commonly used to interpret faults, but the process can be labor-intensive and subjective. Structural attributes such as variance, curvature, fault likelihood, and ant-tracking have improved fault detection, but still require substantial manual editing to create a comprehensive fault framework. To overcome these challenges, we propose a novel approach using 3D deep convolutional neural networks (DCNN) with transfer learning from six dip lines to incorporate local structural knowledge. This method generates fault probability volumes with high training and validation accuracies (0.98 and 0.965). Ant-tracking is used as a post-processing step to enhance fault discontinuities both vertically and horizontally, allowing for the automatic extraction of fault planes ranging from 58 km to 1.25 km in length with minimal manual intervention. We successfully applied this approach to the Poseidon 3D dataset, detecting major and minor discontinuities, demonstrating its robustness and effectiveness in complex fault systems. 1. INTRODUCTION The interpretation of faults has traditionally represented a fundamental aspect of subsurface characterization within the fields of geomechanics, hydrocarbon exploration, geothermal and assessments of CO2 storage and sequestration. Initially, the identification of faults was accomplished through manual interpretation techniques, wherein skilled geoscientists visually analyzed seismic sections for structural discontinuities. This process required the identification of abrupt changes in reflector continuity, displacements within stratigraphic layers, or other geophysical indicators that implied fault activity. However, manual interpretation is inherently laborious and subjective, leading to variability in results that depend on the interpreter's proficiency and understanding of the geological framework. Despite these limitations, manual interpretation laid the groundwork for the advancement of more refined fault detection techniques by establishing criteria for fault recognition in complex structural settings.