Smart Modular Seismic Fault Segmentation via Deep Learning Frameworks

Thesis Abstract

This study addresses the escalating need for rapid and accurate delineation of seismic fault geometries in heterogeneous lithologies, recognizing that conventional manual interpretation is time-consuming and prone to subjective bias, particularly in data-sparse regions. The aim is to develop a robust, modular deep learning framework that automates seismic fault segmentation from 3D seismic volumes and intercalibrates with well-log and surface geology data to enhance reliability and scalability for exploration and hazard assessment. Specific objectives include (1) designing a modular architecture that integrates convolutional neural networks (CNNs) with transformer-based components to capture both local texture cues and long-range structural correlations, (2) evaluating multi-attribute input pipelines (amplitude, phase, coherence, curvature) and their fusion strategies to improve fault boundary delineation, (3) implementing uncertainty-aware predictions through Bayesian deep learning and Monte Carlo dropout to quantify segmentation confidence, (4) benchmarking the framework against expert-labeled seismic fault maps across eight field datasets with varying tectonic regimes, and (5) delivering an open-source toolkit with documented APIs and reproducible workflows for industry and academia. The research adopts a pragmatic mixed-methods design, combining quantitative performance evaluation with qualitative expert validation to assess reliability in operational settings. The population comprises 3D seismic datasets from eight oil and gas basins with publicly available well correlations and regional fault maps, totaling approximately 2,400 square kilometers of volume coverage and over 1,200 labeled fault slices. A stratified sample of 1,200 labeled slices, balanced across fault-present and fault-absent regions, is used for supervised training, with an additional 400 unlabeled volumes reserved for semi-supervised refinement. Data collection leverages publicly accessible seismic surveys, synthetic data generated through forward modeling to augment rare fault geometries, and corresponding borehole and surface mapping for ground-truth validation. Instruments include high-performance GPUs for model training, standardized annotation protocols for fault delineation, and software tools incorporating TensorFlow and PyTorch, complemented by domain-specific plugins for seismic attribute extraction. The methodology comprises a two-stage model development process (i) a modular segmentation backbone employing a 3D U-Net variant to extract multiscale fault features from input attributes, and (ii) a contextual refinement module using a vision-transformer to capture long-range coherence along fault planes, fused through attention-based gating mechanisms. Training employs supervised cross-entropy with focal loss to address class imbalance, supplemented by a semi-supervised consistency loss for unlabeled data. Model validation uses k-fold cross-validation and a held-out test set, with performance metrics including intersection-over-union (IoU), Dice coefficient, precision, recall, and area under the precision-recall curve. Uncertainty is quantified via Monte Carlo dropout and deep ensembles to produce probabilistic fault probability maps. Comparative analyses incorporate regression-based calibration of predicted fault geometries against well-log-derived fracture densities, and Bayesian model comparison using Bayes factors to assess model adequacy. The study anticipates that the modular framework will achieve IoU improvements of 8–12 percentage points over baselines and provide reliable uncertainty estimates that reduce misclassification rates in complex fault zones. Expected findings include that the fusion of coherence attributes with transformer-based context yields more continuous and geologically plausible fault surfaces, particularly in chaotic seismic facies, and that uncertainty maps correlate with regions of sparse control. The contribution to knowledge lies in delivering a generalizable, open-source, modular deep learning framework for seismic fault segmentation that integrates attribute fusion, long-range spatial reasoning, and probabilistic outputs, supported by a rigorous validation across diverse tectonic settings. The study will inform best practices for deploying AI-assisted interpretation workflows in exploration geoscience and hazard assessment, and will offer guidelines for integrating segmentation outputs with structural restoration and reservoir modeling. The main conclusion is that modular deep learning architectures with contextual transformers and uncertainty quantification can substantially outperform conventional methods in seismic fault delineation, enabling faster, more consistent interpretations. Recommendations include extending the framework to time-lapse seismic monitoring, incorporating multi-physics data (gravity, magnetics), and developing standardized benchmarks for cross-region evaluation to foster broader adoption in industry and research.