Automated Seismic Anomaly Detection via Cloud-Based Inversion Frameworks

Thesis Abstract

This study addresses the gap in rapid, scalable interpretation of seismic data by leveraging cloud-based inversion frameworks to automatically detect anomalies indicative of subsurface heterogeneity, faulting, and potential resource anomalies. The aim is to develop a robust, cloud-enabled pipeline that integrates advanced inversion algorithms with real-time data ingestion, enabling near-real-time anomaly detection with quantified uncertainty. Specific objectives include (1) designing a modular cloud architecture that supports seismic data preprocessing, model-based inversion, and anomaly scoring; (2) evaluating multiple inversion strategies—including full-waveform inversion (FWI) and travel-time inversion—under probabilistic Bayesian frameworks to quantify uncertainty; (3) implementing machine learning-driven feature extraction and anomaly classification to improve detection sensitivity; (4) validating the framework against synthetic and real-world seismic datasets with known anomalies; and (5) delivering a decision-support interface for geophysicists with interpretable results and provenance tracking. The methodology adopts a mixed-methods research design combining quantitative algorithmic evaluation with qualitative assessment of interpretability and user workflow integration. The population comprises archival and newly acquired 3D seismic surveys from two regional basins with well-characterized geological settings. A stratified sample of 40 seismic lines from 5 datasets is selected to ensure diverse wavelet characteristics, noise levels, and geologic complexity. Data collection involves (i) assembling high-resolution post-stack and pre-stack seismic traces, (ii) acquiring well-log based lithology and porosity labels for supervised learning, and (iii) collecting expert annotations of representative anomalies for validation. Data collection instruments include an open-source cloud platform (e.g., Google Cloud Platform or AWS) for storage and processing, standardized data formats (SEG-Y and HDF5), and software toolchains implementing FWI, adjoint-state methods, Bayesian inference modules, and convolutional neural networks. Analytical methods comprise (i) preprocessing steps including amplitude normalization, deconvolution, and ambient noise attenuation; (ii) inversion procedures—FWI and travel-time inversion—embedded within a probabilistic framework to produce posterior distributions of velocity and density models; (iii) feature extraction using deep learning-based encoders to derive anomaly scores from reconstructed models; (iv) statistical evaluation of inversion performance via metrics such as root-mean-square error (RMSE) against synthetic truth, structural similarity (SSIM) with reference models, and calibration of predictive intervals; (v) hypothesis testing to compare the efficacy of inversion approaches under varying signal-to-noise ratios; and (vi) interpretability analysis employing SHAP values to relate model outputs to geophysical drivers. Uncertainty management relies on Bayesian model averaging and Markov Chain Monte Carlo (MCMC) sampling to quantify posterior confidence in anomaly detections. Expected findings include (a) a scalable cloud-based inversion pipeline that reduces end-to-end processing time from hours to minutes per 3D block while maintaining or improving anomaly detection accuracy; (b) demonstrated superiority of Bayesian FWI over deterministic counterparts in uncertainty quantification and risk ranking of anomalies; (c) robust integration of ML-derived features with physics-based inversions yielding higher true-positive rates with controlled false positives; and (d) a user-centric interface that presents interpretable, provenance-rich results to geophysicists, supporting rapid decision-making in exploration and hazard assessment contexts. The study contributes to knowledge by bridging seismic inversion science with cloud-native compute paradigms and machine learning to deliver repeatable, transparent anomaly detection workflows with quantified uncertainty. It advances practical methodologies for automating seismic interpretation while preserving geological plausibility and traceability, thus enabling broader adoption in industry and academia. The main conclusion is that cloud-based inversion frameworks, when combined with probabilistic inversion and interpretable ML components, can deliver timely, reliable seismic anomaly detection suitable for real-time decision support. Recommendations include extending the framework to multi-physics data fusion (gravity, magnetotellurics), expanding validation to diverse tectonic settings, and developing standardized benchmarks and open datasets to facilitate cross-institutional replication and benchmarking.