AI-Driven Seismic Tomography for Real-Time Subsurface Mapping

Thesis Abstract

The rapid integration of artificial intelligence (AI) with seismic tomography offers transformative potential for real-time subsurface mapping, addressing persistent challenges in latency, resolution, and uncertainty that constrain conventional seismic imaging in hydrocarbon, geothermal, and geotechnical applications. This study investigates an AI-driven seismic tomography framework designed to deliver high-resolution subsurface velocity models in near-real time, providing actionable insights for hazard assessment, resource exploration, and rapid decision-making under dynamic subsurface conditions. The aim is to develop, validate, and operationalize an integrated workflow that combines neural network inversion with physics-informed constraints to produce stable velocity tomograms from sparse seismic arrays. Specific objectives include (i) to design a multi-fidelity data assimilation pipeline that fuses passive and active seismic data with prior geological information; (ii) to implement a physics-informed neural network (PINN) architecture that enforces wave propagation physics and seismological invariants in the inversion process; (iii) to quantify uncertainty in generated tomograms using Bayesian neural networks and ensemble methods; (iv) to evaluate real-time performance against conventional tomography in controlled synthetic scenarios and field datasets; (v) to assess the robustness of the approach under varying noise levels, data gaps, and heterogeneous subsurface conditions; and (vi) to develop a deployment blueprint for operational environments including data governance, computational requirements, and user interfaces for decision-makers. Methodologically, the research adopts an embedded, mixed-methods design combining quantitative model development and qualitative validation of interpretability and decision-support value. The population comprises seismological datasets from three pilot sites with differing geology a sedimentary basin, a volcanic complex, and a fault-controlled geothermal field. The sample includes 12 synthetic scenes generated from high-fidelity forward models and 18 real-world campaigns totaling 4,500 seismic shot records and 9,000 receiving stations. Data collection instruments consist of portable broadband seismometers, rapid-deploy arrays, and publicly available regional velocity models, supplemented by geological maps and borehole logs. The core analytical methods include (1) physics-informed deep learning for inverse problem solving, (2) Bayesian neural networks for probabilistic tomography with credible intervals, (3) multi-task learning to jointly invert for P- and S-wave velocities, (4) graph-based regularization to preserve geological continuity, and (5) comparative evaluation against traditional traveltime and full-waveform inversion using metrics such as mean absolute error, structural similarity index, and posterior predictive checks. Uncertainty is propagated through hierarchical Bayesian updating as new data arrive, enabling real-time model refinement. The study also employs sensitivity analysis and ablation experiments to identify the contribution of data types and regularization terms. Validation utilizes four criteria fidelity to known subsurface features, resolution tests via point-spread functions, computational performance benchmarks (aiming for sub-2-minute updates on a 32-core CPU and GPU-accelerated environment), and interpretability assessments with domain experts. Expected findings include (i) improved tomographic resolution in complex geologies relative to baseline velocity models, (ii) robust real-time performance with average update latencies under 2 minutes for moderate arrays, (iii) quantified uncertainty maps that correlate with data density and model nonlinearity, (iv) demonstrated stability of PINN-based inversions under moderate noise and missing data, and (v) enhanced decision-support value evidenced by expert judgments of model usefulness in hazard modeling, resource characterization, and rapid response scenarios. The contribution to knowledge encompasses advancing AI-driven, physics-constrained tomography as a scalable approach for real-time subsurface mapping, integrating uncertainty quantification into operational workflows, and providing a replicable methodology applicable across geological settings. The main conclusion posits that an AI-driven, physics-informed tomography framework can deliver accurate, timely, and interpretable subsurface velocity models under real-time constraints, with uncertainty explicitly communicated to end-users. Recommendations include extending the framework to incorporate anisotropy and attenuation, integrating with surface-infrastructure monitoring for hazard early warning, expanding to multi-physics data fusion (e.g., gravity, InSAR), and establishing standardized benchmarks and open datasets to foster broader adoption and comparative evaluation within the geoscience community.