High-Resolution 3D Seismic Inversion for Sub-basement Fault Imaging Using Machine Learning-Driven Regularization

Project Abstract

This study introduces a novel framework for high-resolution 3D seismic inversion aimed at sub-basement fault imaging, leveraging machine learning–driven regularization to enhance subsurface delineation in complex geological settings. We address the persistent challenge of resolving small-scale fault networks and heterogeneities beneath thick sedimentary sections where conventional inversion methods suffer from trade-offs between resolution, noise amplification, and computational feasibility. Our approach fuses physics-informed inversion with data-driven priors derived from deep learning models trained on multi-parameter datasets, including rock physics constraints, seismic impedance, and well-log calibrations. The core methodology integrates a forward seismic simulator with an end-to-end learning module that adaptively regularizes the inverse problem through learned priors and sparsity-enforcing penalties. Specifically, we develop a tiered regularization strategy that combines (i) a spectral sparsity constraint to promote compact fault feature representation in the wavenumber domain, (ii) a learned prior on geological facies and fault geometry obtained from a convolutional neural network trained on synthetic and field-derived examples, and (iii) a physics-based constraint enforcing consistency with available well-log and crosshole datasets. The model is trained to minimize a composite loss that balances data misfit with these priors, while incorporating variance-based uncertainty quantification through a Bayesian framework and ensemble perturbations. A multi-scale inversion workflow is implemented to progressively recover large-scale stratigraphic architecture first, followed by targeted high-resolution updates in fault zones, thereby stabilizing convergence and reducing computational burden. We validate the framework on synthetic models featuring layered sediments overlying complex fault networks and on field datasets from mature hydrocarbon and CO2 storage environments with known sub-basement discontinuities. The results demonstrate substantial improvements in fault detection sensitivity, continuity, and accuracy of lateral and vertical dip estimation, compared with traditional Tikhonov-regularized and total-variation inversions. Notably, the machine learning–driven regularization demonstrates robustness to noise, multiple reflections, and model misspecifications, enabling reliable imaging of fault cores and fracture networks at depths where data quality is typically degraded. We further perform a rigorous uncertainty analysis to quantify confidence in fault delineation and to identify regions where additional data acquisition or borehole calibration would most effectively reduce ambiguity. The implications of this work extend to hazard assessment, reservoir characterization, and sub-surface risk mitigation by providing high-fidelity vertical and horizontal fault maps and associated property models. The methodology is generalizable to other geophysical modalities, enabling integrated, uncertainty-aware subsurface characterization under challenging data conditions. Overall, the study contributes a scalable, interpretable, and physics-consistent machine learning paradigm for high-resolution 3D seismic inversion in complex sub-basement environments.

Project Overview

What This Project Is About

---

This project explores how to create detailed 3D images of faults beneath the earth’s surface using seismic data. It combines traditional seismic inversion with smart computer techniques (machine learning) to improve the clarity and accuracy of fault images, especially in deep or complex areas.

---

---

The Problem It Addresses

---

In geophysics, turning seismic recordings into clear pictures of underground faults is hard. Standard methods can be slow or produce blurry results when data are noisy or the subsurface is complicated. This project aims to make fault images sharper and more reliable, helping scientists and engineers assess risks and plan activities safely.

---

---

Objectives of the Project

---

1. Learn how seismic data are collected and processed to create images of the subsurface.
2. Implement a 3D seismic inversion workflow with regularization to stabilize results.
3. Incorporate machine learning to guide the inversion toward clearer fault features.
4. Evaluate the method on synthetic and real datasets to demonstrate improvements.
5. Assess computational requirements and practical feasibility for field-scale applications.

---

---

What You Will Do Step by Step

---

1. Review basic seismic concepts and data formats.
2. Set up a simple 3D inversion framework and define regularization terms.
3. Train or apply a machine learning model to assist inversion parameters or feature extraction.
4. Test on simulated data with known faults to measure accuracy.
5. Apply to real seismic data and compare results with conventional methods.
6. Analyze sensitivity to noise, data coverage, and model assumptions.
7. Document procedures, results, and limitations for reproducibility.

---

---

Expected Outcome

Clearer 3D images of sub-basement faults with quantified uncertainty, along with a practical workflow that researchers can adapt to similar subsurface imaging tasks. The project should show tangible improvements over standard methods and provide guidance for future work.