Author ORCID Identifier
Semester
Spring
Date of Graduation
2024
Document Type
Dissertation
Degree Type
PhD
College
Eberly College of Arts and Sciences
Department
Geology and Geography
Committee Chair
Dengliang Gao
Committee Member
Jaime Toro
Committee Member
Timothy Carr
Committee Member
Aaron Maxwell
Committee Member
Haibin Di
Abstract
Geophysical data are used to “remotely sense” the subsurface to illuminate structures, stratigraphy, and features that are of interest in the exploration for energy, suitable storage for carbon dioxide, geothermal development, and environmental considerations. The interpretations of these data typically are regarded as partially an art, as it depends on the experience of the interpreter. This is further exacerbated with the big data era as this data is large and takes several days, even longer, to manually interpret. To bridge this gap of speed and subjective bias, machine learning has been successfully applied to automate the geophysical interpretation process. This study improves the application of machine learning systems for practical field deployment through the proposal of novel workflows. This study also documents the first application of diffusion models to solve seismic inverse problems. Traditional seismic inversion lacks a unique solution and suffers from cycle skipping. Here, I propose a new data-driven framework for the direct inversion of velocity from seismic amplitudes, potentially overcoming the limitations and shortcomings associated with traditional methods.
In the second chapter, I propose a novel workflow called SeisSegDiff that utilizes the semantic representations learned by a diffusion model to aid in seismic facies classification with limited training data. I demonstrate that with as little as 5 training cross-sections, the seismic facies can be accurately classified even for facies with limited occurrence. The third chapter investigates the nature of these learned representations and uncovers the optimal model specification for SeisSegDiff. This chapter also discusses the practical usage of this model for field applications. The fourth chapter reformulates the diffusion model used for generative tasks to predict the subsurface velocity from pre-stack seismic data. I demonstrate that this low-resolution velocity prediction from the diffusion model can aid in accelerating classical full waveform inversion. In the fifth chapter, I propose a committee machine that aggregates the gradient boosting, support vector machine, and neural network to improve prediction of the brittleness estimate for mudstone.
Recommended Citation
Ore, Tobi Micheal, "Advances in machine learning aided seismic interpretation and inversion for subsurface characterization" (2024). Graduate Theses, Dissertations, and Problem Reports. 12440.
https://researchrepository.wvu.edu/etd/12440
Embargo Reason
Publication Pending