Ambient seismic noise tomography of the southern United States and seismic inversion with dictionary learning using unsupervised machine learning.

Seismic tomography and inversion are well-established methods to understand deep subsurface structures. These techniques are particularly effective in delineating complex crustal architectures, which often deviate from simplified textbook models. High-resolution seismic tomography enables the interpretation of such intricate geological features. In the context of the Alleghenian orogeny, understanding Laurentia’s tectonic evolution depends on unraveling the structural relationships between the Suwannee terrane and adjacent crustal blocks—spanning from Texas in the west to Georgia and Florida in the east. This study is presented in three chapters. In Chapter 1, our goal is to identify the suture zone between the Gondwanan Suwannee terrane, adjacent terranes, and the Laurentian continental margin by developing an accurate shear wave velocity model. This helps improve our understanding of the lithology and formation of the southeastern U.S., including the Florida Peninsula. Since the region is seismically quiet, we use vertical component data from over 280 broadband stations (USArray, Transportable Array, SESAME Array, and regional networks) to estimate the Empirical Green’s Function (EGF) through cross-correlation and cross-coherence. This mimics Rayleigh wave phase characteristics. We address the problem of non-stationary phases in Rayleigh wave extraction from ambient noise. Seismic interferometry assumes stationary phase dominance in EGF estimation, but real-world, non-uniform source distributions introduce inaccuracies. To mitigate this, we use cross-coherence to improve signal-to-noise ratio and apply double-beamforming to accept only contributions aligned with the Great Circle Path (GCP) between station pairs. Group velocity dispersion curves of Rayleigh waves are used to estimate 1D shear wave velocity models via Markov Chain Monte Carlo sampling, revealing tectonic boundaries. In the next chapter, we apply the Double-sided Optimized Multiple Signal Characterization (DOPMUSIC) algorithm to estimate EGFs by modeling significant noise sources across frequency bands along the GCP. In Chapter 4, to address the computational intensity of seismic inversion for thin-layered structures, we incorporate unsupervised machine learning. CNN and U-Net decompose seismic traces into dictionary and coefficients, reconstruct reflectivity, and convolve it with a wavelet. Lasso regularization aids training. We also use Variational Autoencoders (VAEs) for efficient high-frequency reflectivity recovery. Seismic examples validate our approach.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Doctoral

Grantor

Baylor University.

Year dc:date.issued

2025

Author dc:creator

• Barman, Debajeet, 1988-

Advisors dc:contributor.advisor

• Pulliam, Jay.
• Sen, Mrinal K.

• Tomography.
• Inversion.
• Deep subsurface structures.
• Crustal structure.
• Lithospheric structure.
• Empirical Green's function (EFGs)
• Broadband seismic stations.
• Great circle path (GCP)
• Shear wave velocity models.
• Double-sided optimized multiple signal characterization (DOPMUSIC) algorithm.

Statement dc:rights

• Baylor University works are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission. Contact libraryquestions@baylor.edu for inquiries about permission.

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/2104/13809

OAI identifier oai:identifier

oai:baylor-ir.tdl.org:2104/13809