Yunlin Zeng - Thesis Dissertation Defense

Bayesian Inference for Physical Inverse Problems: Amortized Simulation-Based Inference in Seismic Imaging

In partial fulfillment of the requirements for the degree of 

Doctor of Philosophy in Physics 

 

School of Physics Dissertation Defense

 

Yunlin Zeng
Dr. Felix Herrmann, School of Computational Science and Engineering, School of Earth and Atmospheric Sciences, and School of Electrical and Computer Engineering (Advisor)

 

Bayesian Inference for Physical Inverse Problems: Amortized Simulation-Based Inference in Seismic Imaging

Virtual Link: https://teams.microsoft.com/meet/2135859065505996?p=0IRPGzEwn5Dj9Upy1y

 

Thesis Committee:

Dr. Tamara Bogdanovic, School of Physics, Georgia Institute of Technology

Dr. James C. Gumbart, School of Physics, Georgia Institute of Technology

Dr. Gongjie Li, School of Physics, Georgia Institute of Technology

Dr. Ignacio Taboada, School of Physics, Georgia Institute of Technology

 

Abstract:

In geophysical exploration, sources at the surface send waves into the ground, and receivers record the echoes that come back. Turning these recordings into a picture of the subsurface is hard. The computation is expensive, and many different Earth models can fit the same data. Decisions such as where to drill or where to store carbon dioxide need more than one best picture. They also need to know how uncertain that picture is. Standard methods measure this uncertainty by repeating costly wave simulations many times for every new dataset, which is too slow for realistic problems.

 

This work uses amortized inference instead. A neural network is trained ahead of time on many simulated pairs of Earth models and their seismic data. The network is a generative model that learns to turn new seismic data into many plausible Earth models within seconds, and the spread among them shows where the answer is uncertain. This work uses two such networks. The first is a conditional normalizing flow, which estimates the speed of seismic waves underground. Training it with many possible starting guesses of this wave speed, instead of one, makes its answers depend less on that guess, which is never known exactly in practice. It also makes its uncertainty estimates more reliable. The second is a diffusion model, which builds a sample by removing noise step by step. It recovers two seismic properties, wave speed and acoustic impedance, at once, along with their uncertainty. Tests on a realistic synthetic benchmark based on North Sea geology show that these methods give reliable images of the subsurface with uncertainty estimates.

Event Details

Date/Time:

  • Date: 
    Friday, December 4, 2026 - 3:00pm to 4:00pm

Location:
Howey N201/202