School of Physics Thesis Proposal Exam 

Presenter:        Yunlin Zeng

Title:                   Enhancing Full-Waveform Variational Inference through Stochastic Resampling and Data Augmentation

Date:                  Monday, July 8, 2024

Time:                  9:00 a.m.

Place:                Howey N201/202

Virtual Link:     https://gatech.zoom.us/j/96158574435  / Meeting ID: 961 5857 4435

Committee:     Dr. Felix Herrmann,  School of Earth & Atmospheric Sciences, Georgia Institute of Technology (advisor)

Dr. Tamara Bogdanovic, 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:

Recent developments in simulation-based inference [Cranmer et al. 2020], like the full-Waveform variational Inference via Subsurface Extensions (WISE, Yin et al. [2023]), enable rapid online estimation of subsurface velocities by leveraging pre-trained models. To achieve this, WISE employs subsurface-offset Common Image Gathers (CIGs) to convert shot data into physics-informed summary statistics. While CIGs effectively retain critical information even when initial velocity estimates are inaccurate, WISE’s performance remains dependent on these initial velocity model assumptions. In this work, we explore multiple strategies to reduce the dependency on the initial velocity and improve the quality of posterior samples. These methods include the stochastic resampling and data augmentation of migration-velocity models, the use of multiple sets of compressed statistics, and the reuse of latent variables saved in a trained conditional normalizing flow network.