WITHDRAWN Auto-linear: A Self-supervised Framework for Robust Subsurface Imaging Through Latent Space Correlations
WITHDRAWN Subsurface imaging, central to geophysics, involves solving Full-Waveform Inversion (FWI) to predict geophysical properties from measurements. Conventional methods rely on encoder-decoder networks trained with paired data from two domains: geophysical properties and measurements. Building on recent work (InvLINT), we demonstrate that only the linear mapping between latent spaces requires paired data, while the encoder and decoder can be trained independently through self-supervised learning.
This discovery, termed Auto-Linear, reveals that self-learned features from two domains are inherently linearly correlated. The Auto-Linear framework offers key advantages: (a) simultaneous forward and inverse modeling, (b) applicability across diverse subsurface imaging tasks with superior performance, (c) robustness with limited paired data and noisy inputs, and (d) strong generalization across datasets. These findings advance data-driven subsurface imaging and open new avenues for geophysical modeling.
Session: Scientific Machine Learning for Forward and Inverse Wave Equation Problems [Poster]
Type: Poster
Room: Exhibit Hall
Date: 4/15/2025
Presentation Time: 08:00 AM (local time)
Presenting Author: Youzuo Lin
Student Presenter: No
Invited Presentation:
Poster Number: 137
Additional Authors
Yinan Feng ynf@unc.edu University of North Carolina at Chapel Hill |
Yinpeng Chen Yinpengchen.work@gmail.com Google DeepMind |
Youzuo Lin Presenting Author Corresponding Author yzlin@unc.edu University of North Carolina at Chapel Hill |
|
|
|
|
|
|
WITHDRAWN Auto-linear: A Self-supervised Framework for Robust Subsurface Imaging Through Latent Space Correlations
Category
Scientific Machine Learning for Forward and Inverse Wave Equation Problems
Description