Open Collections

Xu, Anran

University of British Columbia

Artificial intelligence (AI) has become a driving force for innovation, and Canada has been at the forefront of this movement. One area where AI shows great promise is the earth sciences, particularly in understanding the subsurface of our planet, which is crucial for mineral exploration, groundwater studies, and more. Exploring the subsurface relies on geophysical surveys that measure data sensitive to the Earth’s physical properties. Geophysical inversions are crucial for understanding subsurface structures by converting measurements into models of subsurface physical properties. However, geophysical inverse problems are ill-posed; therefore, inverse modeling focuses on designing a regularization to recover geologically plausible models. Conventional regulation typically penalizes spatial variations and/or promotes proximity to a reference model. Supervised machine learning approaches often rely on the availability of extensive training datasets, which are limited in geophysics. This research focuses on leveraging machine learning (ML) methods in a test-time learning manner to improve geophysical inversions. The research explores the potential of implicit regularization effects inherent in certain ML models, which do not require training datasets, so these methods are test-time learning methods. These effects have shown promise in various inverse problems in fields like computer vision, computer graphics, and biomedical imaging. For instance, methods like Deep Image Prior (DIP) and Neural Fields (NFs), which are categorized as test-time learning methods, have been effective in tasks such as de-noising and 3D rendering without prior training. The hypothesis is that these implicit regularization effects can also improve geophysical inversions. My research involves applying multiple test-time learning methods to these inversions and lets us re-think how we can construct regularization in an inverse problem. The goal is to develop better subsurface models by integrating implicit regularizations. In this research, I mainly test the test-time learning methods on the Direct Current (DC) resistivity inverse problem, but it can be applied to other geophysical methods. While the primary application is in mineral exploration, the research has broader implications for groundwater studies, environmental investigations, and oil and gas exploration, potentially benefiting these fields by improving the quality of subsurface models.

Text

2024-10-22

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/89472

Geophysics

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/