Upcoming Event: Jackson School of Geosciences & Oden Institute

Synthetic Geology Meets Deep Learning and Inverse Problems

Eldad Haber, Professor, University of British Columbia

Abstract

Geophysical inversion is a complex challenge that requires both a deep understanding of physical systems and the design of appropriate priors. This task becomes even more demanding due to the scarcity of high-resolution 3D subsurface images needed to construct meaningful priors, as well as the computational challenges of training such priors in 3D.

This talk is structured into three key parts. First, we introduce synthetic geology, a framework for generating realistic 3D geological models that can be leveraged to train powerful priors. Next, we present Multilevel SGD, a novel training algorithm designed to efficiently train large-scale 3D networks with extensive datasets. Finally, we demonstrate the practical applications of these priors, showcasing how they can be utilized to solve inverse problems through flow matching algorithms sampling the posterior and exploring the uncertainty in the inversion.

 

 

Biography

Eldad Haber is a NSERC Industrial Research Chair at the University of British Columbia. Eldad is working in the field of computational inverse problems with applications in machine learning, geosciences and medical imaging. Over the last 20 years, Eldad has written various commercial software packages that have been widely adopted by industry. Eldad has written or co-authored over 150 peer reviewed publications on computational problems and is a U.S. Department of Energy Career Award recipient. After completing his Ph.D, he spent several years as a research scientist with Schlumberger and nine years at Emory University in Atlanta at the Department of Mathematics and Computer Science. In 2011, Eldad co-founded Computational Geosciences Inc and in 2017 he co-founded Xtract.ai. In 2022 Eldad was awarded a SIAM Fellow on his work on computational inverse problems.