Enhancing Geophysical Inverse Problems with Physics-informed Machine Learning

Bryan Riel, Associate Professor

School of Earth Sciences, Zhejiang University

2023.06.02（Friday）15:00, Science Building II, 2821

Abstract:

Recent advances in physics-informed machine learning (PIML) have provided new opportunities to study complex geophysical systems within both small- and large-data regimes. PIML utilizes the flexibility of machine learning regression models to combine information extracted from data with information from known physical constraints. These methods can be used to infer large-scale parameter fields for physical systems governed by partial differential equations, such as mantle flow, glacier flow, and groundwater motion. However, it is still unclear how to assess the robustness of the inferred parameter fields in the context of data noise and model uncertainties. Moreover, it is not yet obvious how these methods differ from or are complementary to existing geophysical inverse methods.

Here, I will present a preliminary approach to address these questions through rigorous uncertainty quantification via Bayesian inference. Using the estimation of ice shelf rigidity in Antarctica as a motivating example, I will introduce the use of variational Gaussian Processes to model and sample from probability distributions of ice shelf rigidity, conditional on remote sensing observations. This approach is scalable to large datasets within a PIML framework and allows for stochastic predictions of future sea level rise. I will then discuss how recent developments in deep generative modeling and differentiable programming can be combined to efficiently perform Bayesian inference within the classical framework of inverse modeling. I show that, in the context of inferring ice shelf rigidity, the two approaches are complementary and may be combined to reduce overall uncertainties in rigidity values. Thus, PIML combined with Bayesian inference can be a powerful new tool for geophysical inverse problems.

Bio:

Bryan Riel is an Associate Professor at the School of Earth Sciences at Zhejiang University. He obtained a B.S. and M.S. in Aerospace Engineering at the University of Texas at Austin. Afterwards, he obtained an M.S. and Ph.D. in geophysics at the California Institute of Technology. Following his graduate studies, he worked at the NASA Jet Propulsion Laboratory as a signal analysis engineer, primarily working on synthetic aperture radar (SAR) development. He then joined the Massachusetts Institute of Technology as a research scientist to study glacier dynamics using remote sensing. He joined Zhejiang University in 2022. Bryan's research is broadly focused on utilizing large-scale geodetic data to extract meaningful insights into the behaviors and properties of physical systems. These systems include glaciers, ice sheets, groundwater systems, volcanic systems, and active faults. He is also interested in utilizing scientific machine learning to address questions in geophysical inverse problems and uncertainty quantification.