Upcoming Event: Babuška Forum

Scientific Machine Learning for Surrogate Modeling, Parameter Identification and Transfer Learning of Multiphase Flows in Porous Media

Hannah Lu, Assistant Professor, Aerospace Engineering and Engineering Mechanics

Abstract

Scientific machine learning—a modeling approach in which a deep neural network representation of a process is informed by both data and physics (usually in the form of PDEs)—has received increased attention in the field of geophysics, from subsurface flows to
earthquake forecasting and coupled thermo-hydro-mechanical processes.

Here, we develop scientific machine learning approaches to build surrogate models of CO2 injection and migration in porous media in realistic geologic settings. In contrast with most explorations of ML techniques, here we rely on intermediate-scale physical experiments
conducted in laboratory. We explore and compare different ML approaches in their ability to robustly: (1) emulate, via training with full-physics simulations, the complex fluid flow dynamics of the experiments (which feature multiphase flow migration, trapping and convective dissolution) in terms of the full state variables (pressure, CO2 saturation and CO2 concentration in brine); (2) determine the sensitivity of quantities of interest (such as CO2 leakage rate, leakage volume, and amount of dissolved CO2) to the parameters of the physics-based model (such as permeability and capillary entry pressure of the geologic layers, and injection rate and duration); (3) solve the inverse problem to identify the underlying parameters in the ground-truth physical experiments; and (4) transfer learning from the detailed surrogate model building in one geometry of the geologic layers to a different geometry and injection location.

We comment on the power, current limitations, and opportunities of machine learning approaches for modeling, forecasting and uncertainty quantification of subsurface CO2 sequestration and energy storage.

Biography

Hannah Lu is a tenure-track assistant professor at UT-Austin. Before joining UT, she was a postdoc associate at MIT, affiliated with the Department of Aeronautics and Astronautics, Department of Civil Environmental Engineering, Earth Resources Laboratory and Laboratory for Information and Decision Systems. She obtained her Ph.D. from Energy Science and Engineering at Stanford Doerr School of Sustainability. Her research interests lie in the field of scientific computing, reduced order modeling, uncertainty quantification and machine learning in applications of environmental fluid mechanics.