skip to main content

RSI Research Seminar

Monday, April 17, 2023
12:00pm to 1:00pm
Add to Cal
Jorgensen 109
ML-enabled model of underground fluid transport in the presence of chemical reactions
Mina Karimi, Postdoctoral Scholar Research Associate in Mechanical and Civil Engineering,

Join us every other Monday at noon for lunch and a 30-minute research talk, presented by Resnick Sustainability Institute Graduate Fellows and Caltech researchers funded by the Resnick Sustainability Institute. To see the full schedule of speakers, visit the RSI Research Seminar web page. Seminars currently take place in a hybrid format, both in-person (Jorgensen building first-floor conference room) and via Zoom. For more information and to get the Zoom login info, please reach out to ramonae@caltech.edu

Topic: ML-enabled model of underground fluid transport in the presence of chemical reactions

Abstract: The transport of water through permeable geological formations couples various phenomena. As the water flows through the permeable medium, water reacts with the medium changing the morphology, mechanical properties, and permeability of the medium; this in turn affects the flow and chemistry. These varied phenomena occur at various length and time scales, which makes the problem extremely difficult. Multiscale modeling approaches have been developed for particular processes to pass the information from one scale to another. However, the actual implementation of this technique is prohibitively expensive. We present a methodology to overcome this challenge by creating a high-fidelity, computationally efficient surrogate of the lower scales behavior that can directly be used at the upper scale to prevent repeatedly solving equations. We particularly focus on the transport of flow through the porous medium in the presence of chemical reactions, where due to the change of internal variables and evolving microstructure, the convection-diffusion equations are complicated. We generate one-time off-line data from the lower scale to train the solution to the partial differential equations over a neural network and obtain a learned surrogate. The surrogate is an inexpensive, approximate solution to the lower-scale problem, which can be used to solve the macroscopic problem without further modeling.