LANS Seminar

Seminar Title: Parametric Sensitivity of Ocean Modeling through Neural Surrogates

Speaker: Yixuan Sun is a postdoctoral appointee at the Mathematics and Computer Science (MCS) division at Argonne National Laboratory

Date/Time: Thursday, February 27, 2025/ 2:30 PM – 3:30 PM (In-Person)
Location: See Meeting URL on the cels-seminars website which will require an Argonne login.

Description: Ocean processes regulate climate by absorbing heat, transporting salt, and sequestering carbon dioxide. Simulating these processes requires parameterizations for unresolved dynamics. However, the uncertainties in these parameterizations often lead to discrepancies with observations. Traditional sensitivity analysis methods, like adjoint state techniques, are computationally expensive for complex models. In this talk, I will present our work on developing differentiable neural network surrogates for an idealized wind-driven baroclinic double-gyre setting. Our approach leverages the efficient computation of the surrogate’s Jacobian as the sensitivity measure of model outputs to various parameterizations. We also investigate the influence of surrogate hyperparameters using large-scale multi-objective Bayesian optimization for streamlined surrogate modeling. Finally, I will introduce our current and future efforts involving a novel local differential consistency training strategy and graph-based learning for more accurate and reliable sensitivity estimates. This framework is expected to offer a pathway of AI-driven sensitivity analysis for complex systems.

Bio: Yixuan Sun is a postdoctoral appointee at the Mathematics and Computer Science (MCS) division at Argonne National Laboratory. He received his B.S. degree from the School of Energy and Power Engineering at Shandong University, China, and his M.S. and Ph.D. degrees from the School of Mechanical Engineering at Purdue University, USA. He was a Givens Associate at MCS in 2020 and 2021 prior to his current position. His current research interests include physics-guided machine learning and composable generative modeling for complex dynamical systems.

Please note that the meeting URL for this event can be seen on the cels-seminars website which requires an Argonne login.