Seminar 2023-10-27

A Novel Extreme Value Autoencoder Framework for Probabilistic Model Emulation and Calibration 

Likun Zhang

Assistant Professor

Department of Statistics

University of Missouri

Date: Friday, October 27, 2023

Time: 11:00 A.M. – 12:00 P.M. Eastern Time

Location: Johnson Center, Room 325 Meeting Room A

Large physics-based simulation models are crucial for understanding complex problems related to energy and the environment. These models are typically quite computationally expensive and there are numerous computational and uncertainty quantification (UQ) challenges when using these models in the context of calibration, inverse problems, UQ for forward simulations, and model parameterization. Surrogate model emulators have proven to be useful in recent years to facilitate UQ in these contexts, particularly when combined with Bayesian inference. However, traditional methods for model emulation such as Gaussian processes, polynomial chaos expansions, and more recently, neural networks and generative models do not naturally accommodate extreme values, which are increasingly relevant for many complex processes such as environmental impacts due to climate change and anomaly detection. Many statistical methods have been developed to flexibly model the simultaneous occurrences of extremal events, but most of them assume that the dependence structure of concurrent extremes is time invariant, which is unrealistic for physical processes that exhibit diffusive dynamics at short-time scales. We propose to develop a novel probabilistic statistical framework to explicitly accommodate concurrent and dependent extremes within a conditional variational autoencoder (CVAE) engine for enabling fast and efficient uncertainty quantification in model calibration, inverse modeling, ensemble prediction, and parameter estimation contexts. We also propose a new validation framework that is tailored to assess skill in fitting extreme behavior in model outputs. Our approach addresses, for the first time, the need to have efficient surrogate emulators of expensive simulation models that can accurately characterize, in a rigorous probabilistic manner, extreme values that are dependent in space and time and across processes.

Dr. Zhang is currently an assistant professor in the Department of Statistics at the University of Missouri. He received his Ph.D. degree in Statistics from Penn State in 2020, after which he worked with climate scientists at Lawrence Berkeley National Laboratory for two years. His research focuses on extreme value theory and flexible spatial extremes modeling, which has been used to study a variety of weather processes and to detect changes in their long-term climatology. He also incorporates deep learning techniques in spatial extremes modeling so domain scientists can study the dependent extremes on datasets with massive number of locations. 

David Kepplinger

Nicholas Rios