What to Expect
Developing reliable and robust surrogate models for uncertainty quantiﬁcation and propagation is a key problem in many scientiﬁc and engineering applications, such as structural health monitoring, natural hazard modeling, and environmental process understanding, to name but a few. In this talk, we address two problems regarding surrogate modeling. First, how to construct a surrogate model when numerical simulators are conﬁgured with a large number of parameters? Inspired by the achievements in computer vision, an end-to-end ﬁeld-to-ﬁeld surrogate is introduced to provide pixel-wise predictions. More speciﬁcally, a multi-level network is designed to directly capture the complex mapping for high-dimensional data without using any explicit dimension reduction method. The second question is how do we objectively measure the eﬀects of the input uncertainties on model predictions? To tackle this problem, hierarchical Bayesian modeling is utilized to describe model uncertainties. A recently introduced variational gradient descent algorithm, stochastic gradient variational Bayes, is scaled to deep neural networks to perform Bayesian inference on millions of model parameters. Various examples are provided to illustrate the concept, methodology, and related algorithms of the proposed modeling techniques.
Luo is a Ph.D. candidate from Department of Civil & Environmental Engineering & Earth Sciences at the University of Notre Dame. He works with the National Academy of Engineering (NAE) member Dr. Ahsan Kareem at the Natural Hazards Modeling Laboratory. Xihaier’s research lies at the intersection of machine learning, computer vision and physical systems. He has mainly worked on data-driven and physics-reinforced deep learning for predictive modeling and uncertainty quantiﬁcation of PDE systems.