08-27-2019 | Xun Huan: 122nd NIA CFD Seminar: Uncertainty Quantification via Optimal Experimental Design and Bayesian Neural Networks for Aerospace Applications

122nd NIA CFD Seminar: Uncertainty Quantification via Optimal Experimental Design and Bayesian Neural Networks for Aerospace Applications

Date: Tuesday, August 27, 2019
Time: 11:00 a.m.-Noon (EDT) 
Room: NIA, Rm137
Speakers: Xun Huan

Link: http://nia-mediasite.nianet.org/NIAMediasite100/Catalog/Full/fe54023273ef446084620d8a1a25ea5821

Speakers Bio:
Dr. Xun Huan is an Assistant Professor of Mechanical Engineering at the University of Michigan—Ann Arbor, and affiliated faculty to the Michigan Institute for Computational Discovery & Engineering (MICDE), the Michigan Institute for Data Science (MIDAS), and the U-M Applied Physics Program. Dr. Huan received a Ph.D. in Computational Science and Engineering from MIT Department of Aeronautics and Astronautics and was a postdoctoral researcher at Sandia National Laboratories in Livermore, California. His research broadly revolves around uncertainty quantification, machine learning, and numerical optimization, with a focus on aerospace and mechanical engineering applications. His current projects involve optimal experimental design for identifying and acquiring the most useful data, and methods for quantifying uncertainty and trust for machine learning models. Outside work, Dr. Huan is passionate about aviation and holds a private pilot certificate.

Abstract:
Models and data are two pillars of scientific research: models make predictions, and data help calibrate existing models and develop new ones. In this talk, we focus on two important interactions between models and data: (1) optimal experimental design (OED) for identifying the most useful data, and (2) Bayesian neural networks (BNNs)—a class of data-driven models with quantified uncertainty—for accelerating expensive predictions.
First, we present the OED framework that systematically quantifies and maximizes the value of experiments. Indeed, some experiments can produce more useful data than others, and well-chosen experiments can lead to substantial savings. We describe a general mathematical framework that accommodates nonlinear and computationally intensive (e.g., ODE- and PDE-based) models. The formalism employs Bayesian statistics and an information-theoretic objective, and we develop tractable numerical methods with demonstrations on designing combustion kinetic experiments and sensor placement for contaminant source inversion.
Next, we introduce BNNs as data-driven surrogate models. The capability of rapid predictions with quantified uncertainty makes them excellent tools for supporting real-time, high-consequence decision-making. In a proof-of-concept application on in-flight detection of rotorcraft blade icing, we build a BNN from a database of SU2 simulations to directly map noise signals to aerodynamic performance metrics, thus bypassing the expensive inverse problems. The BNN is able to produce a distribution of predictions instead of a single-value output, reflecting the quality and confidence of the machine learning model, and offering valuable information for pilot decision-making.
Lastly, we present preliminary results on machine learning reconstruction of turbulent fluctuations for stochastic noise generation in RANS-based aeroacoustic simulations using SU2, where in particular we investigate a data-driven characterization of turbulence energy spectrum from LES computations.