AME Seminar: Kevin Quinlan
Thursday, March 3, 2022, 4:00 p.m.
Kevin R. Quinlan
Applied Statistician
Lawrence Livermore National Laboratory
"Active Learning for Multi-Fidelity Aerodynamic Databases of Hypersonic Design Spaces"
AME Lecture Hall, Room S212
Zoom Link | Password: 2022
Abstract
Predicting flight characteristics of a hypersonic flight vehicle requires aerodynamic databases that cover a wide design space, across multiple dimensions and wide dynamic ranges within those dimensions. In developing these databases, it becomes critical to conserve limited computational resources while ensuring quantities of interest are well resolved. Efficient generation of databases is especially useful in the early stages of development where designs may be subject to change. To achieve this goal, we present a multi-fidelity cokriging model with an associated active learning approach. In this multi-fidelity framework, we utilize fast flow solvers which can be densely sampled to improve prediction using a limited number of samples from a more accurate but more computationally intensive flow solver. Using fewer expensive flow solver runs, we can augment the response surface to capture the physics that were simplified in the lower fidelities. Our novel approach to adding additional samples accounts for both model uncertainty as well as the practical state space of the vehicle, as determined by forward modeling within the active learning loop. The combination of these two methodologies can greatly reduce the computational efforts allowing for a faster turn-around time.
This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. LLNL-ABS-831940.
Bio
Kevin Quinlan is a staff member in the Applied Statistics Group at Lawrence Livermore National Laboratory. Previously, he completed his Ph.D. in statistics at Penn State. His main research interests are in the design of experiments, specifically computer experiments, and Gaussian process modeling.