Multi-Fidelity Surrogate Models for Computational Aerodynamics
Many types of aerodynamic analyses rely on surrogate models to provide a continuous representation of discrete measurements obtained by simulation or experiment. Generating high-fidelity datasets can be prohibitively expensive and degrade surrogate model performance in the absence of sufficient training data. Multi-fidelity models that use Gaussian process regression can enhance the cost-accuracy tradeoff when generating information used for decision-making in DoD aerodynamics programs. The adoption of multi-fidelity modeling is hindered by the availability of a unified software framework capable of handling the complexities of both generating data from multiple computational models and training surrogates. This presentation will highlight advancements made towards generalizing multi-fidelity modeling capabilities intended for HPCMP CREATE Sage. Two relevant aircraft geometries, a generic fighter jet and hypersonic waverider, will be used to demonstrate various computational aerodynamics models from HPCMP CREATE Kestrel, NASA Cart3D and NASA CBAero that can be used to train multi-fidelity surrogates.
IMPACT
Accomplishment: Demonstrated the capability to create a unified multi-fidelity framework in HPCMP CREATE Sage that can perform the entire multi-fidelity workflow. Result: Significantly reduces the effort required to perform a multi-fidelity analysis through software automation, abstraction and systematic definition of the computational models used. Multi-fidelity models can achieve greater than 2x accuracy compared to high-fidelity counterparts at the same computational expense depending on the application.
PRESENTER
Simin, Andrew
andrew.simin@gdit.com
667-228-5507
HPCMP PET / GDIT
CO-AUTHOR(S)
Boyer, Mathew
mathew.boyer@gdit.com
CATEGORY
Surrogate Modeling for HPC
SECONDARY CATEGORY
Comp Fluid Dynamics
SYSTEM(S) USED
Raider