Adaptive Multi-Fidelity Sampling Using HPCMP CREATE™ Sage

The applications of surrogate models span a wide range of research fields where fundamentally some set of parametrizable inputs are predicting some quantifiable output. Surrogate models particularly shine when high-frequency evaluation is desired. In many instances, the computational cost of high-fidelity physics-based data generation for the creation of high-fidelity parametric surrogate models can be extreme. Using multi-fidelity models rather than solely high-fidelity models can dramatically reduce overall cost of training but maintain accuracy comparable to models that exclusively use high-fidelity data. Additionally, intelligent sampling algorithms that consider sampling costs across the multi-fidelity data sources can further reduce costs and increase efficiency. The DoD HPCMP CREATE™ Sage surrogate modeling framework is utilized for multi-fidelity model creation and adaptive model refinement. The model archetype available in Sage that this paper utilizes is scalar-to-field, where a set of scalar input conditions are used to predict values over a field in some solution domain. Specifically, the surface temperature and pressure on a generic waverider are predicted as a function of the freestream Mach Number and angle of attack. CREATE AV Kestrel is used with different air models to represent different fidelities, perfect gas and 5-species air.