Abstract

The design of turbomachinery is typically sequential with gradual increment of increasing fidelity, from preliminary design to identify velocity triangles with a 1D meanline model, to detailed design optimization iterations with a 3D Computational Fluid Dynamics (CFD) solver. Multi-fidelity learning is a machine learning technique that enables training a surrogate model from multiple fidelity of data sources with different accuracy and varying cost of data generation. It can contribute to accelerate the design of turbomachinery, by leveraging information from conventionally distinct design phases into a combined and unified design process. One of the common assumptions in multi-fidelity learning is the presence of homogeneous input domain across all fidelities. However, the different levels of fidelities involved in the design of turbomachinery have inconsistent inputs representations (e.g. 1D chord and aspect ratio for meanline and 3D blade shape for CFD). This paper proposes an approach to extend the capability of multi-fidelity learning towards incorporating input domains which are heterogeneous in their dimensionality across the fidelities. The proposed approach is tested on a transonic airfoil problem where a multi-fidelity surrogate is generated from different discretizations of computational fluid dynamics (CFD) solvers with variable input domains and single target performance target. It is demonstrated to exhibit better predictive accuracy compared to the surrogate learned only from a sparse data generated by the highest discretization.