Alternative Cokriging Model for Variable-Fidelity Surrogate Modeling

摘要

In this article, an alternative cokriging method for variable-fidelity surrogate modeling is proposed. This cokriging formulation is obtained by redefining the cokriging weights, which allows for factoring out the process variances from the cokriging covariance matrix. Additionally, a scaling factor σ_1t/σ_2 was introduced in (he cokriging predictor to account for the influence of the low-fidelity data on the prediction of the high-fidelity function, and a closed-form solution for its maximum likelihood estimation was given. The proposed cokriging predictor was then demonstrated for analytical examples and an aerodynamic example of predicting the aerodynamic coefficients as well as the drag polar of an RAE 2822 airfoil. The examples show that the proposed cokriging method can potentially be applied to efficient aerodynamic data production as well as shape optimization or any other research areas where computer codes of varying fidelity are in use.