Numerical investigation for erratic behavior of Kriging surrogate model

Abstract

Kriging model is one of popular spatial/temporal interpolation models in engineering field since it could reduce the time resources for the expensive analysis. But generation of the Kriging model is hardly a sinecure because internal semi-variogram structure of the Kriging often reveals numerically unstable or erratic behaviors. In present study, the issues in the maximum likelihood estimation which are the vital-parts of the construction of the Kriging model, is investigated. These issues are divided into two aspects; Issue I is for the erratic response of likelihood function itself, and Issue II is for numerically unstable behaviors in the correlation matrix. For both issues, studies for specific circumstances which might raise the issue, and the reason of that are conducted. Some practical ways further are suggested to cope with them. Furthermore, the issue is studied for practical problem; aerodynamic performance coefficients of two-dimensional airfoil predicted by CFD analysis. Result shows that such erratic behavior of Kriging surrogate model can be effectively resolved by proposed solution. In conclusion, it is expected this paper could be helpful to prevent such an erratic and unstable behavior.