An efficient surrogate model construction strategy for large-scale output problems

Applying common surrogate models to problems have numerous output variables is computational expensive, since the number of surrogate models should be constructed equals to the number of output variables. This paper presents an efficient strategy to solve this problem. For that, snapshot Proper Orthogonal Decomposition (POD) is used to extract a few main basis modes from certain number of samples. The predicted result of a large-scale output problem comes from the linear superposition of these basis modes. Common surrogate models just need to predict the coefficients for these basis modes. Through this strategy, The Mach numbers at 36864 points around an airfoil are predicted by just constructing 12 kriging surrogate models. The predicted Mach number distributions fit with the CFD results very well, that proves the efficiency of this strategy.