Uncertainty Quantification (UQ) is a computationally expensive process requiring a large number of system evaluations with identified random variables. Multidisciplinary Design Optimization (MDO) is also a computationally demanding process, requiring iterative system evaluations with design variables of interest. Surrogate models can alleviate the computational burden in both MDO and 1 Q activities; however, the accuracy of interpolated system responses may deteriorate. In this paper, the Non-Deterministic Kriging (NDK) method is utilized to construct a surrogate model that alleviates numerical instabilities inherent to conventional deterministic kriging such as overfitting. NDK captures epistemic and aleatory uncertainties separately for uncorrelated and correlated stochastic variables, producing more physically meaningful predictions. This study introduces the incorporation of correlation length estimations into NDK and investigates the convergence rates of the NDK framework's mean and variance estimations for independent random variables using a practical engineering example. First, the fundamental numerical behavior of NDK is discussed using two mathematical examples. Then, a three-input Numerical Propulsion System Simulation (NPSS) UQ problem is analyzed.
展开▼
