Local maximum-entropy based surrogate model and its application to structural reliability analysis

A novel surrogate model based on the Local Maximum-Entropy (LME) approximation is proposed in this paper. By varying the degrees of locality, the LME-based surrogate model is constructed according to the local behavior of the response function at the prediction points. The proposed method combines the advantages of both local and global approximation schemes. The robustness and effectiveness of the model are systematically investigated by comparing with the conventional surrogate models (such as Polynomial regression, Radial basis function, and Kriging model) in three types of test problems. In addition, the performance of the LME-based surrogate model is evaluated by an industry case of turbine disk reliability analysis (TDRA) involving random geometric parameters. In TDRA, two LME-based surrogate models are built including a 1 (s t) surrogate model employed in the sensitivity analysis to determine the key random variables and a 2 (n d) surrogate model utilized in Monte-Carlo Simulations (MCS) to predict the Low Cycle Fatigue (LCF) life of turbine disks. Finally, a model-based Uncertainty Quantification (UQ) analysis is performed to rigorously quantify the uncertainties of the physical system and fidelity of surrogate model predictions simultaneously. Results show that the LME-based surrogate model can achieve a desirable level of accuracy and robustness with reduced number of sample points, which indicates the proposed method possess the potential for approximating highly nonlinear limit state functions and applicable for structural reliability analysis.