工学的実問題における大域的感度解析手法の適用性評価

抄録

This study explores the applicability of Global Sensitivity Analysis (GSA) methods to practical engineering applications in terms of computational cost and accuracy. We focus on four GSA methods: (1) ANalysis Of VAriance (ANOVA) based on an orthogonal-array in a Design of Experiment (DoE), (2) the Morris method, (3) a Monte Carlo (MC)-based Sobol' method, and (4) a Polynomial Chaos Expansion (PCE)-based Sobol' method. The test cases are the Ishigami function, which exhibits strong nonlinearity and interaction effects, and a 1D-CAE model of a laboratory-scale experimental system that represents a spacecraft propulsion system. The ANOVA approach is suitable for identifying the most influential variable with lower computational costs. The Morris method and PCE-Sobol' method allow one to prioritize the variables by influence with moderate computational costs on the order of 10² model evaluations. The PCE-Sobol' method is suitable to find the insignificant variables. This study provides practical guidelines for applying GSA in practical engineering applications.