Models with high-dimensional parameter spaces are common in many applications. Global sensitivity analyses can provide insights on how uncertain inputs and interactions influence the outputs. Many sensitivity analysis methods face nontrivial challenges for computationally demanding models. Common approaches to tackle these challenges are to (i) use a computationally efficient emulator and (ii) sample adaptively. However, these approaches still involve potentially large computational costs and approximation errors. Here we compare the results and computational costs of four existing global sensitivity analysis methods applied to a test problem. We sample different model evaluation time and numbers of model parameters. We find that the emulation and adaptive sampling approaches are faster than Sobol' method for slow models. The Bayesian adaptive spline surface method is the fastest for most slow and high-dimensional models. Our results can guide the choice of a sensitivity analysis method under computational resources constraints.

翻译：具有高维参数空间的模型在许多应用中十分常见。全局敏感性分析能够揭示不确定输入及其交互作用对输出的影响。对于计算要求较高的模型，许多敏感性分析方法面临重大挑战。应对这些挑战的常用方法是：(i) 使用计算效率高的仿真器，以及(ii) 自适应采样。然而，这些方法仍然可能涉及较大的计算成本和近似误差。本文比较了四种现有全局敏感性分析方法应用于一个测试问题的结果和计算成本。我们对不同的模型评估时间和模型参数数量进行了采样。结果发现，对于慢速模型，仿真和自适应采样方法比Sobol'方法更快。其中，贝叶斯自适应样条曲面方法在大多数慢速和高维模型中速度最快。我们的研究结果可为计算资源受限条件下敏感性分析方法的选择提供指导。