An efficient data-driven method for variance-based global and regional sensitivity analysis

Variance-based global sensitivity analysis represents a robust technique within the domain of scientific computing and advanced simulation. In response to this, the present study develops an efficient, data-driven methodology integrated with a density-based framework. This approach needs only a limited number of samples for execution and demonstrates broad applicability to problems characterized by high computational cost. Additionally, a novel regional sensitivity analysis technique, termed the Individual Contribution to Sample Variance (ICSV) plot, is introduced. This technique quantifies the distinct influence of specific input value ranges on the output variance, thereby offering valuable insights for reducing uncertainty in model predictions. For all input parameters, the ICSV representation can be constructed from the identical sample data applied in first-order effect index evaluation, ensuring no additional computational cost. A current limitation of the proposed data-driven approach is its restriction to computing first-order effect indices. This constraint may be addressed in future research through advancements in high-dimensional density estimation techniques. The proposed density-based method and ICSV plots are validated on benchmark test cases to illustrate their efficiency, accuracy, and practical applicability.