Sensitivity analysis is a critical tool in reactor safety assessments, as it evaluates the impact of uncertainties in input parameters, identifies key factors, and highlights potential safety risks and measures. Conventional sensitivity methods, such as Spearman, Pearson, or Kendall, while straightforward, are typically limited to linear relationships and independent input parameters. Shapley values offer a more advanced, model-agnostic approach to sensitivity analysis, making them particularly valuable in scenarios with dependent parameters or nonlinear systems.

This study not only applies variance-based sensitivity methods, including Sobol indices and Shapley values, but also introduces the development of a reduced-order model (ROM) based on deep neural networks (DNNs) combined with Shapley values for time-dependent reactor simulations. This approach addresses the computational challenges of traditional methods, especially in cases involving correlated parameters, providing a more efficient and accurate sensitivity analysis. Sensitivity indices are calculated for the TWIGL benchmark, with two-group cross sections as the input parameters and core power during the ramp reactivity insertion transient as the output.

The results demonstrate that Shapley values, combined with the DNN-based ROM, yield robust, accurate, and physically meaningful indices, especially in models with dependent parameters where Sobol indices may lead to over- or underestimation and might even result in negative indices. This highlights the advantages of Shapley values for comprehensive and reliable sensitivity analyses in complex reactor simulations.