Gaussian process-based methodology to analyze influencing factors on numerical fatigue assessment

Abstract The FKM guideline [1] provides a robust framework for numerical fatigue assessment, ensuring component safety against fatigue failure. Its accuracy depends critically on the underlying input parameters and their interactions. In practice, the parameters are subject to a high degree of uncertainty. Global sensitivity analysis offers a rigorous, variance-based approach to quantify the influence of individual parameters and their interactions on the output of a model. This work proposes a workflow for numerical fatigue assessment under uncertainty. The workflow integrates a latent variable Gaussian process (LVGP) surrogate that simultaneously captures the effects of continuous FKM parameters, such as the ultimate tensile strength ( R m ), and the categorical influence of mesh refinement strategies. Finally, a Sobol sensitivity analysis is applied to the trained surrogate to quantify the contribution of each parameter to the variance in the utilization of the components. The methodology is demonstrated in three representative case studies, each using a different material and component. The resulting sensitivity rankings highlight the dominant role of material strength and mean stress effects under different loading ratios on the durability of components. Additionally, the results indicate that, within practical bounds, mesh discretization has a minimal impact on numerical fatigue assessment. The findings offer valuable insights for both uncertainty quantification and optimization of experimental campaigns in fatigue design, and they underscore the potential of mixed-variable surrogate-based sensitivity analysis in engineering practice.