Abstract
To address the efficiency limitations of the Kriging (KG) model caused by single-point infill criteria, which add only one sample per iteration, and the excessive sample size requirement of multi-point infill criteria, which add a fixed number of samples per iteration, this paper proposes an AMSS-QBC-KG framework. This framework allows the sample size to decrease as model accuracy improves in the later stages of KG modeling, thereby enhancing modeling efficiency while controlling sample size. The effectiveness of the AMSS-QBC-KG framework was validated using standard test functions, with its predictive performance compared against existing infill criteria: Expected Improvement (EI), q-EI, Query-by-Committee (QBC), Adaptive Multi-Scale Sampling (AMSS), and Latin Hypercube Sampling (LHS). The results demonstrate that while achieving accuracy comparable to other methods for low-dimensional weakly nonlinear problems, the AMSS-QBC-KG framework exhibits superior predictive accuracy and reduced computational cost across diverse problem types. Furthermore, the framework's effectiveness was verified through an engineering case study on fatigue life analysis of vortex-induced vibration in submarine pipelines.