PSI - Issue 75

Philippe AMUZUGA et al. / Procedia Structural Integrity 75 (2025) 53–64 Author name / Structural Integrity Procedia 00 (2025) 000–000

55

3

1. Generation of geometric, loading, and FAT class variables; 2. Stress calculation via FE analysis (Abaqus) and extraction of the hot-spot stress; 3. Fatigue life prediction N cycles using S–N curves; 4. Construction of SVR, MLP, RFR, and GLM metamodels.

The correlations between input variables and the response log 10 ( N cycles ), shown in Figure 2, highlight the hetero geneity in parameter influence. Among the eight variables, only t 2 ( r = 0 . 41), F ( r = − 0 . 39), and FAT ( r = 0 . 23) show moderate correlation with fatigue life. Other parameters such as a , h 1 , h 2 , t 1 , and θ exhibit near-zero correlation, indicating limited direct linear influence. This confirms the relevance of incorporating nonlinear transformations and automatic feature selection techniques, as implemented in the enhanced GLM, to capture more complex or interaction e ff ects. Sample results from the parametric model a t2 F FAT h1 θ t1 h2 log 10 ( N ) . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 12 2000 36 30 25 16 100 4.15 2 5 1000 50 20 30 5 60 4.32 5 4 750 90 45 75 8 100 5.14 10 4 750 90 45 75 24 100 5.19 15 4 1000 90 45 75 8 50 4.81 2 5 1000 80 40 30 5 40 4.94 10 24 2000 36 30 50 24 100 5.05 . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 5 500 806045 5 40 5.82 10 4 750 160 45 75 8 100 5.91 4 10 500 80 20 60 20 60 6.72 15 12 1000 160 45 50 16 100 6.97 4 5 500 125 40 30 5 40 6.42 10 12 750 90 45 50 16 100 6.58 . . . . . . . . . . . . . . . . . . . . . . . . . . .

Fig. 2: Combined visualization of a subset of parametric data used to train the model, including input variables and the response log 10 ( N cycles ).

The modeling pipeline adopted in previous work is based on a modular architecture, illustrated in Figure 3. It en compasses all necessary steps to build a GLM that is both accurate and interpretable, starting from a dataset generated through finite element (FE) simulations. This process includes: • Generation of input data via a parametric FE model ( Database );

• Logarithmic transformation of the target variable to stabilize variance and improve linearity; • Polynomial transformation of explanatory variables to capture nonlinear relationships; • Feature selection via RFECV to eliminate redundant or non-contributive variables; • Validation curve evaluation to balance bias and variance; • Custom performance metric based on domain-specific indicators.