Welded assemblies are commonly used in various industrial sectors. They constitute a versatile solution, notably due to their ability to withstand static and dynamic loads and temperature variations, all at reasonable costs. Their design and sizing are often governed by construction guides and standards to meet safety requirements, ensuring the integrity and durability of metallic structures. Fillet welds, commonly found in T-joints, are among the most common types. Fatigue represents the most frequent failure mode in welded assemblies. Designing these joints to resist fatigue is a major challenge, as the lifespan of the entire structure is reduced to that of the weld when it represents the most critical area. The evaluation and validation of the fatigue resistance of welded assemblies are not new and date back to the earliest industrial applications. Current methodologies, both analytical and numerical, aim for reliable accuracy to lighten structures without oversizing them, although they are often costly and time- consuming due to the need to adopt sophisticated techniques integrating specialized software. This study presents the application of Machine Learning to the metamodeling of a complete finite element analysis process by automating modeling, simulation, and fatigue analysis to estimate the fatigue life of T-welded joints, considering geometric parameters, loading conditions, and weld quality classes. Comparing the effects of configuration and data preprocessing of the training dataset for several regression estimators led to a reduced model providing explicit empirical rules for design and fatigue life evaluation. Our results demonstrate the potential of artificial intelligence in mechanical engineering practices

This study assesses the robustness of a polynomial Generalized Linear Model (GLM) for predicting fatigue life of T-welded joints under controlled Gaussian noise. Based on finite element and fatigue simulation data, various noise levels were injected into the target variable. Despite perturbations, the GLM maintained structural stability and showed improved performance under moderate noise (30\% amplitude, 10\% proportion), reducing RMSE to 0.041 and increasing accuracy to 80\%. Results suggest a critical noise threshold that enhances generalization, confirming the GLM’s reliability for early design under realistic industrial conditions.