Welded joints are critical load bearing components in marine structures, automotive, and structural engineering, where local geometry significantly influences fatigue performance due to stress concentration. This work presents a data driven framework to quantify the effect of local weld geometry on fatigue life using automated finite element (FE) simulations and machine learning (ML) based geometric modeling. Two dimensional cross sectional slices of butt welded joints are modeled with high fidelity FE simulations, systematically varying key geometric parameters such as weld toe radius, reinforcement height, and undercut depth to capture realistic local imperfections. A fully automated pipeline enables rapid generation and analysis of thousands of geometric variations with minimal manual intervention, drastically reducing computational overhead. To explore the design space beyond measured geometries, we employ a machine learning generative model to synthesize diverse, physically plausible weld geometries. These synthetic geometries are then fed into the automated FE solver to compute local stress fields. From the resulting stress distributions, probabilistic stress concentration factors (SCFs) are derived, capturing the statistical variability induced by geometric uncertainty. The framework enables the construction of probabilistic fatigue life prediction models based on geometric variations, providing a robust statistical basis for reliability assessment. The approach bridges the gap between microstructural geometry and macroscopic performance, offering a scalable alternative to traditional parametric studies or expensive full 3D simulations. Results demonstrate correlations between geometric features and SCF distributions, with the ML generated geometries revealing non intuitive failure sensitive configurations. This methodology supports the design of more fatigue resistant welds and informs quality control standards by quantifying the impact of manufacturing variability.