Predicting Fillet Stresses from Singular Fields Using Machine Learning

A precise evaluation of peak stresses in fillet regions traditionally requires high-resolution finite element meshes, which substantially increases computational cost. This work introduces a new stress-prediction approach that integrates singular-field analysis with machine learning (ML). Rather than explicitly modeling the fillet, the method uses linearized stress components obtained from simplified models in the vicinity of the geometric singularity as input features for a regression model. A study performed on 2D plane-strain models demonstrated that linearized components yield the highest prediction accuracy, achieving a coefficient of determination of R2=0.98 and a mean absolute percentage error (MAPE) of 6.1%. The key advantage of the proposed approach is its ability to rapidly assess structural strength using simplified geometric representations without the need for detailed submodeling.