Prediction of stress concentration at fillets using a neural network for efficient finite element analysis

Abstract

In finite element analysis, small fillets make mesh generation difficult and accurate evaluation of stress concentration at fillets requires refined meshes. Simplified analysis is often performed using a corner model where the fillets are removed. In the analysis using a corner model, mesh division becomes easier and the number of elements is reduced, which shortens the calculation time. However, the stress concentrations cannot be evaluated, and stress singularities occur at corners. We have developed a method for predicting the stress at a fillet based on the simulation of a simplified corner model and the use of a neural network. We use the stress distribution at a corner as the neural network input such that the method can be applied to arbitrary object shapes, loading, and boundary conditions. We trained and validated the neural network using simple corner and fillet models. It was shown that stress distribution at a corner can express the difference in loading conditions. In addition, we found that the method can predict stress at fillets of models that were not used for the neural network training. These results show the possibility that the method enables efficient stress concentration evaluation in finite element analysis.