ANN-based metamodels for effective stress determination

Please cite: https://doi.org/10.1016/j.ijfatigue.2026.109692
This repository contains three artificial neural network (ANN) based metamodels, training datasets as well as lightweight python-based helper scripts to load and implement the metamodels.
- ANN to predict effective stresses based on first principal stresses
- ANN to predict effective stresses based on von Mises stresses under plane strain
- ANN to predict effective stresses based on von Mises stresses under plane stress
The metamodels are used to predict effective stress in notched flat bars based on the Theory of critical distances (TCD) methodology with outputs for the point method (PM), line method (LM), and area method (AM). The models are trained with the MATLAB Deep Learning toolbox using data from a parametric study with finite element analysis (FEA) simulations using a range of notch geometrical parameters under varied loading conditions. These metamodels can simplify the process of estimating notch failure, through eliminating the need for FEA analysis and stress field integration for each individual notch geometry.