Estimation of elastic-plastic notch strains and stresses using artificial neural networks

The prediction of fatigue lives of metallic components using the notch strain approach is based on local stresses and strains. These can be determined either by analytical approximation formulas such as Neuber's rule based on linear elastic stresses or by finite element analysis with elastic-plastic material behavior. The latter has the disadvantage of longer computation times. This paper discusses a draft for application of artificial neural networks as an alternative approach for estimating uniaxial and multiaxial proportional elastic-plastic stresses and strains in notch roots based on linear elastic stresses. These networks are trained on results from elastic-plastic finite element analyses on flat bars notched on both sides and circumferentially notched shafts to predict local stresses. The resulting accuracy is compared to the analytical approximation formulas. The research findings confirm that neural networks can estimate local stresses with similar scatter and better mean value compared to analytical formulas.