Machine learning assisted calibration of a ductile fracture locus model

Abstract

While several different specimen geometries are typically required to calibrate a ductile fracture locus model, this article presents for the first time a calibration methodology that uses one single specimen geometry. This is accomplished by a computational framework that combines finite element modelling (FEM) and artificial neural network (ANN). The combinations of the model parameters are used to generate the training database. The local displacement fields and global force-displacement histories are extracted throughout the complete numerical experiment and passed to the ANN. Therefore, the influence of the local stress state on the evolution of the local deformation is implicitly taken into account. The trained ANN is verified by evaluating its predictability of material parameters of FE simulations unseen in the training stage. The experimental data obtained from the shear tensile test using Digital Image Correlation is introduced to the trained ANN to identify the parameter set that predicts the real mechanical response of the shear specimen. Three different ANN architectures with distinguished input representations are studied. It turns out that all of them can acceptably describe the experimental behaviour of not only the calibration specimen but also the specimens not used for training the model.