Machine Learning Assisted Modelling of a Ductile Fracture

This work presents a computational framework for the determination of material parameters by a machine-learning method where finite element modelling (FEM) is supported by artificial intelligence of a neural network. An artificial neural network (ANN) model is developed and trained using merely the numerical experiments on a shear tension specimen. For the training database generated by FEM, the combinations of 3 unknown material parameters of a regularized ductile damage model are considered following a full factorial design of a 3-variable system with 3 levels (minimum, medium, maximum). Finite element (FE) simulations allowed to extract (i) the local displacement fields and (ii) the global force-displacement curves resulting from various parameter combinations. The ANN was trained by passing the displacement field images and force values throughout the complete numerical experiment in specified increments. In this way, valuable information about the local complex stress state evolution in a shear tensile test, from low over moderate to high stress triaxiality, is taken into account.<br/>The ANN model is verified by evaluating the model predictability of material parameters of FE simulations unseen in the training stage. Finally, the experimental data obtained from shear tensile test and Digital Image Correlation (DIC) analysis is introduced to the trained ANN to identify the parameter set that predicts the real mechanical response of the shear specimen in terms of local strain distribution, global force-displacement curve and the instant of fracture. Three different ANN architectures are studied and compared.<br/>It turned 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 ANN model. Therefore, this study presents a calibration methodology that uses one single specimen geometry instead of several different specimen geometries characterizing various stress states that are typically required to calibrate a ductile fracture model.