Machine learning-assisted parameter identification for constitutive models based on concatenated loading path sequences

A hybrid strategy for the identification of material parameters of constitutive models is presented. One main challenge in the context of classic optimisation-based parameter identification schemes is the generation of adequate starting values for the multi-objective optimisation procedure. We address this issue by employing an artificial neural network that is trained with data obtained from the material model, more precisely speaking, from different solutions of the direct problem for homogeneous states of deformation. As a result, a solution of the inverse problem of parameter identification can be approximated with the help of a neural network processing experimental data. This approximate solution is used as a starting value for a subsequent, classic optimisation-based parameter identification approach. One key advantage of this strategy is that a network only has to be trained once per material model and can subsequently be applied for different materials. A rather sophisticated material model, incorporating gradient-enhanced damage coupled to plasticity in a geometrically non-linear setting, is used to demonstrate the capabilities of this approach. Two different strategies for the generation of training data are investigated and compared, and the neural network's hyperparameters are optimised for improved prediction capabilities. In view of the calibration of material parameters related to the non-locality of the model, the obtained set of parameters is again considered as a starting value for a final optimisation step, where we consider inhomogeneous states of deformation and digital image correlation data.

Automated summary

A hybrid strategy for identifying material parameters is proposed, which uses an artificial neural network to approximate the solution of the inverse problem and generate starting values for an optimization-based approach. The advantage of this strategy is that the network only needs to be trained once per material model and can be applied to different materials.