ForschungForschungsprojekte
A Consistent Hybrid Approach with Machine Learning for Time Dependent Constitutive Modelling

A Consistent Hybrid Approach with Machine Learning for Time Dependent Constitutive Modelling

Leitung:  Udo Nackenhorst
Team:  Darcy Beurle
Jahr:  2020

Machine learning is currently uncovering new possibilities in data- driven and meta-modelling for the field of computational mechanics. In particular, material modelling can be augmented or completely replaced with experimental results. Such techniques exist in the literature, such as data-driven material modelling. However, combining both a physical model and machine learning techniques in the form of an artificial neural network (ANN), results in so called ‘Physically Informed Neural Networks’ (PINNs), offering better accuracy over purely data driven data with imperfect observations. The handling of time- dependent data with a neural network in material modelling is still an area of active research, with a Long Short-Term Memory (LSTM) topology providing a promising approach.

The aim of this research project is to predict time dependent behaviour in a one-dimensional model with a combination of experimental data and classical material modelling, whilst ensuring thermodynamic consistency. Differential equations for the internal variables will be determined by machine learning and MOR techniques will be used to simplify the internal variable parameter space. Variants of ANN topologies will be investigated to determine which architecture type is suited to the underlying physics. Once the one-dimensional model is able to predict the response of a time-dependent material, this will be extended into a three-dimensional formulation.