A machine learning material model for structural adhesives in finite element analysis

The computational analysis of adhesives within finite element simulations is important for many engineering applications, from automotive structures to civil engineering. The underlying material descriptions are closed, analytically formulated models representing the stress-response of the material for given strains. In this publication, we investigate and present a methodology for adopting artificial neural networks for this purpose. The aim is to show feasibility and performance of such models to be able to uncouple the material description from fixed formulations to reap the benefits of neural networks for materials modelling. A methodology is presented for creating training data for a structural adhesive, whereby the data is obtained from specimen and from generic strain paths of a baseline material model. From this, feed-forward neural networks are trained and evaluated using machine learning tools as well as from an engineering perspective by pre-defined test strain paths. It is shown that, given the quality and the quantity of training data is adequate, the model is able to generalize with respect to the response. Especially, the coverage of complete strain space is shown to be crucial. A framework for using machine learning models in a commercial finite element code is presented, integrating models trained through high-level programming interfaces. This enables validating the performance and accuracy of such models in explicit simulations, which is shown to be a crucial step in evaluation of machine learning material models. The model is confirmed to be able to represent structural adhesives with high precision comparable to an analytical model and to generalize for unknown geometries and loading conditions. An approach to predict element failure with the machine learning model is shown and finally, the neural network model response is made explainable from an engineering point of view using feature importance.