Recognition of mechanical behaviour laws through the supervised learning of an artificial neural network

Abstract

In this paper, we put forward a classifier to automatically identify the most suitable mechanical behaviour law among two candidates, enabling optimal modelling of experimental data from a uniaxial tensile test, represented as elongation–specific stress curves. An ArtificialNeuralNetwork (ANN) is employed to performthis classification task and assist the modeler, even when the resulting curves from different models are very close. This paper compares two different methods that enable supervised learning of the neural network from a training dataset labelled with specific stress values as a function of elongation, without requiring any external data. The learning process is then validated by testing the network’s ability to correctly identify the most suitable model from experimental data it has never encountered before. This approach could potentially pave the way for adaptations to more complex tasks in a multidimensional context or to feature recognition in imaging, in the frame of various fields of materials mechanics.