From data to reality: Criteria for the validation of data-driven models for history-dependent materials

This paper addresses the issue of validating identified data-driven material models for history-dependent materials, which are nowadays typically represented using neural networks. For this purpose, we introduce an a posteriori general acceptability criterion based on the Constitutive Relation Error (CRE) . A computed solution using the identified material model is deemed acceptable if it closely approximates the experimental data when using the CRE metric. • We define a posteriori criteria for the validation of identified history-dependent material models. • A CRE metric is used to assess proximity to the experimental constitutive manifold. • Additional technicalities are given on the computation of data-driven material models. • Insights are also given on the determination of the hidden variable dimension.