ON THE ACCURACY AND EFFICIENCY OF CONVOLUTIONAL NEURAL NETWORKS FOR ELEMENT-WISE REFINEMENT OF FEM MODELS

In this paper, a new methodology for the choice of the best structural theories through Machine Learning (ML) techniques is described, with a particular focus on composite shells. The identification of the most adequate theory can be operated very efficiently using Convolutional Neural Networks (CNN) as surrogate models to replicate the performances of a Finite Element (FE) formulation, although requiring only a small fraction of the usual amount of analyses. Enhanced by the introduction of the Carrera Unified Formulation (CUF), the FE Method (FEM) provides the results necessary for the training of the networks, while the Node Dependent Kinematics (NDK) approach opens to the practical implementation of local refinement capabilities. The evaluation of different structural theories is carried out with the Axiomatic/Asymptotic Method (AAM) and this can be done for both static and dynamic analyses, with The Best Theory Diagrams (BTD) being the outcome of this rating procedure. As shown in the results, CNNs can properly identify and reproduce the underlying connections between different sets of problem features and the accuracy of a given structural theory with just a very small amount of available reference data.