Data-driven Computational Homogenization Using Neural Networks

Abstract

Fusion of data mining and computational mechanics is a modern approach for the exploitation of available data within rigorous modeling. First steps in this direction have been focused on the usage of neural networks and other soft computing tools as metamodeling tools. This framework seems suitable for numerical homogenization techniques realized within the so-called FE 2 environment, where the lower-level analysis of a detailed representative volume element is replaced by a prediction based on a previously prepared database. Numerically prepared data are used here, although the method can be used with experimental data as well. In this case, the need for a constitutive description of the fine scale is bypassed. Extraction of material properties from the database, required by the upper-level finite element analysis, is based on backpropagation artificial neural networks. The method is applicable to monuments and masonry structural systems. We investigate this approach here for the analysis of masonry structures with elastoplastic behavior. Results indicate a satisfactory comparison with published research.