Neural networks as material models within a multiscale approach

摘要

This paper shows the application of neural networks in a multiscale analysis of a reinforced concrete beam. A mesoscale model is presented, which simulates the pullout test of a reinforcement bar in concrete. By applying a homogenization procedure, a macroscopic stress vs. crack opening response is obtained from the mesoscale simulations. The neural network is used to approximate this relation in a macroscale simulation and replaces the material formulation of the interface layer between concrete and reinforcement, thus avoiding the computationally expensive parallel simulation on different scales.