A Recurrent Neural Network-based Surrogate Model for History-Dependent Multi-scale Simulations of Composite Materials

In order to make Computational homogenization affordable, pre-off-line finite element simulations are conducted on the mesoscale problem in order to build a synthetic database that can, in turn, be used to train surrogate models, which can be used as a constitutive law on a classical finite element simulation, speeding up the multi-scale process by several orders.

Artificial neural networks (NNWs) offer the possibility to serve as a surrogate model, but a difficulty arises for elasto-plasticity because of its history-dependency. This difficulty can be solved by considering a Recurrent Neural Network (RNN), which uses sequential information [1]. Nevertheless, in order to be accurate under multi-dimensional non-proportional loading conditions, a sufficiently wide database is required in order to perform the training. To this end, a sequential training synthetic database is obtained from finite element simulations on an elasto-plastic RVE subjected to random loading paths. The RNN predictions are thus found to be in agreement with the FE2 simulations, while reducing the computational cost by 4 orders.

Nevertheless, such a paradigm is essentially used as a mapping between the macro-stress and macro-strain tensors of the micro-scale boundary value response and the micro-structure information could not be recovered in a so-called localization step. We thus also develop Recurrent Neural Networks (RNNs)-based surrogate of the local micro-structure state variables for complex loading scenarios [2]. In order to address the curse of dimensionality arising because of the large amount of internal state variables in the micro-structure, we enrich the RNN with PCA dimensionality reduction and dimensionality break down, i.e. the use of several RNNs instead of a single one. The sequential training strategy is optimized to allow for GPU usage.