A multi-fidelity data-driven model for highly accurate and computationally efficient modeling of short fiber composites

Abstract

To develop physics-based models and establish a structure–property relationship for short fiber composites, there are a wide range of micro-structural properties to be considered. To achieve a high accuracy, high-fidelity full-field simulations are required. These simulations are computationally very expensive, and any single analysis could potentially take days to finish. A solution for this issue is to develop surrogate models using artificial neural networks. However, generating a high-fidelity data set requires a huge amount of time. To solve this problem, we used transfer learning technique, a limited amount of high-fidelity full-field simulations, together with a previously developed recurrent neural network model trained on low-fidelity mean-field data. The new RNN model has a very high accuracy (in comparison with full-field simulations) and is remarkably efficient. This model can be used not only for highly efficient modeling purposes, but also for designing new short fiber composites.