An efficient multiscale surrogate modelling framework for composite materials considering progressive damage based on artificial neural networks

Modelling of the progressive damage behaviour of large-scale composite structures presents a significant challenge in terms of computational cost. This is due to the detailed description in finite element (FE) models for the materials, i.e., with each unidirectional layer defined as required by the applicability of laminate failure criteria, and numerous iterations required to capture the highly nonlinear behaviour after damage initiation. In this work, we propose a method to accelerate the nonlinear FE analysis by using a pre-computed surrogate model which acts as a general material database representing the nonlinear effective stress-strain relationship and the possible failure information. Developed using artificial neural network algorithms, the framework is separated into an offline training phase and an online application phase. The surrogate model is first trained with a vast number of sampling data obtained from mesoscale unit cell models offline, and then used for online predictions on a macroscale FE model. The prediction accuracy of the surrogate model was examined by comparing the results with conventional FE modelling and good agreement was observed. The presented method enables progressive damage analysis of composite structures with significant savings of the online computational cost. Lastly, the surrogate model is only based on material designs and is reusable for other structures with the same material.