ECCOMAS 2024

Data augmentation for deep-learning modelling of composite materials

  • Uvdal, Petter (University of Gothenburg)
  • Cheung, Hon Lam (Chalmers University of Technology)
  • Mirkhalaf, Mohsen (University of Gothenburg)

Please login to view abstract download link

Recent trends confirm an increasing demand for Short Fibre Reinforced Composites (SFRCs), partially because they are suitable for injection moulding fabrication processes and possess a high strength-to-density ratio compared to unfilled polymer matrices, making them applicable in lightweight applications. To develop physics-based models and establish structure-property relationships for SFRCs, micro-mechanical models are used. However, traditional high-fidelity micro-mechanical methods are computationally expensive, especially for non-linear path-dependent modelling [Schneider et al., (2016)]. Given the significance of numerical modelling in iterative design processes, there is a current need for more efficient approaches for computational modelling of elasto-plastic behaviour of SFRCs. Recently, data-driven Artificial Neural Networks (ANNs) have been used as surrogate models, since they enable parallel computing, and conduct fast and efficient calculations. However, the data-hungry nature of ANNs remains a challenge, since creating large datasets of high-fidelity simulations requires extensive computational efforts and time. As a possible solution, transfer learning has been used recently, where multiple datasets are used to train an ANN [Cheung and Mirkhalaf (2024)]. While this technique produces accurate models, it is limited to scenarios containing multiple datasets. To address this challenge, we have developed a novel data augmentation approach by introducing random rotations. This method effectively expands a dataset of high-fidelity simulations of non-linear elasto plastic response of SFRCs without a need for additional simulations. Recurrent Neural Networks (RNNs) trained on the augmented datasets exhibit significant improvement in predicting high-fidelity micro-mechanical simulations. Our data augmentation method may also be used for modelling other materials, and hence, offers a promising solution for data scarcity issue. This opens possibilities for innovative advancements in deep learning modelling in materials science.