A Data-driven Approach to predict Strain Rate Effect of Carbon/epoxy Composites incorporating Constitutive Artificial Neural Networks (CANNs)

Kurzfassung

The design of polymer composite structures for crashworthiness is significantly affected by strain-rate effects, through which mechanical properties vary as a function of the local loading rate of the material. This is especially crucial when considering crash or impact loaded structures, particularly in the range of intermediate strain rates. In recent years, the application of a data-driven approach has increased to predict the mechanical properties of fibre reinforced plastics (FRP) using machine learning models [1]. The dataset size and quality are crucial for a data-driven approach to accurately analyse and predict material behaviours. However, the acquisition of such data through physical experiments alone is often constrained by several factors, such as cost and time. To address these challenges, we propose developing a novel data augmentation framework that combines Constitutive Artificial Neural Networks (CANNs) [2] to find the constitutive law of strain-rate effects automatically with transfer learning. This framework will serve two essential components. First, it will generate synthetic data that complements existing experimental results on the dynamic compressive behaviour of carbon/epoxy composites. This approach will incorporate constitutive material models into neural networks to ensure the meaningful prediction of strength properties. Second, the framework will bridge gaps in experimental datasets by learning underlying patterns from available test data. By developing a base model for well-characterised carbon/epoxy composites and implementing transfer learning techniques, we can efficiently extend a compressive strength prediction to new composite materials with limited test data. The developed approach will enhance the predictability of strain-rate effects in FRP by integrating generative modelling for data augmentation and applying transfer learning. By systematically expanding the dataset and leveraging existing models, we can improve the accuracy of strain rate predictions for new composite materials.