A Generative Adversarial Networks (GANs)-based data augmentation in the context of material testing and predicting strain rate effect

Kurzfassung

The aerospace industry is showing increasing interest to accelerate the certification process of composite aero-structure using machine learning (ML) models since certification often requires extensive testing to validate the safety and reliability. For example, the number of experimental tests, especially at the coupon level, can be reduced using ML models by identifying patterns and relationships while still ensuring the necessary levels of reliability in material characterisation. However, a large amount of data is essential to train ML models for learning the patterns to correctly predict the target values. The acquisition of training data, more particularly for the identification of strain-rate effects in composite materials, solely through material testing is often difficult due to high costs and time constraints, which often result in data scarcity. The lack of data presents considerable challenges when training ML models, as it can lead to overfitting and poor generalisation. This study reports the development of a data augmentation technique with a generative model Generative Adversarial Networks (GANs) as one of the potential solutions to alleviate the issues of data scarcity. Firstly, the capabilities of the GANs model are investigated for the generation of synthetic material test data, which should capture realistic physical phenomena. Secondly, the quality of the generated data and its similarity to available experimental datasets are evaluated. Evaluation metrics based on Principle Component Analysis (PCA) are developed to determine the acceptability of the generated synthetic data. Finally, the strain-rate effects on carbon/epoxy composites are exemplarily predicted using a Gaussian Process Regression (GPR). Furthermore, the management of a high number of heterogeneous datasets raises additional challenges to guarantee FAIR principles. In this work, experimental and synthetic data are parsed and converted into structured JSON data and labelled using an internal ontology. The storage in the research data management system of DLR, Shepard, allows to increase data findability and interoperability. The predicted results indicate that GPR can accurately forecast the in-plane shear strength of carbon/epoxy composites across various strain rates. This accuracy is achieved using a training dataset generated through a developed data augmentation technique. This framework supports a data-driven approach for creating synthetic data and establishes its acceptance criteria while also facilitating the prediction of the mechanical properties of carbon/epoxy composites.