Addressing data scarcity in deep learning: Leveraging real and artificial datasets to predict compaction of composites
А. А. Коптелов, Hanna Beketova, Jonathan P.-H. Belnoue, Stephen R. Hallett, Iryna Tretiak
Abstract
• Predictive framework forecasted composite compaction response to processing conditions with an average final thickness error of 5.5% • Compaction dataset of real experimental data of over 1,100 tests was created for toughened carbon/epoxy prepregs. • Real data is important for capturing all complex aspects of compaction response like springback. • Artificial data can be tailored to reflect specific processing conditions and can improve robustness of predictions. • Combining real and artificial data improved target metrics, helping address data scarcity in modelling. This study focuses on predicting the compaction behaviour of composite materials under processing conditions using Long Short-Term Memory (LSTM) neural networks. The bespoke predictive system accurately captured key stages of material behaviour with an average final thickness prediction error of 5.5%. A major outcome of this research was the creation of the large real compaction dataset for toughened prepreg materials, offering a valuable resource for the development of new material models. The study also explored aspects of leveraging both real and artificial data in training predictive models. While real data remains essential for capturing the full complexity of the studied system, its availability is often limited. Incorporating artificial data together with real data in the training set enhanced the overall prediction robustness, offering a potential solution to the issue of data scarcity. However, such technique introduced biases in predicting certain aspects of material behaviour, such as springback, highlighting the importance of a balanced approach when assembling training data.