Few-shot augmentation based on variational auto-generative adversarial network with moving losses: Application to the variable stiffness prediction in composites

The dataset of carbon fiber reinforced plastics (CFRP) materials presents characteristics of high-dimensional, low-rank, and sparse, which pose difficulties in the combination of mechanical modeling. In this paper, a Variational Auto-Generative Adversarial Network (VAGAN) with moving losses is proposed as a data augmentation method, which extends the size of the CFRP dataset covering components, processes, elasticity, and strengths factors, and increases the information conveyed in the surrogate modeling. A compression strength prediction model for CFRP laminates was constructed by combining the component, process, and mechanical tensor with a neural network optimized by the search algorithm. Combined with the data augmentation strategy, not only was the amount of data expanded, but the prediction accuracy was also significantly improved. The allowable value of compression strength is analyzed and calculated by the predicted values, which brings direct benefits in simplifying the test.