A spatiotemporal prediction model for rapid prediction of delamination growth in open-hole composite laminates

Delamination of laminates around the hole edge is one of the most common but most threatening forms of damage in composites. The numerical solutions based on the finite element method (FEM) are difficult to meet the requirements of online evaluation and decision-making of the structure operation in terms of prediction efficiency. In this study, an end-to-end spatiotemporal prediction model with Encoder-Translator-Decoder architecture is developed to rapidly predict the delamination growth of laminates under compressive loading. The prediction results of single-hole delamination growth at different locations show that the model has good accuracy and generalization. The trained model can complete an inference prediction at the millisecond level, which is hard to achieve with FEM. Furtherly, a transfer framework is designed to promote the forecasting of multi-hole delamination growth. The framework is designed with a coupling coding block for learning the complex relationship of delamination damage superposition, which can effectively accelerate the training and achieve better prediction accuracy and inference speed. The developed deep learning model has good potential to be extended and applied to predict the spatiotemporal evolution of other physical fields.