Deep learning for traction field prediction in delaminations with large-scale bridging

Abstract

Accurate modelling of delamination interfaces in fibrous laminated composites is computationally expensive, particularly when large-scale bridging zones form. Smeared-out approaches fail to capture the localised nature of the bridging fibres, while models that discretise the bridging ligaments become computationally intractable. Consequently, high-fidelity interface modelling approaches cannot be applied to large structures, limiting the full potential of laminated composites. This paper proposes a proof-of-concept for a hybrid approach that replaces the delamination interface with a physics-aware convolutional neural network structured as a modified U-Net architecture. The machine learning model is trained using data generated from a high-fidelity physics-based mechanical model. It takes as inputs the deformation field of each crack face and maps it to the corresponding discrete traction field of the bridging zone. Two models are compared: a physics-guided model, whose loss function minimises the mean squared error of the traction field, and a physics-informed model, in which the loss function is augmented by the J-integral of the region of interest. The evaluation shows that both models accurately capture the bridging tractions pattern and the global physical response of the region of interest, with an average J-integral error of less than 3.3% relative to the maximum J-integral value in the dataset. The physics-informed model demonstrates superior performance in capturing the underlying mechanics, providing accurate J-integral results even when the bridging pattern is poorly predicted. Once trained, both models can generate predictions across all simulation time steps in 1.5 s, potentially enabling new possibilities in the design of large composite structures.