Rapid prediction of full-field stress evolution during drilling of carbon fiber reinforced polymer: A finite element database-driven graph recurrent surrogate model

Abstract

Stress evolution during carbon fiber reinforced polymer (CFRP) drilling governs drilling-induced damage and strongly affects hole quality and the structural reliability of aerospace components. Rapid and accurate prediction of stress-field evolution is essential for online process optimization and damage control. Experimental methods have restricted access to internal field variables, and finite element (FE) simulations are computationally expensive. This study proposes a FE database-driven surrogate modeling framework that couples a graph neural network (GNN) with a gated recurrent unit (GRU) for rapid prediction of full-field equivalent stress evolution during CFRP drilling. Rather than reproducing the detailed FE solving process, the proposed model transforms high-dimensional field-variable prediction into graph-based feature propagation and temporal state updating. Graph-structured spatiotemporal samples are constructed from the FE database, with nodal equivalent stress responses as targets. The GNN captures spatial stress propagation patterns over the FE mesh, while the GRU models node-level temporal stress evolution. A dual-head output mechanism predicts stress increments and nodal survival states, the survival output gating the stress updates during progressive material removal. Under unseen cutting speeds, unseen feed rates and their combination, the model achieves an average mean absolute error (MAE) of 10.00 MPa (MPa) and nodal survival classification accuracy of 99.82%. Single-condition full-field prediction takes 0.199 s (s), an average speedup of about 3.2 × 10 3 over FE simulation. Baseline comparison and ablation results confirm the contribution of graph-based spatial modeling, node-level recurrent temporal modeling, and dual-head design. The proposed model can support online hole-quality assessment, damage control, and virtual-real interaction model development in CFRP drilling.