Damage-field-informed hybrid learning for fatigue life prediction of CFRP bolted joints under small-sample conditions

Fatigue life prediction of carbon fiber-reinforced polymer (CFRP) bolted joints remains challenging under small-sample conditions, as the macroscopic fatigue response is closely associated with hole-edge damage evolution that is difficult to quantify from limited tests. In this study, a damage-field-informed hybrid learning framework, termed BFP-Net, is developed to bridge mesoscale damage fields and macroscopic fatigue life. An experimentally validated progressive damage finite element model is first used to generate damage fields around the bolt hole. A conditional generative adversarial network is then introduced using structural parameters and fatigue-life labels as conditions to augment the damage-field dataset while maintaining the learned parameter–damage correspondence. Fatigue-relevant damage features are extracted from the simulated and generated fields using a convolutional neural network and fused with structural parameters in a back-propagation neural network for fatigue life regression. The proposed framework achieves a coefficient of determination of 0.9525 on the test set and an average relative error of 3.95% across 12 independent fatigue tests; compared with the parameter-only baseline model, the test-set coefficient of determination is increased by more than 25%. These results indicate that damage-field representations derived from physically based simulations can provide an effective route for fatigue life prediction of CFRP bolted joints when experimental fatigue data are limited.