Peridynamics-driven two-stage conditional diffusion generative model for discontinuous fracture prediction in porous structure

• Developed a peridynamics-guided two-stage conditional diffusion generative model. • Bypassed complex physical calculations to predict fracture paths from geometry. • Introduced an attention-enhanced dual-branch U-Net to improve denoising capability. • Generated a fracture dataset of 2000 porous samples for training the model. • Delivered technical insights for precise fracture path prediction in engineering. The fracture behavior of microstructures is critical in determining the failure mechanisms of porous structure. A comprehensive understanding of the damage response in porous structures is essential for designing crack-resistant materials. To enable real-time and efficient prediction of fracture responses in porous structure, this study develops a peridynamics-driven two-stage conditional diffusion deep generative model. The model leverages strain energy density as an informative bridge to accurately reconstruct fracture paths directly from geometric information. Unlike conventional U-Net architectures, an attention-enhanced dual-branch U-Net is proposed as the denoiser within the conditional diffusion framework. A discontinuous fracture dataset comprising 2000 porous samples with varying porosities was generated using a bond-based peridynamic method to support physics data-driven learning. A multi-metric evaluation framework was then established to comprehensively assess the training dynamics and predictive performance. The proposed model achieved structural similarity scores of 88.92 % and 83.31 % in the two respective stages, demonstrating strong predictive capabilities. Moreover, from the perspective of data regeneration, the model exhibited excellent reconstruction performance. To quantify reliability, a set of metrics was designed to evaluate the fidelity and robustness of the generated results, revealing the profundity of the multi-step generative process. Additionally, the generalization of the model was thoroughly assessed, and an effective implementation framework for engineering applications was proposed. This study presents a powerful framework for predicting fracture responses in porous structure and offers novel insights into the failure behavior of porous materials.