Graph neural networks have demonstrated notable advantages in modeling structured data for material property prediction. Nevertheless, their deployment in predicting the creep life of nickel-based single-crystal superalloys presents key challenges, as current models often fail to effectively capture the intricate nonlinear and synergistic interactions among multiple alloying elements. To address this, this paper proposes a Deep Learning-based Graph Attention Network v2(DL-GATv2) prediction framework that integrates physical metallurgy knowledge with an enhanced graph neural network. The model constructs an element graph structure and introduces the GATv2 dynamic attention mechanism to accurately capture the dependencies between elements. Simultaneously, it incorporates multidimensional physical metallurgy knowledge, including microstructure, electronic structure, and thermodynamics, to achieve multiscale modeling of the creep behavior of nickel-based single-crystal superalloys. Experimental results show that the model achieves a prediction accuracy of R² = 0.8218, with errors reduced by 24.7% to 38.0%, significantly outperforming existing methods. Ablation experiments and sensitivity analyses confirm the model's effectiveness and stability under complex operating conditions. More importantly, the weight distribution of element aggregation successfully reveals the key strengthening elements and their synergistic mechanisms across different generations of alloys, enabling a physically interpretable prediction of material behavior. This framework overcomes the limitations of traditional black-box models, providing an alternative paradigm for the composition optimization and performance regulation of high-temperature alloys, with both high prediction accuracy and physical interpretability.

Close

PrevNext

View figure in article

• Creep
• Material failure
• Structural properties
• Alloys
• Single crystal materials
• Deep learning