Pgcnn: an interpretable graph convolutional neural network for predicting the mechanical properties of Ti-6Al-4V alloy

This study introduces a polycrystalline graph convolutional network (PGCNN) to predict the mechanical properties of Ti-6Al-4V alloy’s dual-phase polycrystalline microstructure. The model captures complex inter-grain interactions. It integrates node features and graph structural information to map microstructures to macroscopic mechanical properties. The PGCNN model demonstrated exceptional predictive performance (mean absolute relative error, MARE = 0.369%). It remained robust in handling nonlinear relationships and capturing high-order inter-grain interactions, even with limited datasets (MARE = 1.985%). We evaluated the interpretability of the PGCNN model through analyses at the node, edge, and graph structure levels, offering comprehensive insights. At the node level, the influence of each grain (node) on the output was quantified, clarifying the direct link between individual grains and macroscopic performance. Edge level analysis emphasized the importance of inter-grain interactions. It laid the groundwork for identifying grain boundaries that significantly affect mechanical properties. Graph level analysis quantified the overall impact of microstructural features on macroscopic performance. This provided insights into the complex “microstructure–mechanical property” relationship in dual-phase polycrystals.