Predicting grain boundary damage by machine learning

Damage in metallic materials usually originates at grain boundaries where deformation leads to local strain incompatibility and stress concentration. Characteristics of a grain boundary affects its tendency to damage, but a quantitative understanding is still lacking. In this work, we utilize machine learning to predict grain boundaries susceptible for damage nucleation. To this end, synthetic 3D microstructures representing single-phase magnesium polycrystals with different textures were created. A Crystal Plasticity Fast Fourier Transform (CPFFT) framework was used to simulate the uni-axial tensile deformation of each microstructure. From the simulation results, those grain boundaries (GB) having strain incompatibility or stress concentration can be identified. We selected 46 crystallographic and geometric features for each GB to build classification models using the XGBoost algorithm. The model for predicting strain incompatible grain boundaries (SI-GBs) and the model for predicting stress concentration grain boundaries (SC-GBs) achieve the ROC-AUC (area under the receiving operating characteristic curve) scores of-87% and-82%, respectively, among all textures. From the classification models, the most important features that lead to SI-GBs or SC-GBs were identified. This work demonstrates that the combination of traditional simulation approaches and machine learning can yield new knowledge in fundamental material behaviors.

Automated summary

In this study, machine learning was used to predict grain boundaries susceptible to damage, and the characteristics and importance of these boundaries were analyzed through simulations and classification models, providing new knowledge for understanding the behavior of metallic materials.