A combined machine learning and EBSD approach for the prediction of {10-12} twin nucleation in an Mg-RE alloy

[Display omitted] The room-temperature ductility of the Mg alloys is closely related to the deformation behavior of the twin. However, there are currently no effective criteria that can accurately predict in which grains twins will nucleate during plastic deformation. With the rapid development of artificial intelligence technology, the applications of machine learning in the microstructure design and predictions have become the highlight. In the present study, a novel Mg-4Y-3Nd-2Sm-0.5 Zr alloy is prepared so as to explore the twin nucleation behavior of Mg alloys by combining machine learning along with electron backscattered diffraction (EBSD) techniques. At a true strain of 0.05, twins are found in 68 grains of the 297 grains which are counted from the initial microstructure. Eight features that may affect the twin nucleation are selected, including the grain diameter, the number of neighboring grains, the Schmid factor, and so on. Furthermore, the relevant importance of eight features on twin nuclei are also sorted; the grain diameter of original grains and the Schmid factor of the tensile twins have the greatest influence on the twin nucleation. Three machine learning algorithms including XGBoost, ANN, and the proposed relevance based ensemble scheme are used to model the prediction of the twin nucleation. The proposed relevance based ensemble scheme achieved an AUC score of 0.880, which is higher than that of the ANN (0.879) and XGBoost (0.756). A better ROC and PR curve also validate the feasibility of the proposed scheme.