Machine learning-driven high-accuracy prediction and rapid design of mechanical properties in Mg-TM-RE alloys

Achieving a desirable balance between strength and ductility in magnesium alloys remains a major challenge due to the complex nonlinear interactions between alloy composition and thermo-mechanical processing parameters. In this work, a machine learning-driven framework was developed for the Mg-TM-RE alloy system to enable accurate prediction of mechanical properties and accelerated materials design. A comprehensive database containing 571 experimental samples was established, including alloy compositions, processing parameters, and mechanical properties. Five nonlinear regression models (SVR, MLP, RFR, GBR, and XGBoost) were constructed and evaluated. XGBoost achieved the best predictive performance for yield strength (YS) and ultimate tensile strength (UTS), while GBR showed superior accuracy in predicting elongation (EL). SHAP analysis revealed that thermo-mechanical processing parameters, particularly extrusion ratio and extrusion temperature, dominate strength-related properties, whereas alloy composition, especially Mg and Zr, play a critical role in regulating elongation and the strength-ductility balance. High-throughput exploration of the alloy-process design space identified promising Mg-TM-RE compositions capable of achieving YS≥300 MPa, UTS≥400 MPa, and EL>15%. Representative high-performance compositions fall within the range of (86-88)Mg-(1-3.5)Zn-0.5Zr-(0-5)Y-(0-1)Nd-(8-9)Gd. This work demonstrates the potential of machine learning-enabled materials design for accelerating the discovery of high-performance rare-earth magnesium alloys.