Active Learning-Based Alloy Design Strategy for Improving the Strength/Ductility Balance of Al-Mg-Zn Alloys

Abstract

Al-Mg-Zn alloys, designed to combine the formability of 5xxx alloys with the high strength of 7xxx alloys, still face challenges in achieving an optimal strength-ductility balance. This study presents an active learning-based alloy design strategy to guide experiments aimed at enhancing the strength-ductility balance in Al-Mg-Zn alloys. Firstly, a sub-dataset comprising ultimate tensile strength (UTS) and elongation (EL) data with optimal generalization ability was identified from the small and disordered Al-Mg-Zn dataset using the bagging method. Subsequently, the bagging model of this sub-dataset was employed to construct a Pareto front based on the Upper Confidence Bound for UTS and EL, providing guidance for alloy composition design. Through experimental validation and iterative optimization, the strength-ductility balance of Al-Mg-Zn alloys was significantly improved, with the designed Al-5.27Mg-2.8Zn-0.44Cu-0.19Ag-0.15Sc-0.05Mn-0.01Zr alloy (wt.%) exhibiting superior mechanical properties with the measured UTS of 602 MPa and EL of 15.1%. Microstructural analysis using SEM, EBSD and TEM revealed that the improved strength/ductility balance of the alloy is attributed to its optimized composition, which results in the minimal micron phases, numerous fine Al3Sc particles, low-recrystallization grains, and a high density of precipitates. This active learning-based design strategy offering a novel approach for material development in systems with limited data.

Keywords: Aluminum alloy, Machine Learning, Active learning, Alloy design, Mechanical property