Accelerated discovery of high-performance Al-Si-Mg-Sc casting alloys by integrating active learning with high-throughput CALPHAD calculations

Abstract

Scandium is the best alloying element to improve the mechanical properties of industrial Al-Si-Mg casting alloys. Most literature reports devote to exploring/designing optimal Sc additions in different commercial Al-Si-Mg casting alloys with well-defined compositions. However, no attempt to optimize the contents of Si, Mg, and Sc has been made due to the great challenge of simultaneous screening in high-dimensional composition space with limited experimental data. In this paper, a novel alloy design strategy was proposed and successfully applied to accelerate the discovery of hypoeutectic Al-Si-Mg-Sc casting alloys over high-dimensional composition space. Firstly, high-throughput CALculation of PHAse Diagrams (CALPHAD) solidification simulations of ocean of hypoeutectic Al-Si-Mg-Sc casting alloys over a wide composition range were performed to establish the quantitative relation ‘composition-process-microstructure’. Secondly, the relation ‘microstructure-mechanical properties’ of Al-Si-Mg-Sc hypoeutectic casting alloys was acquired using the active learning technique supported by key experiments designed by CALPHAD and Bayesian optimization samplings. After a benchmark in A356-xSc alloys, such a strategy was utilized to design the high-performance hypoeutectic Al-xSi-yMg alloys with optimal Sc additions that were later experimentally validated. Finally, the present strategy was successfully extended to screen the optimal contents of Si, Mg, and Sc over high-dimensional hypoeutectic Al-xSi-yMg-zSc composition space. It is anticipated that the proposed strategy integrating active learning with high-throughput CALPHAD simulations and key experiments should be generally applicable to the efficient design of high-performance multi-component materials over high-dimensional composition space.

Aging heat treatment design for Haynes 282 made by wire-feed additive manufacturing using high-throughput experiments and interpretable machine learning

“…Certainly, a more extensive investigation assessing different critical properties, such as the creep resistance and ductility, is necessary to fulfill the industrial application requirements. However, such work will requires more extensive experiments and implementation of multi-objective optimization techniques [ 38 ]. Instead of establishing optimal heat treatment parameters that satisfy all properties, this study aims at demonstrating that the combination high-throughput experimentation, precipitation model, and interpretable machine learning technique can rapidly enhance our understanding of how heat treatment parameters affect specific properties of additive manufactured parts and to assist in the design process.…”