On-the-Fly Machine Learning of Interatomic Potentials for Elastic Property Modeling in Al-Mg-Zr Solid Solutions

The development of resilient and lightweight Aluminum alloys is central to advancing structural materials for energy-efficient engineering applications. To address this challenge, in this study, we explore the elastic properties of Al-Mg-Zr solid solutions by integrating advanced machine learning (ML) techniques with quantum-mechanical (QM) atomistic simulations. For this purpose, we develop accurate and transferable machine-learned interatomic potentials (MLIPs) using two complementary approaches: (i) an on-the-fly learning scheme combined with Bayesian linear regression during ab initio molecular dynamics simulations, and (ii) the equivariant neural network architecture MACE. Both MLIPs facilitate the prediction of composition-dependent elastic properties while drastically reducing the computational cost compared to conventional QM methods. Comparison with ultrasonic measurements shows that the deviation between simulation and experiment remains within a few GPa across all Al-Mg-Zr systems investigated. These potentials also enable the systematic exploration of the Al-Mg-Zr solid solution phase space and provide insights into the elastic behavior as a function of alloying element concentration. Hence, our findings demonstrate the reliability and transferability of the parameterized on-the-fly MLIPs, making them valuable for accelerating the design of Al alloys with tailored physicomechanical properties in complex compositional spaces. While the present study focuses on homogeneous phases, it establishes a foundation for future multiscale simulations that include microstructural features such as precipitates and grain boundaries.