Presentation Information
[9a-S301-7]Stable and Functional Quaternary Heusler Alloys: A Machine Learning-Assisted High-Throughput First-Principles Exploration
〇(P)Enda Xiao1, Terumasa Tadano1 (1.NIMS)
Keywords:
Machine Learning,Machine Learning force field,Heusler
Heusler alloys are promising materials for applications in spintronic devices, shape-memory devices, and thermoelectrics. Building upon our DXMag comprehensive density functional theory (DFT) database of ternary Heusler compounds [1], we extend our exploration to quaternary systems with the formula XX′YZ. Given the vast chemical space of over 131,545 potential compositions, an efficient pre-screening method is essential to mitigate computational costs.
We employ a machine learning (ML)-assisted strategy to overcome this problem. Pretrained ML force fields (MLFFs) are utilized for structural optimization and stability estimation. The models trained on our database are used to do further estimation of stability and property. Successive filters based on formation energy and convex hull distance reduce the candidate pool to about 10%. Further screening, incorporating ML-predicted phonon stability and Curie temperatures, narrows the selection to a manageable subset for first-principles validation.
In this talk, we assess the reliability of different pretrained MLFFs in structural optimization and energetic stability evaluation. We also demonstrate the predictive accuracy of our custom ML models in estimating various properties. Our findings underscore the potential of this ML-assisted data-driven approach in accelerating the discovery of stable, multifunctional magnetic materials.
[1] E. Xiao and T. Tadano. arXiv preprint arXiv:2502.17946 (2025)
We employ a machine learning (ML)-assisted strategy to overcome this problem. Pretrained ML force fields (MLFFs) are utilized for structural optimization and stability estimation. The models trained on our database are used to do further estimation of stability and property. Successive filters based on formation energy and convex hull distance reduce the candidate pool to about 10%. Further screening, incorporating ML-predicted phonon stability and Curie temperatures, narrows the selection to a manageable subset for first-principles validation.
In this talk, we assess the reliability of different pretrained MLFFs in structural optimization and energetic stability evaluation. We also demonstrate the predictive accuracy of our custom ML models in estimating various properties. Our findings underscore the potential of this ML-assisted data-driven approach in accelerating the discovery of stable, multifunctional magnetic materials.
[1] E. Xiao and T. Tadano. arXiv preprint arXiv:2502.17946 (2025)