Machine learning potentials for modeling alloys across compositions

Area of Science:

• Materials Science
• Computational Materials Science
• Chemical Physics

Background:

• Materials properties are fundamentally linked to chemical composition and arrangement.
• Predicting materials behavior across the full compositional spectrum, from ordered compounds to disordered solid solutions, is a significant challenge.
• Existing machine learning potentials (MLPs) struggle with accurately capturing diverse chemical arrangements, limiting predictive power in materials modeling.

Purpose of the Study:

• To develop advanced machine learning potentials (MLPs) capable of accurately modeling metallic alloys across their entire compositional and structural landscape.
• To enhance the predictive accuracy of MLPs by optimizing the sampling of chemical motifs using information theory.
• To enable high-fidelity materials modeling by effectively capturing complex chemical arrangements.

Main Methods:

• Integration of information theory with machine learning techniques to optimize the sampling of chemical motifs.
• Design and application of novel MLPs tailored for metallic alloys.
• Systematic prediction of materials properties, including stacking-fault energies, short-range order, heat capacities, and phase diagrams.

Main Results:

• Demonstrated effectiveness of the developed MLPs in predicting the compositional dependence of key materials properties for binary (AuPt, CuAu), ternary (CrCoNi), and high-entropy alloys (TiTaVW).
• Accurate prediction of stacking-fault energies, short-range order, heat capacities, and phase diagrams across diverse alloy systems.
• Validation against extensive experimental data confirmed the robustness and high physical fidelity of the approach.

Conclusions:

• The combined information theory and machine learning approach significantly improves the ability of MLPs to model materials properties.
• This method provides a robust framework for designing MLPs that accurately capture the behavior of metallic alloys across their full compositional range.
• The developed approach enables materials modeling with unprecedented physical fidelity, advancing the field of computational materials science.