Machine Learning Model for the Prediction of Hubbard U Parameters and Its Application to Fe-O Systems

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Density functional theory (DFT) methods often struggle with electron self-interaction errors.
• The Hubbard U parameter is crucial for correcting these errors in local and semilocal functionals.
• Accurate and efficient determination of the Hubbard U parameter remains a significant challenge.

Purpose of the Study:

• To develop a machine learning-based approach for predicting the Hubbard U parameter in iron oxides.
• To establish a correlation between structural fingerprints and U values obtained from linear response methods.
• To improve the accuracy and efficiency of DFT calculations for iron oxide properties.

Main Methods:

• Machine learning fitting of structural fingerprints to U values.
• Evaluation of U using linear response-constrained density functional theory.
• Application of the developed method to wüstite, hematite, and magnetite.
• Redefinition of convex hulls for the Fe-O system at various pressures.

Main Results:

• The machine learning method accurately predicts Hubbard U for iron oxides.
• Calculated properties for wüstite, hematite, and magnetite show good agreement with experimental data and hybrid functional results.
• Redefined convex hulls for the Fe-O system at 0, 50, and 100 GPa align with experimental observations.
• Insights provided into high-pressure phase debates for Fe2O3 and Fe3O4.

Conclusions:

• The developed machine learning approach offers a computationally efficient and accurate way to determine the Hubbard U parameter.
• This method enhances the predictive power of DFT for iron oxide materials.
• The findings contribute to a better understanding of iron oxide phase stability under high pressure.