Machine learning for improved density functional theory thermodynamics.

Area of Science:

• Computational Materials Science
• Machine Learning in Chemistry
• Thermodynamics of Alloys

Background:

• Density functional theory (DFT) is a powerful tool for predicting material properties but suffers from energy resolution errors.
• These errors significantly impact the accuracy of alloy formation enthalpies, especially in complex ternary systems.
• Reliable prediction of phase stability is critical for designing new materials for demanding applications.

Purpose of the Study:

• To develop a machine learning (ML) approach to systematically correct DFT energy errors in alloy formation enthalpies.
• To enhance the predictive accuracy of first-principles calculations for binary and ternary alloys and compounds.
• To improve the reliability of phase stability predictions for materials relevant to aerospace and protective coatings.

Main Methods:

• A neural network model was trained to predict the discrepancy between DFT and experimental enthalpies.
• The model used a feature set including elemental concentrations, atomic numbers, and interaction terms.
• Supervised learning, rigorous data curation, and cross-validation (LOOCV, k-fold) were employed to ensure model robustness and prevent overfitting.

Main Results:

• The ML model successfully learned to predict and correct DFT energy errors for alloy formation enthalpies.
• Application to Al-Ni-Pd and Al-Ni-Ti systems demonstrated significant improvement in prediction accuracy.
• The corrected predictions offer more reliable insights into ternary phase stability.

Conclusions:

• The presented ML approach provides a robust method for correcting DFT energy errors in alloy formation enthalpies.
• This significantly enhances the reliability of first-principles predictions for alloy phase stability.
• The corrected predictions are valuable for accelerating the discovery of new materials for high-temperature applications.