Active learning potentials for first-principles phase diagrams using replica-exchange nested sampling.

Area of Science:

• Computational Materials Science
• Materials Informatics
• Statistical Mechanics

Background:

• Predicting materials phase diagrams from first principles is computationally demanding.
• Machine-learning interatomic potentials require extensive, representative training data for accuracy.
• Current methods struggle with efficiently generating diverse and thermodynamically relevant data.

Purpose of the Study:

• To develop a fully automated active-learning (AL) strategy for generating training data.
• To compute complete pressure-temperature phase diagrams using machine-learning potentials.
• To enhance the reliability and transferability of machine-learning interatomic potentials.

Main Methods:

• Implemented an active-learning (AL) strategy integrated with replica-exchange nested sampling (RENS).
• RENS served as both an exploration engine and data acquisition mechanism.
• Selected configurations for DFT labeling based on RENS's diversity and likelihood-constrained sampling.

Main Results:

• Successfully applied the RENS-based AL approach to silicon, germanium, and titanium.
• Achieved convergence in approximately 10-15 AL iterations.
• Generated transferable potentials that accurately reproduce known phase transitions and thermodynamic trends.

Conclusions:

• The RENS-based AL strategy offers a general and autonomous method for creating machine-learning interatomic potentials.
• This approach enables accurate prediction of materials phase diagrams across wide thermodynamic conditions.
• Demonstrates a significant advancement in computational materials science for materials discovery and design.