On-the-fly training of polynomial machine learning potentials in computing lattice thermal conductivity

Area of Science:

• Materials Science
• Computational Physics
• Solid-State Chemistry

Background:

• First-principles calculations and the linearized phonon Boltzmann equation are key for predicting lattice thermal conductivity (LTC).
• Accurate force constant determination is crucial for reliable LTC predictions.
• High-throughput material exploration demands efficient LTC calculation methods.

Purpose of the Study:

• To develop and validate an efficient workflow for first-principles LTC calculations.
• To reduce the computational expense of LTC predictions.
• To evaluate the performance of integrated polynomial machine learning potentials.

Main Methods:

• Integration of polynomial machine learning potentials within first-principles LTC calculations.
• Development of an optimized, modular workflow for automated calculations.
• Application to 103 compounds across wurtzite, zinc blende, and rocksalt crystal structures.

Main Results:

• Demonstrated significant reduction in computational resources required for LTC predictions.
• Successfully evaluated the performance of machine learning potentials in LTC calculations.
• Generated LTC data for a diverse set of crystalline compounds.

Conclusions:

• The integrated approach offers a computationally efficient alternative for predicting LTC.
• Polynomial machine learning potentials show promise for accelerating materials discovery through rapid LTC assessment.
• The developed workflow facilitates high-throughput screening of materials for thermal properties.