Data-Driven Exploration and Insights Into Temperature-Dependent Phonons in Inorganic Materials.

Area of Science:

• Materials Science
• Condensed Matter Physics
• Computational Materials Science

Background:

• Phonons (quantized lattice vibrations) are crucial for material properties but are often approximated, neglecting temperature-dependent anharmonic effects.
• Existing materials databases typically use the harmonic approximation, limiting predictions of real-world material behavior at finite temperatures.

Purpose of the Study:

• To develop a scalable computational framework for predicting finite-temperature phonons in crystalline solids.
• To improve the accuracy and efficiency of phonon predictions by incorporating machine learning and anharmonic lattice dynamics.

Main Methods:

• A machine learning interatomic potential (M3GNet) was fine-tuned with high-quality phonon data.
• The refined model was integrated with high-throughput calculations using the stochastic self-consistent harmonic approximation.
• Phonon predictions were computed for 4669 inorganic compounds.

Main Results:

• Phonon prediction accuracy was improved fourfold while maintaining computational efficiency.
• The study revealed systematic trends in anharmonic phonon renormalization across various material classes.
• Machine learning identified weak bonding, large atomic radii, and specific coordination as drivers of anharmonicity.
• Anharmonic effects were shown to significantly alter lattice thermal conductivity (2-4x).

Conclusions:

• The developed framework provides an efficient, data-driven platform for predicting finite-temperature phonon behavior.
• This approach can guide the discovery of novel materials with tailored thermal and vibrational characteristics.