Lattice Thermal Conductivity: An Accelerated Discovery Guided by Machine Learning

Area of Science:

• Materials Science
• Computational Materials Science
• Solid State Physics

Background:

• Predicting lattice thermal conductivity (LTC) is crucial for designing materials with specific thermal properties.
• Traditional methods like density functional theory (DFT) and phonon calculations are computationally intensive.
• Accurate and efficient prediction of LTC is needed for materials discovery.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting the LTC of crystalline materials.
• To identify novel compounds with ultralow LTC at room temperature.
• To establish a predictive framework for LTC-temperature behavior without requiring DFT-relaxed structures.

Main Methods:

• Collected LTC data for 119 compounds from DFT and phonon calculations across temperatures (100-1000 K).
• Trained various ML algorithms, with the random forest (RF) model showing the highest accuracy (R² = 0.957).
• Validated the RF model with new compounds and screened over 32,000 compounds from the Inorganic Crystal Structure Database.

Main Results:

• The RF-based ML model achieved high accuracy in predicting LTC, validated against theoretical calculations.
• Identified Cs₂SnI₆ and SrS as potential candidates with ultralow LTC.
• The model successfully predicted LTC-temperature behavior using only prototype structures and chemical compositions.

Conclusions:

• Machine learning, particularly the RF algorithm, provides an accurate and efficient method for predicting LTC.
• The developed model enables rapid screening of vast material databases for desired thermal properties.
• This approach accelerates the discovery of new materials with tailored thermal conductivity, such as those with ultralow LTC.