Feature mining for thermoelectric materials based on interpretable machine learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Traditional material optimization relies heavily on empirical data and extensive trial-and-error, consuming significant time and resources.
• Machine learning (ML) offers a powerful alternative for complex optimization problems in materials science.

Purpose of the Study:

• To establish the relationship between physical characteristics and the thermoelectric figure of merit (zT) for thermoelectric materials.
• To identify critical features influencing experimental outcomes and analyze indirect effects.

Main Methods:

• Utilized an interpretable machine learning approach on a thermoelectric material database.
• Employed feature engineering to construct and optimize ML models based on identified key features.
• Evaluated different feature combinations to determine optimal descriptors for the experimental system.

Main Results:

• Identified key molecular features directly and indirectly affecting experimental results.
• Optimized ML models by comparing the efficiency of various feature combinations.
• Determined optimal feature descriptors for the specific experimental system.

Conclusions:

• The developed ML models enable high-throughput screening of thermoelectric materials.
• This data-driven approach significantly enhances the efficiency of experimental optimization for thermoelectric materials.