Thermoelectric Material Performance (zT) Predictions with Machine Learning.

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning

Background:

• Machine and deep learning models show promise in predicting material properties like thermoelectric (TE) performance.
• Existing models often rely on diverse data sources including experimental, crystallographic, and DFT data.
• Accurate prediction of the figure of merit (zT) is crucial for advancing TE materials.

Purpose of the Study:

• To develop an interpretable machine learning model for predicting TE material performance.
• To utilize a large experimental dataset for training and validation.
• To directly predict the complex zT property across a wide range of TE materials.

Main Methods:

• Developed an interpretable machine learning model.
• Trained the model on a substantial experimental dataset of approximately 160,000 data points.
• Validated the model's predictive accuracy on three independent test sets.

Main Results:

• The model achieved high accuracy in predicting TE material performance, with RMSE values between 0.15 and 0.20.
• Evaluation coefficients (R²) ranged from 0.80 to 0.67 across test sets.
• Identified potential sources of deviation between predicted and experimental zT values.

Conclusions:

• The developed interpretable model effectively predicts the figure of merit (zT) for thermoelectric materials.
• The study provides insights into factors influencing prediction accuracy, such as experimental variability.
• This work represents a significant step in directly predicting complex TE properties across diverse materials using experimental data.