MACHINE LEARNING-ASSISTED COMPOSITION OPTIMIZATION OF TERNARY CHALCOGENIDES FOR ENHANCED THERMOELECTRIC EFFICIENCY

This Paper presents a machine learning approach for optimizing the power factor in ternary chalcogenide thermoelectric materials, using a dataset of 67,860 data points across 27 features. Key methodologies include data preprocessing, outlier removal via the IQR method, and feature importance analysis. The Random Forest model resulted in an R-squared value of 0.985 and effective predictive performance, particularly compared to linear regression. The top 10 influential parameters identified include electrical conductivity, Seebeck coefficient, temperature, and carrier concentration, which account for over 90% of the predictive importance. The models showcased robust generalization, validated through cross-validation and SHAP analysis to enhance interpretability by clarifying the nonlinear relationships between material features and power factor. This research establishes a comprehensive computational framework for the rapid screening of ternary chalcogenides, significantly reducing the experimental search space by over 95%. It demonstrates that machine learning models can predict power factor with high accuracy, facilitating virtual material screening. Additionally, feature importance analysis provides insights into power factor optimization mechanisms. The models identified promising candidates for experimental validation, while the methodology is generalizable to other thermoelectric materials, marking a significant advancement in the discovery and optimization of high-performance thermoelectric.