Rationally Design Thermoelectric Materials Based on Ingenious Machine Learning Methods

Data quality, feature interpretability, and model generalization are critical and challenging for applying machine learning (ML) in the design of high-efficiency materials. In this work, an ML framework with integrating multi-step feature engineering is constructed for predicting the figure of merit (ZT) values of thermoelectric materials. By incorporating thermoelectric material data from the Starrydata2 database and implementing rigorous data cleaning, a high-quality ZT prediction dataset is established. An integrated strategy of feature extraction with combining Magpie and CBFV methods is utilized, followed by feature selection via Pearson correlation analysis and LassoCV cross-validation. Finally, the deep neural network model (Model-I) demonstrates excellent predictive performance (R² = 0.95 on the training set and R² = 0.90 on the test set), as well as identified successfully promising candidates such as CsCdBr₃ and TlBSe₃ in screening chalcogenide and halide perovskites. Combined with Density Functional Theory (DFT) calculation, the outstanding thermoelectric performance of CsCdBr₃ under p-type doping (ZT_max = 1.64) and the bipolar thermoelectric characteristics of TlBSe₃ (ZT_max = 1.04 for n-type and ZT_max = 0.99 for p-type) at 800K are successfully demonstrated, further confirming the reliability of our method. This study provides an applicative data-driven approach for functional material design, balancing predictive accuracy and physical interpretability.