Machine Learning Prediction of Bulk Bi2Te3 Thermoelectric Properties: Model Benchmarking, Transport-Physics Interpretation, and Calibration to Unseen Laboratories

Abstract Bismuth telluride alloys are the main thermoelectric materials for cooling and power generation near room temperature. We used published data on bulk Bi2Te3 alloys to test how well machine learning (ML) predicts their Seebeck coefficient, electrical conductivity, power factor (PF), thermal conductivity, and zT, and to examine what the models learn. An audited data set of 740 samples from 151 publications was compiled, and 15 regression models were trained on the 681 literature samples and compared under random, publication-grouped, and laboratory-grouped cross-validation. With random splits, the best models reach R2 = 0.72 for zT and 0.78 for PF, but for an unseen study R2 falls to 0.43 and 0.48, because 56 to 73% of the variance lies between publications. For zT and PF, bagging and boosting ensembles are the most reliable. Three measured samples from that study raise the R2 of the random forest model from 0.43 to 0.64 for zTand from 0.46 to 0.72 for PF, even when the study comes from a laboratory that is absent from the training data. A test on 31 later publications shows the same pattern (R2 for zT of 0.22 without and 0.67 with three measured samples). Transport analysis explains the model attributions: relative density and grain size act mainly through the weighted mobility, which is 1.7 times higher in p-type than in n-type samples, partly because untextured polycrystals penalize the more anisotropic n-type alloys; zT correlates with the quality factor (Spearman ρ = 0.70 and 0.78 for p- and n-type samples, respectively). The single-parabolic-band model places the zT optimum at 1.5 × 1019 to 1.8 × 1019 cm–3, below the PF optimum and below the carrier density of most n-type samples. The results are summarized as design guidelines for Bi2Te3 alloys at 290 to 340 K.