Machine Learning-Assisted Validation and Composition-Property Mapping of Thermoelectric Properties in the Ba1– x Sr x Zn2– y Mn y Sb2 System

Abstract A series of machine learning (ML) investigations were performed to validate experimentally obtained thermoelectric (TE) property trends and guide composition-property exploration in the Zintl-phase Ba1–xSrxZn2–yMnySb2 system. As an experimental test case, the three Mn or Sr/Mn-substituted BaCu2S2-type compounds, BaZn1.83(3)Mn0.17Sb2, BaZn1.71(3)Mn0.29Sb2, and Ba0.88(1)Sr0.12Zn1.75(3)Mn0.25Sb2, were successfully synthesized and structurally characterized by powder and single-crystal X-ray diffraction analyses. Their phase selectivity for the BaCu2S2-type phase was consistent with the r+/r– radius ratio criterion, and temperature-dependent transport property measurements confirmed the p-type behavior of the title materials. Given that the Mn or Sr/Mn substitutions systematically altered the balance between S and σ, the title system was considered as a useful benchmark for the ML-based trend validation for TE properties. Therefore, a customized TE dataset containing 5923 temperature-dependent entries from 949 materials was then constructed to train the three ML models for σ, S, and κtot predictions. First, seven different regression algorithms were comprehensively evaluated based on a coefficient for determination values, and XGBRegressor was selected for the further analysis since it offered the best balance between predictive accuracy and computational efficiency. The selected models were rigorously assessed using 10-fold cross-validation, residual and error analyses, learning curves, bootstrapped uncertainty estimation, Shapley additive explanations, and partial dependence plot/individual conditional expectation curve analyses of temperature-dependent property responses. The ML models reproduced the experimental performance hierarchy, correctly identifying BaZn2Sb2 as the highest PF compound among the studied samples. The 2-dimensional composition-property contour maps further predicted that moderate Sr substitution may lower κtot and improve projected ZT, indicating the Ba-rich/Sr-substituted BaZn2Sb2 derivatives as data-guided targets for future experimental optimization.