Ensemble learning-based investigation of thermal conductivity of Bi2Te2.7Se0.3-based thermoelectric clean energy materials

Abstract

Bi2Te3-based materials are remarkable semiconducting compounds widely employed for clean energy harvesting via thermoelectric effects. Their energy conversion efficiency is assessed based on their thermoelectric figure of merit which depends on the electrical and thermal properties. The measurement of electrical conductivity is often straightforward; however, the thermal conductivity (κ) measurement is incredibly arduous and requires sophisticated experimental devices. Herein, we propose a pioneering machine-learning technique that estimates the value of κ for such materials based on their electrical properties and structural lattice constants (a and c). Decision tree regression (DTR) and Support vector regression (SVR) algorithms were employed to build the thermal conductivity estimators. Afterward, the concept of adaptive boosting was applied to the conventional regressors to improve their prediction accuracy. The performance of the developed models was evaluated based on the R2-value, coefficient of correlation (CC) between the actual and experimental values of κ, mean absolute error, and mean square error. The results obtained based on these metrics indicate that the boosted decision tree outperforms the rest of the models with a CC of 99.4% and an R2-value of 98.8% in the testing phase. The models were also utilized to make predictions in actual physical situations, such as forecasting the thermal conductivities of compounds doped with transition metals or non-metals, and examining how the values of κ are affected by substrate temperature during pulsed laser deposition. The remarkable performance of the models in predicting the thermal conductivity of different configurations of the material with high accuracy underscores its importance to researchers and industry for renewable energy applications.