BENCHMARKING MACHINE LEARNING MODELS FOR POWER FACTOR PREDICTION IN BINARY THERMOELECTRIC COMPOUNDS

Abstract

Thermoelectric materials have the ability to directly convert waste heat into electricity, providing a sustainable energy solution. Electrical performance of these materials is determined by the power factor, which is an important parameter of the thermoelectric figure of merit. Nonetheless, high-performance binary thermoelectric compounds are sparsely found experimentally and are costly to synthesize. Machine learning has now become an influential instrument to speed up the discovery of materials, but a systematic benchmarking of algorithm currently existing in binary thermoelectric data is unavailable. In this work, we come up with a thorough machine learning benchmarking system to predict the power factor of binary thermoelectric compounds by relying on both compositional and elemental descriptors. An analysis of a large binary data set consisting of 22750 samples and 26 input variables obtained as the result of first-principles calculations was performed. Machine learning models used in predicting the power factor of cubic binary thermoelectric compounds are compared, such as linear regressors, support vector machines, tree-based ensembles, and neural networks. Following the hyperparameter tuning using randomized search, CatBoost has the highest predictive accuracy with a test R² of 0.9897, followed by Gradient Boosting (0.9869), LightGBM (0.9861), and Random Forest (0.9804). The poor performance of linear models and the support vector regressors implies the non-linearity of the structure-property relationships. The computational efficiency is evaluated by analyzing the training time of each model. The findings indicate that ensemble and gradient boosting models have a better predictive performance than linear and kernel based models. The additional study of model interpretability is conducted with the help of SHAP analysis to reveal the most significant descriptors that control the power factor prediction. Lastly, the importance of the polynomial feature expansion is examined to determine the effects of interaction between the features. The study offers a methodical reference of machine learning models to predict the power factor of binary thermoelectric materials and outlines the promising inquiry of data-driven methods in initial screening of materials.