AI-driven ensemble learning for accurate Seebeck coefficient prediction in half-Heusler compounds based on chemical formulas

Half-Heusler compounds with their unique crystal structure and tunable properties show excellent thermoelectric properties. As a result, these compounds hold great potential for many applications such as thermoelectric devices and energy conversion technologies. This paper presents a machine learning study on developing ensemble learning models to accurately predict the Seebeck coefficient at ambient temperature and arbitrary carrier concentration for both n-type and p-type half-Heusler compounds, based only on their chemical formula. Ensemble learning algorithms proved to be effective regression models for identifying hidden relationships. All models displayed satisfactory performance under 10-fold cross-validation and achieved high R-squared values ranging from 0.87 to 0.95 and low MAE values ranging from 20.8 μV/K to 37.04 μV/K. Among ensemble models, the GBoost shows the best performance for p-type with an R2 value of 0.95. On the other hand, for n-type the Light GBoost and CatBoost models yielded the best results with R2 = 0.94. Ultimately, we validated the accuracy of the model outputs through rigorous Density Functional Theory (DFT) calculations, wherein these models demonstrated their remarkable effectiveness in precisely predicting the Seebeck coefficients for half-Heusler compounds. This research holds significant potential in facilitating the screening of materials for thermoelectric devices, which heavily rely on the critical Seebeck coefficient parameter. The ability to accurately predict this key property can accelerate the identification of promising candidate materials, driving advancements in thermoelectric technology and paving the way for more efficient energy conversion and management solutions.