Presentation Information
[8a-N302-1]How Reliable Are Machine-Learning Predictions of Thermoelectric Performance?
〇(PC)Andrei Novitskii1, Vladimir Baturin2,3, Guillaume Lambard3, Jean-Claude Crivello2, Takao Mori1,4 (1.MANA, NIMS, 2.LINK, NIMS, 3.CBRM, NIMS, 4.Tsukuba Univ.)
Keywords:
thermoelectric performance,machine learning
Machine learning is increasingly applied to accelerate thermoelectric materials discovery, yet the reliability of models trained on experimental data remains difficult to assess. In this presentation, we examine how conventional benchmark metrics relate to practical predictive performance using two large experimental thermoelectric databases. We show that models based only on chemical composition can achieve benchmark scores comparable to those using more sophisticated physically motivated descriptors, despite exhibiting different predictive behavior. Furthermore, the reliability and transferability of ML models strongly depend on dataset composition and chemical-space coverage, which are not captured by standard metrics such as R2 or RMSE. These findings demonstrate that benchmark scores alone are insufficient for evaluating ML models trained on heterogeneous experimental datasets and highlight the importance of considering data quality, dataset diversity, and transferability when applying data-driven approaches to thermoelectric materials discovery.