Accelerated Discovery of Advanced Thermoelectric Materials via Transfer Learning

Abstract Thermoelectric (TE) technology can realize direct conversion of widely distributed heat into useful electricity, which provides a promising route to solve the global energy crisis that is increasingly severe. However, it is extremely complex and time‐consuming to discover advanced TE materials via conventional trial‐and‐error approaches. In this work, using a pre‐trained neural network architecture for the electronic bandgap, a transfer learning (TL) strategy that allows ready and accurate prediction on the ZT values of any TE materials at arbitrary temperature is proposed. Compared with direct machine learning algorithms, the TL‐driven model exhibits significantly enhanced predictive power beyond the initial dataset, as characterized by improved Pearson correlation coefficient (reduced mean absolute error) from 23% to 95% (0.35 to 0.07) for the p ‐type systems, and 46% to 94% (0.23 to 0.06) for the n ‐type systems. By screening 6353 possible candidates in the AFLOW repository that having relatively smaller gaps, 925 p ‐type and 788 n ‐type systems are quickly identified to exhibit ZT exceeding 2.0. Equally importantly, the established TL model is highly adaptable to ZT prediction in an even larger search space, where the constituent atoms and/or stoichiometric compositions of the screened systems may be variously tuned to further optimize their TE performance.