Predicting doped thermoelectric properties with multitask attention

Abstract

Thermoelectric (TE) materials, capable of converting temperature gradients into electricity, and vice versa, emerge as a promising class of sustainable materials because of their pollution-free operation. However, their efficiency remains lower than that of conventional heat engines and pumps, limiting their applications. Thus, much research is geared towards discovering high performing TE materials. One strategy pertains to impurity doping, which have the potential to drastically augment material property despite small amounts of elements added. Experimentally, it is not feasible to synthesize every possible doped material due to the large chemical space involved, necessitating an alternative procedure. Recently, machine learning (ML) has emerged as a powerful tool to accelerate property prediction and the discovery of new materials. Yet, it is challenging to predict doped material properties solely from composition. Typical ML featurization techniques rely on stoichiometric prevalence, resulting in doped materials having similar vectors to their pure forms. Consequently, typical ML models struggle to predict the complex, nonlinear effects of dopants. This work addresses these limitations by enhancing the predictive accuracy of 7 key TE transport property prediction, via the modification of the Compositionally restricted attention-based Network (CrabNet). CrabNet predicts properties of compositions using the attention mechanism, which can be leveraged to learn dopant-host interactions implicitly. First, a comprehensive experimental TE dataset is collated from recent literature, providing a source of high-fidelity data. Second, by utilizing multitask learning to exploit the interdependence of different TE transport properties, and encoding temperature information, the modified CrabNet model demonstrates improved prediction accuracy over conventional and existing ML models geared towards TE property prediction.