Predicting thermoelectric properties from chemical formula with explicitly identifying dopant effects

Abstract Dopants play an important role in synthesizing materials to improve target materials properties or stabilize the materials. In particular, the dopants are essential to improve thermoelectic performances of the materials. However, existing machine learning methods cannot accurately predict the materials properties of doped materials due to severely nonlinear relations with their materials properties. Here, we propose a unified architecture of neural networks, called DopNet, to accurately predict the materials properties of the doped materials. DopNet identifies the effects of the dopants by explicitly and independently embedding the host materials and the dopants. In our evaluations, DopNet outperformed existing machine learning methods in predicting experimentally measured thermoelectric properties, and the error of DopNet in predicting a figure of merit (ZT) was 0.06 in mean absolute error. In particular, DopNet was significantly effective in an extrapolation problem that predicts ZTs of unknown materials, which is a key task to discover novel thermoelectric materials