Machine learning for accelerated prediction of the Seebeck coefficient at arbitrary carrier concentration

Abstract

Listen

Translate article

The convenient determination of Seebeck coefficient is a major challenge in the thermoelectric field, either for experimental metrology or theoretical prediction. Taking Heusler compounds as a prototypical class of examples, we propose a neural network model by which the Seebeck coefficient can be quickly evaluated at arbitrary carrier concentration. Using only a few elemental properties of the constituent atoms as input features, the derived model exhibits a high Pearson correlation between the real and predicted Seebeck coefficients for both the n- and p-type systems. Beyond the training data, we have checked the Seebeck coefficients of four randomly selected Heusler compounds and found strong predictive power of our model, which can be in turn utilized to screen a vast number of Heusler compounds for accelerated discovery of new materials with desired Seebeck coefficients. As an addition, we adopt the sure independence screening and sparsifying operator to obtain a three-dimensional descriptor, which is physically interpretable and shows comparable predictive power as our neural network model.