Predicting Lattice Constants of Half-Heusler Alloys Through Deep Neural Network Models Using Symbols of Elements and/or Ionic Radii

In this paper, we have developed models based on Deep-Learning Neural Network (DNN) for accurately predicting lattice constants of half-Heusler alloys. A commercial software WIEN2k employing the Density Functional Theory (DFT) is first used to generate data of the lattice constants of 377 half-Heusler alloys for the training, testing and validation of the models. These models use elemental symbols or/and ionic radii as input parameters. The model that uses only symbols of the constituent element and the model that uses symbols in combination with the radii of the ions predict lattice constants of half-Heusler alloys with an average error of less than 1% to the data obtained from the WIEN2k calculations. The average error stays below 2% in the prediction by the model that uses radii of the ions alone. These results show a great promise for these models to be extended for the prediction of structural and elastic properties of not only new half-Heusler alloys but also other new materials such as full-Heusler, spinels, perovskites, etc. with greater convenience, saving hours of computation time.