The 9th International Conference on Multiscale Materials Modeling

講演情報

Poster Session

D. Data-Driven and Physics-Informed Materials Discovery and Design

[PO-D2] Poster Session 2

Symposium D

2018年10月31日(水) 17:45 〜 20:00 Poster Hall

[P2-32] Development of artificial neural network model for prediction of electronic density of states in atomistic systems

Atsushi Kubo, Yoshitaka Umeno (Institute of Industrial Science, the University of Tokyo, Japan)

Recently, the artificial neural network (ANN) model has been intensively applied to interatomic potentials for atomistic simulations. ANN-based potential functions possess basic characteristics; i.e., (1) they do not have any physical background besides the least geometrical conditions, e.g., physical quantity conservation against any coordinate transformation; (2) theoretically they can mimic any continuous functions. These features make ANN potentials applicable to complex atomistic systems, where various crystal structures and phases are relevant.
Since an ANN potential can be simply regarded as a mapping from atomistic structure to a real number, it is found that the application of ANNs does not have to be limited only for prediction of potential energy; i.e., ANNs are applicable to prediction of other physical quantities or material properties of the atomistic systems, e.g., electronic density of states (DoS), magnetic moment, etc. However, to the best of our knowledge, there has been no attempt to apply ANNs to prediction of physical quantities beside potential energy in atomistic systems. It will be of great impact if we can evaluate physical quantities such as electric or magnetic properties in huge atomistic structures using ANNs with the accuracy of the first-principles calculation.
In this study, we developed an ANN model to predict the DoS, as an exemplified case of prediction of general physical quantities in atomistic systems with ANNs. The ANN was optimized to reproduce the DoS in various crystal structures of silicon-carbon system that was obtained by first-principles calculations based on the density functional theory.