Recent application of neural networks (NNs) to modeling interatomicinteractions has shown the learning machines' encouragingly accurateperformance for select elemental and multicomponent systems. In this study, weexplore the possibility of building a library of NN-based models by introducinga hierarchical NN training. In such a stratified procedure NNs formulticomponent systems are obtained by sequential training from the bottom up:first unaries, then binaries, and so on. Advantages of constructing NN setswith shared parameters include acceleration of the training process and intactdescription of the constituent systems. We use an automated generation ofdiverse structure sets for NN training on density functional theory-levelreference energies. In the test case of Cu, Pd, Ag, Cu-Pd, Cu-Ag, Pd-Ag, andCu-Pd-Ag systems, NNs trained in the traditional and stratified fashions arefound to have essentially identical accuracy for defect energies, phonondispersions, formation energies, etc. The models' robustness is furtherillustrated via unconstrained evolutionary structure searches in which the NNis used for the local optimization of crystal unit cells.
展开▼