Stratified construction of neural network based interatomic models for multicomponent materials

TL;DR

This paper introduces a hierarchical, stratified training approach for neural network interatomic models in multicomponent materials, improving training efficiency while maintaining accuracy across various properties.

Contribution

It proposes a novel stratified training method for neural network interatomic models, enabling efficient and accurate modeling of multicomponent systems with shared parameters.

Findings

Stratified NNs achieve similar accuracy to traditional models.

Shared parameter models accelerate training.

Models are robust in structure searches.

Abstract

Recent application of neural networks (NNs) to modeling interatomic interactions has shown the learning machines' encouragingly accurate performance for select elemental and multicomponent systems. In this study, we explore the possibility of building a library of NN-based models by introducing a hierarchical NN training. In such a stratified procedure NNs for multicomponent systems are obtained by sequential training from the bottom up: first unaries, then binaries, and so on. Advantages of constructing NN sets with shared parameters include acceleration of the training process and intact description of the constituent systems. We use an automated generation of diverse structure sets for NN training on density functional theory-level reference energies. In the test case of Cu, Pd, Ag, Cu-Pd, Cu-Ag, Pd-Ag, and Cu-Pd-Ag systems, NNs trained in the traditional and stratified fashions are found to have essentially identical accuracy for defect energies, phonon dispersions, formation energies, etc. The models' robustness is further illustrated via unconstrained evolutionary structure searches in which the NN is used for the local optimization of crystal unit cells.