Atomistic Graph Neural Networks for metals: Application to bcc iron

Abstract

The prediction of the atomistic structure and properties of crystals including defects based on ab-initio accurate simulations is essential for unraveling the nano-scale mechanisms that control the micromechanical and macroscopic behaviour of metals. Density functional theory (DFT) can enable the quantum-accurate prediction of some of these properties, however at high computational costs and thus limited to systems of ~1,000 atoms. In order to predict with quantum-accuracy the mechanical behaviour of nanoscale structures involving from thousands to several millions of atoms, machine learning interatomic potentials have been recently developed. Here, we explore the performance of a new class of interatomic potentials based on Graph Neural Networks (GNNs), a recent field of research in Deep Learning. Two state-of-the-art GNN models are considered, SchNet and DimeNet, and trained on an extensive DFT database of ferromagnetic bcc iron. We find that the DimeNet GNN Fe potential including three-body terms can reproduce with DFT accuracy the equation of state and the Bain path, as well as defected configurations (vacancy and surfaces). To the best of our knowledge, this is the first demonstration of the capability of GNN of reproducing the energetics of defects in bcc iron. We provide an open-source implementation of DimeNet that can be used to train other metallic systems for further exploration of the GNN capabilities.