Prediction of the binding energy of self interstitial atoms in alpha iron by a graph neural network

Abstract

The estimation of neutron irradiation embrittlement is crucial for the long-term operation of nuclear power plants. The jump of self-interstitial atoms is an essential process for understanding embrittlement, and their binding energy is critical to estimating the activation energy for jumping. This study presents a fast estimation method for the binding energy of self-interstitial atoms in iron using a graph neural network. The network assigns a self-interstitial atom to a node and the relationship between self-interstitial atoms to an edge. The graph neural network was trained using a large dataset of binding energies between self-interstitial atoms (∼0.5 % density) calculated using an embedded atom method interatomic potential. The network was also trained using smaller datasets of binding energies between self-interstitial atoms (∼2.0 % density) calculated using first principles calculations. The graph neural network accurately predicts binding energy with a mean absolute error of 0.05–0.06 eV and coefficient of determination (R2) of 0.9 without overfitting, using the large interatomic potential dataset. When trained on smaller first principles datasets, the network predicted binding energies with a mean absolute error of 0.15–0.16 eV and coefficient of determination (R2) of 0.6 with some degree of overfitting. The developed graph neural network provides a fast and accurate estimation method for the binding energy of self-interstitial atoms in iron.