Deep learning interatomic potential for Ca-O system at high pressure

Calcium-containing oxides are fundamental components of the Earth's crust and mantle. Analysis of their structural behavior contributes to our understanding of the Earth's interior, which needs a reliable interatomic potential. Here, we present an interatomic potential for the Ca-O system using the deep neural network method. The initial training data set for the deep interatomic potential (DP) consists of snapshots of prototype binary structures from materials projects and crystal-derived structures based on Ca-O binary compounds, as well as their molecular dynamic trajectories. The accuracy of the DP is evaluated by predicting forces and energies in comparison with those from ab initio calculations. We demonstrate that the vibrational and thermodynamic properties based on DP calculations are in excellent agreement with those from ab initio calculations. Besides, we also construct a temperature-pressure phase diagram of Ca-O compounds with DP at a lower cost compared to ab initio methods. Finally, we use the DP to explore Ca-O structures by combining it with the genetic algorithm, the accuracy of which is validated by a principal component analysis for the local atomic environments. The DP training procedure used in this work is equally applicable to other systems for accurate atomistic simulations as an effective method.

Automated summary

This study presents an interatomic potential for the Ca-O system using deep neural network method, demonstrating its accuracy in predicting forces and energies. It also shows that the potential can be used to construct the temperature-pressure phase diagram of Ca-O compounds at a lower cost compared to ab initio calculations, and explore Ca-O structures by combining with genetic algorithm.