Overcoming Inaccuracies in Machine Learning Interatomic Potential Implementation for Ionic Vacancy Simulations.

Area of Science:

• Computational Materials Science
• Materials Informatics
• Solid State Chemistry

Background:

• Machine learning interatomic potentials, especially deep neural networks (DNNs), accelerate simulations with high accuracy.
• DNNs excel in describing pristine ionic systems with multiple oxidation states.

Purpose of the Study:

• To evaluate the accuracy of deep neural network potentials (DNPs) for calculating vacancy formation energies in the ionic material MgO.
• To compare the performance of DNPs with moment tensor potentials (MTPs) for ionic systems.

Main Methods:

• Implementation and testing of deep neural network potentials (DNPs) for MgO.
• Calculation of vacancy formation energies in MgO using DNPs and MTPs.
• Analysis of DNP errors in relation to the ionic interaction strength in different oxides (MgO, CuO, AgO).

Main Results:

• DNPs exhibited a significant error of approximately 3 eV for vacancy formation energies in MgO.
• Moment tensor potentials (MTPs) accurately predicted vacancy formation energies in MgO.
• Errors in DNPs correlated with the ionic interaction strength, being larger in MgO than in less ionic Cu2O and Ag2O.

Conclusions:

• The descriptors used in current deep neural network potentials may be insufficient for accurately modeling vacancies in ionic systems.
• Moment tensor potentials demonstrate superior accuracy for describing properties, including vacancy formation energies, in ionic oxides like MgO.