Overcoming Inaccuracies\nin Machine Learning Interatomic\nPotential Implementation for Ionic Vacancy Simulations

Machine learning interatomic potentials, particularly\nones based\non deep neural networks, have taken significant strides in accelerating\nfirst-principles simulations, expanding the length and time scales\nof the simulations with accuracies akin to first-principles simulations.\nNotwithstanding their success in accurately describing the physical\nproperties of pristine ionic systems with multiple oxidation states,\nherein we show that an implementation of deep neural network potentials\n(DNPs) yield vacancy formation energies in MgO with a significant\n∼3 eV error. In contrast, we show that moment tensor potentials\ncan accurately describe all properties of the oxide, including vacancy\nformation energies. We show that the vacancy formation energy errors\nin DNPs correlate with the strength of ionic interactions in the system\nas evidenced by contrasting MgO with the less ionic systems Cu<sub><i>x</i></sub>O<sub><i>y</i></sub> and Ag<sub><i>x</i></sub>O<sub><i>y</i></sub>. Our findings\nsuggest that descriptors employed in the DNP may be inadequate and\ncannot accurately describe vacancies in ionic systems.