Do we really need machine learning interatomic potentials for modeling amorphous metal oxides? Case study on amorphous alumina by recycling an existing ab initio database

Abstract

In this study, we critically evaluate the performance of various interatomic potentials/force fields against a benchmark ab initio database for bulk amorphous alumina. The interatomic potentials tested in this work include all major fixed charge and variable charge models developed to date for alumina. Additionally, we introduce a novel machine learning interatomic potential constructed using the NequIP framework based on graph neural networks. Our findings reveal that the fixed-charge potential developed by Matsui and coworkers offers the most optimal balance between computational efficiency and agreement with ab initio data for stoichiometric alumina. Such balance cannot be provided by machine learning potentials when comparing performance with Matsui potential on the same computing infrastructure using a single Graphical Processing Unit. For non-stoichiometric alumina, the variable charge potentials, in particular ReaxFF, exhibit an impressive concordance with DFT calculations. However, our NequIP potentials trained on a small fraction of the ab initio database easily surpass ReaxFF in terms of both accuracy and computational performance. This is achieved without large overhead in terms of potential fitting and fine-tuning, often associated with the classical potential development process as well as training of standard deep neural network potentials, thus advocating for the use of data-efficient machine learning potentials like NequIP for complex cases of non-stoichiometric amorphous oxides.