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.

In this study, we benchmarked various interatomic potentials and force fields in comparison to an ab initio dataset for bulk amorphous alumina. We investigated a comprehensive set of fixed-charge and variable-charge potentials tailored for alumina. We also train a machine learning interatomic potential, using the NequIP framework. Results highlight that the fixed-charge potential by Matsui provides an ideal blend of computational speed and alignment with ab initio findings for stoichiometric alumina. For non-stoichiometric variants, the variable charge potentials, especially ReaxFF, align remarkably well with DFT outcomes. The NequIP ML potential, while superior in some instances and adaptable, might not be the best fit for specific tasks.

在本研究中，我们对各种原子间势和力场进行了基准测试，并将其与用于致密非晶态氧化铝的从头算数据集进行了比较。我们调查了一组针对氧化铝的固定电荷和可变电荷势。此外，我们利用NequIP框架训练了一种机器学习原子间势。结果表明，Matsui提出的固定电荷势在计算速度与从头算结果对致氧铝的化学计量比方面提供了理想的结合。对于非化学计量比变体，可变电荷势，尤其是ReaxFF，与DFT结果高度吻合。尽管NequIP机器学习势在某些情况下表现出色且具有适应性，但它可能并非特定任务的最佳选择。