Ab initio methods, such as Density Functional Theory (DFT), that aim for high accuracy cannot be used to simulate more than a few hundred atoms efficiently. Therefore, they are scarcely used to simulate extended defects, such as dislocations, that require a thousand atoms or more. In [1], we have developed an algorithm that bypasses this limitation by using a machine-learning potential that is automatically trained using active learning on only some small subsets of 100-200 atoms, extracted from the large-scale configuration while running the simulation. We have applied this algorithm to simulate bcc screw dislocations and shown that the potential essentially predicts all relevant properties of the dislocation, i.e., its core structure, Peierls barrier, and Peierls stress.

In this work, we further develop this algorithm to simulate fcc dislocations. The added challenge, compared to bcc dislocations, is the dissociated structure of the dislocations in fcc materials that eliminates the possibility to do even one single-point calculation with DFT. Therefore, our algorithm extracts configurations of 100-200 atoms that contain only one of the partial dislocations. With this algorithm we are able to fully automatically run atomistic simulations of fcc dislocations of, in principle, any material. As a test problem, we show for fcc Aluminum that our MTPs accurately predict the relaxed core structure, the core energy, and the splitting between the two partial dislocations, when compared to the very expensive pure-DFT simulations from [2].

[1] M. Hodapp, A. Shapeev, Machine Learning: Science and Technology, 2020, 1.

[2] C. Woodward et al., Physical Review Letters, 2008, 100.