Plasticity and fracture of zinc are complex and highly anisotropic, because of the unusually large c/a ratio of the hexagonal-closed-packed (HCP) structure. Traditional semi-empirical interatomic potentials fail to capture both HCP stability and c/a ratio, thus they are not suitable to study plastic deformation. Recent machine learning-based potentials, including Rapid Artificial Neural Network (RANN) and Moment Tensor Potential (MTP), address this issue, however they struggle to accurately reproduce the Generalized Stacking Fault Energy (GSFE) surfaces, which control dislocation structure and glide. Also, these potentials do not accurately reproduce the basal traction-separation (T-S) curve, which determines crack propagation. In this work, we develop an Atomic Cluster Expansion (ACE) potential for Zn, trained on an extensive and well-converged DFT database. The training data is optimized using the recent HyperActive Learning (HAL) algorithm, enabling to achieve very low RMSE, accurate phonons for molecular dynamics simulations, and good transferability, assessed using model uncertainty measures, to extended defects such as dislocations and cracks. The validated potential is then used to clarify key observations of crack and dislocation slip behaviour in Zn. First, basal slip activates at a critical resolved shear stress below 0.5 MPa, consistent with single crystal experiments. Second, we find that pyramidal II slip is the next easier slip system, and it shows a pronounced compression/tension asymmetry in line with some experimental findings. Finally, we reveal why prismatic slip does not occur in Zn: screw cores are unstable and dissociate into basal dislocations, confirming previous DFT-based calculations. Our work demonstrates how carefully validated machine learning potentials can be used to unravel atomic-scale mechanisms of slip, that are beyond reach of DFT calculations.
Fioravanti et al. (Sun,) studied this question.