Interatomic potentials are essential for studying fundamental mechanisms of deformation and failure in metals and alloys because the relevant defects (dislocations, cracks, etc.) are far above the scales accessible to first-principles studies. Existing potentials for non-fcc metals and nearly all alloys are, however, not sufficiently quantitative for many crucial phenomena. Here machine learning in the Behler-Parrinello neural-network framework is used to create a broadly applicable potential for pure hcp magnesium (Mg). Lightweight Mg and its alloys are technologically important while presenting a diverse range of slip systems and crystal surfaces relevant to both plasticity and fracture that present a significant challenge for any potential. The machine learning potential is trained on first-principles density-functional theory (DFT) computable metallurgically relevant properties and is then shown to well predict metallurgically crucial dislocation and crack structures and competing phenomena. Extensive comparisons to an existing very good modified embedded atom method potential are made. These results demonstrate that a single machine learning potential can represent the wide scope of phenomena required for metallurgical studies. The DFT database is openly available for use in any other machine learning method. The method is naturally extendable to alloys, which are necessary for engineering applications but where ductility and fracture are controlled by complex atomic-scale mechanisms that are not well predicted by existing potentials.