Magnesium ions play essential roles in many biological processes. Yet, despite their apparent simplicity, they remain remarkably challenging to describe accurately in all- atom molecular dynamics simulations. Over the past years, we have systematically optimized classical ion force fields against experimental data, substantially improving their description of structural, thermodynamic, and kinetic properties. Still, Mg2+ force fields fail to simultaneously reproduce all key experimental properties, likely reflecting their inability to explicitly account for quantum many-body effects.

Machine-learned interatomic potentials provide a promising route toward more accurate simulations of ions in aqueous solution. Here, we develop MACE neural network potentials for aqueous MgCl2 solutions trained on revPBE-D3/zd and revPBE0-D3/zd density functional theory reference data and systematically evaluate their ability to reproduce experimental structural, dynamical, and thermodynamic properties.

Both neural network potentials accurately reproduce the octahedral hydration structure of Mg2+. Transition path-sampling simulations reveal a dissociative water-exchange mechanism via a five-coordinated intermediate, in agreement with experiment, and yield exchange rates within one order of magnitude of experimental values. In contrast, the solvation free energy is significantly underestimated, highlighting limitations of current local models and the need for explicit long-range electrostatic treatments.