Benchmarking on-the-Fly Machine Learning Force Fields for Ion Hydration: Structure, Coordination, and Exchange Dynamics across Monovalent and Divalent Cations.

Machine learning force fields (MLFFs) promise ab initio accuracy at a fraction of the computational cost of first-principles simulation, yet their fidelity for aqueous ion hydration across chemically diverse cation series remains systematically underexplored. Here, we benchmark an on-the-fly MLFF (OT-MLFF) framework, implemented in VASP via kernel-based Bayesian regression with adaptive DFT calls triggered when the estimated force error exceeds a set threshold, together with the pretrained, equivariant MACE foundation model applied off-the-shelf, against ab initio molecular dynamics (AIMD; CP2K, PBE-D3) and classical MD (CMD; Madrid-2019/TIP4P-2005) for aqueous solutions of Li+, Na+, K+, Cs+, Mg2+, and Ca2+. Across all six ions, the OT-MLFF reproduces a bulk water structure and the systematic increase of first-shell M-O distances down both cation series with near-AIMD accuracy, outperforming CMD. Coordination-number distributions reveal a mild OT-MLFF overcoordination of Li+ (modal CN 5 versus 4 in both AIMD and MACE) and Cs+, whereas the two divalent shells, including the previously problematic Ca2+, are reproduced by both machine-learned potentials, all three methods placing the Ca2+ coordination number at ≈6. MACE tracks AIMD most closely overall, its principal bias being a residual overcoordination of K+ and undercoordination of Cs+ in the diffuse monovalent shells. Angular distribution functions confirm accurate reproduction of the rigid octahedral geometry of Mg2+ but reveal systematic OT-MLFF smoothing of three-body features for more labile shells, including the O-Na-O shoulder near 65-70° and the O-Ca-O peak angle. The largest failures are kinetic: the OT-MLFF generates spurious first-shell water exchange for Mg2+, absent in AIMD, and completely suppresses the exchange of the dynamic Ca2+ shell, even though it reproduces the equilibrium coordination number of both ions, a clear dissociation between structural and dynamical fidelity. Velocity autocorrelation functions reveal a related failure, the OT-MLFF underestimating the cooperative cage confinement of the large monovalent ions Cs+ and K+. In each of these cases, the pretrained MACE model recovers the AIMD behavior that the on-the-fly potential misses. Collectively, these results establish that the OT-MLFF is robust for an equilibrium solvation structure but fails for three-body angular sensitivity and exchange kinetics in chemically distinct edge cases and that a deeper, higher-body-order, pretrained model such as MACE systematically narrows these gaps even without system-specific training. Targeted improvements such as augmented transition-state sampling, higher angular-resolution descriptors, and long-range electrostatic corrections are identified as high-priority directions for kinetic fidelity in next-generation electrolyte force fields.