What does this research mean for the field?

While on-the-fly machine learning force fields (OT-MLFF) accurately reproduce equilibrium solvation structures for aqueous cations, they fail to capture three-body angular sensitivity and exchange kinetics, gaps that are systematically narrowed by deeper, pretrained models like MACE. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To evaluate the structural, coordination, and exchange dynamics fidelity of on-the-fly and pretrained machine learning force fields for aqueous monovalent and divalent ion hydration against first-principles simulations.

September 5, 2026The Journal of Physical Chemistry BOpen Access

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

View Full Paper

Ask AI

Bookmark

Authors

MSMohammed S. SalhaQueen Mary University of London WLWuyang Lin GCGregory A. ChassQueen Mary University of London

Discussion

Member takes

Implication

Comparative simulation study demonstrates machine learning force fields reproduce ion hydration structures but misrepresent exchange dynamics, highlighting the need for improved kinetic sampling.

Key Points

To evaluate the structural, coordination, and exchange dynamics fidelity of on-the-fly and pretrained machine learning force fields for aqueous monovalent and divalent ion hydration against first-principles simulations.
Benchmarked an on-the-fly machine learning force field (OT-MLFF) using VASP kernel-based Bayesian regression and the pretrained equivariant MACE foundation model.
Compared performance against ab initio molecular dynamics (AIMD; CP2K, PBE-D3) and classical molecular dynamics (Madrid-2019/TIP4P-2005) across Li+, Na+, K+, Cs+, Mg2+, and Ca2+ aqueous solutions.
Both machine-learned potentials accurately captured bulk water structure and first-shell metal–oxygen distances, outperforming classical molecular dynamics and placing the Ca2+ coordination number at approximately 6.
OT-MLFF displayed kinetic failures by generating spurious first-shell water exchange for Mg2+, completely suppressing exchange in dynamic Ca2+ shells, and smoothing three-body angular features.
The pretrained MACE foundation model systematically narrowed these discrepancies without system-specific training, recovering the dynamic and angular behaviors observed in first-principles simulations.

Cite This Study

Salha et al. (2026) studied this question.

synapsesocial.com/papers/6a9bd49d6b95aff0620ec693 https://doi.org/10.1021/acs.jpcb.6c02837

View Full Paper

Ask AI

Bookmark

Implication

Comparative simulation study demonstrates machine learning force fields reproduce ion hydration structures but misrepresent exchange dynamics, highlighting the need for improved kinetic sampling.

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

Discussion

Member takes

Implication

Key Points

Cite This Study

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

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider