pith. sign in

arxiv: 2604.13659 · v3 · pith:HCBNR4AYnew · submitted 2026-04-15 · ⚛️ physics.chem-ph · cond-mat.soft

Ion-Specific Anomalous Water Diffusion in Aqueous Electrolytes: A Machine-Learned Many-Body Force Field Study with MACE

Massimo Ciacchi , Ilnur Saitov , Nico Di Fonte , Isabella Daidone , Carlo Pierleoni This is my paper

Pith reviewed 2026-05-10 12:38 UTC · model grok-4.3

classification ⚛️ physics.chem-ph cond-mat.soft
keywords aqueous electrolyteswater diffusionmachine-learned force fieldsion-specific effectsmolecular dynamicshydration shellsanomalous diffusionMACE
0
0 comments X

The pith

A many-body machine-learned force field trained on DFT data reproduces the ion-specific anomaly in water diffusion for NaCl and CsI solutions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that classical molecular dynamics with a MACE equivariant neural network force field, fitted to revPBE-D3 energies forces and stresses, produces the observed increase in water diffusion in CsI solutions and the decrease in NaCl solutions across a range of concentrations. This matches experiment more closely than earlier DeePMD models on the same underlying theory, especially for NaCl. The work supplies a shell-by-shell decomposition that links the retardation to Na+ hydration shells and the acceleration to the loose I- shell. A reader should care because the result supplies a concrete microscopic route to the long-standing chaotrope-kosmotrope distinction without relying on empirical pair potentials.

Core claim

The MACE many-body force field reproduces the experimentally observed anomalous diffusion of water: diffusion coefficients rise with CsI concentration yet fall with NaCl concentration. The improvement over prior DeePMD results is traced to a stronger Na+-water interaction in the first shell together with a measurable retarding effect from the second shell. In CsI the acceleration is shown to originate mainly from the anion, whose diffuse hydration shell permits fast water exchange with bulk liquid.

What carries the argument

The MACE equivariant graph neural network trained on DFT energies, forces, and stresses to generate a many-body classical force field for aqueous electrolytes.

If this is right

  • The same MACE-trained force field yields a coherent microscopic picture in which Na+ retards water through its first and second hydration shells while I- accelerates water through rapid exchange.
  • Shell-decomposition of time-dependent diffusivities and ion-oxygen potentials of mean force directly attributes the ion-specific sign of the anomaly.
  • Quantitative improvement is obtained for both salts, with the largest gain for NaCl relative to earlier many-body models trained on the same theory.
  • The acceleration-retardation mechanism holds across the concentration window 0.89-3.56 mol/kg at ambient conditions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same training protocol could be used to predict diffusion anomalies for other salt pairs whose experimental behavior is already known, providing a direct test of transferability.
  • Extension to concentrated or high-pressure regimes would clarify whether the second-shell retardation of Na+ remains dominant when ion pairing becomes frequent.
  • The shell-resolved analysis offers a practical route to parameter-free estimates of viscosity or conductivity in mixed electrolytes once the force field is available.

Load-bearing premise

The revPBE-D3 functional supplies an accurate enough description of the many-body interactions that the trained force field can be trusted for water dynamics in these salt solutions.

What would settle it

If the computed water diffusion coefficient at 1.78 mol/kg NaCl deviates from the experimental value by more than the combined statistical and experimental uncertainty, the central claim of quantitative reproduction would be falsified.

Figures

Figures reproduced from arXiv: 2604.13659 by Carlo Pierleoni, Ilnur Saitov, Isabella Daidone, Massimo Ciacchi, Nico Di Fonte.

Figure 1
Figure 1. Figure 1: FIG. 1: Density as a function of concentration (EoS) at [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: Radial distribution functions (RDFs) involving the oxygen as a function of concentration for NaCl and CsI [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Ion-ion radial distribution functions (RDFs) as a function of concentration. Systems with 250 water [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Total neutron scattering structure factor [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: Reduced structure factor for NaCl (upper [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: Relative diffusion dependence on the [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8: Panel (a): Values of relative diffusion at 2.5 ps for total first solvation shells (pure first solvation shells + [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9: Potential of Mean Force: concentration [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: FIG. 11: Oxygen RDF comparison between DFT [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10: RDF comparison between DFT, M1 and [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 13
Figure 13. Figure 13: FIG. 13: (a) Velocity Autocorrelation Function, (b) [PITH_FULL_IMAGE:figures/full_fig_p011_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: FIG. 14: Comparison between DFT and FT-M1 Model [PITH_FULL_IMAGE:figures/full_fig_p011_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: FIG. 15: Oxygen-Oxygen RDFs comparison with Avula [PITH_FULL_IMAGE:figures/full_fig_p012_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: FIG. 16: Radial distribution function comparison [PITH_FULL_IMAGE:figures/full_fig_p012_16.png] view at source ↗
Figure 19
Figure 19. Figure 19: FIG. 19: The dependence of the diffusion coefficient of [PITH_FULL_IMAGE:figures/full_fig_p013_19.png] view at source ↗
Figure 18
Figure 18. Figure 18: FIG. 18: The dependence of the diffusion coefficient of [PITH_FULL_IMAGE:figures/full_fig_p013_18.png] view at source ↗
read the original abstract

The dynamics of water in electrolyte solutions exhibits a striking, ion-specific anomaly: the diffusion coefficient of water is enhanced relative to the neat liquid in chaotropic CsI solutions, yet suppressed in kosmotropic NaCl solutions. This phenomenon, long challenging for classical force-field-based molecular dynamics, is studied here using classical molecular dynamics simulations with a many-body machine-learned force field (MLFF) trained within the MACE equivariant graph neural network framework. The force field is trained on energies, forces, and stresses computed at the density functional theory level with the revPBE-D3 exchange--correlation functional, which provides a reliable balance between accuracy and computational efficiency for aqueous systems. Simulations of NaCl and CsI aqueous solutions at ambient conditions over a concentration range of 0.89--3.56 mol/kg reproduce the experimentally observed anomalous diffusion and show a quantitative improvement over previous results obtained with the DeePMD framework, trained on the same theory, particularly for NaCl solutions. This improvement is traced to a stronger Na$^{+}$--water interaction in the first hydration shell and the non-negligible retarding contribution of the second hydration shell of Na$^{+}$. For CsI solutions, the water acceleration is shown to be primarily driven by the anion I$^{-}$, whose diffuse and weakly structured hydration shell facilitates rapid water exchange with the bulk. These results are rationalised through a shell-decomposition analysis of time-dependent water diffusivities and ion--oxygen potentials of mean force providing a coherent microscopic picture of the acceleration--retardation mechanism in the studied aqueous electrolytes.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper uses classical MD simulations with a MACE equivariant graph neural network force field trained on revPBE-D3 DFT energies, forces, and stresses to study water diffusion in NaCl and CsI aqueous solutions (0.89--3.56 mol/kg). It claims reproduction of the experimental ion-specific anomaly (suppressed diffusion in NaCl, enhanced in CsI), quantitative improvement over prior DeePMD results on the same reference theory, and a microscopic mechanism via shell-decomposition of time-dependent diffusivities and ion--oxygen PMFs.

Significance. If the quantitative reproduction holds, the work shows that MACE can capture many-body interactions in electrolyte solutions more accurately than DeePMD for this observable, yielding a coherent shell-based rationale for acceleration versus retardation that classical force fields have struggled to reproduce. This advances MLFF methodology for condensed-phase dynamics in physical chemistry.

major comments (3)
  1. [Abstract and Results] Abstract and Results section on diffusion coefficients: the claim of 'quantitative improvement' over DeePMD (particularly for NaCl) is stated without reported statistical uncertainties, error bars, convergence tests with respect to simulation length or system size, or direct side-by-side metrics (e.g., mean absolute deviations from experiment). This leaves the strength of the central reproduction claim difficult to evaluate.
  2. [Methods] Methods section (DFT functional choice): revPBE-D3 is selected for 'reliable balance between accuracy and computational efficiency,' yet no benchmarking against experimental diffusion coefficients, higher-level methods (e.g., hybrid functionals or MP2), or sensitivity tests is provided. Given that GGA functionals can err by 20% or more on hydrogen bonding and ion solvation, this choice is load-bearing for the ion-specific claims.
  3. [Results] Results section (shell-decomposition analysis): the attribution of NaCl retardation to first- and second-shell Na+--water interactions and CsI acceleration to the diffuse I- shell is extracted from the trajectories, but without quantitative decomposition of diffusivity contributions or error estimates on the PMFs and time-dependent diffusivities, the mechanistic picture remains qualitative.
minor comments (2)
  1. [Abstract and Methods] The concentration units (mol/kg) are clear but the manuscript would benefit from also reporting the corresponding molar (mol/L) values at ambient conditions for direct comparison with some experimental datasets.
  2. [Figures] Figure captions for diffusion plots should explicitly state the number of independent trajectories and total sampling time used to compute each D value.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful and constructive review. The comments have prompted us to strengthen the statistical rigor and quantitative aspects of the analysis. We address each major comment below and have revised the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract and Results] Abstract and Results section on diffusion coefficients: the claim of 'quantitative improvement' over DeePMD (particularly for NaCl) is stated without reported statistical uncertainties, error bars, convergence tests with respect to simulation length or system size, or direct side-by-side metrics (e.g., mean absolute deviations from experiment). This leaves the strength of the central reproduction claim difficult to evaluate.

    Authors: We agree that the absence of uncertainties and direct metrics weakens the claim. In the revised manuscript we report diffusion coefficients with standard errors obtained from five independent 10 ns trajectories per system using block averaging. We have added convergence tests demonstrating that values stabilize for simulation lengths beyond 5 ns and system sizes above ~1000 water molecules. A new table in the Results section (and Supplementary Information) provides mean absolute deviations from experiment for both MACE and DeePMD on the same reference data, confirming the improvement for NaCl while remaining comparable for CsI. revision: yes

  2. Referee: [Methods] Methods section (DFT functional choice): revPBE-D3 is selected for 'reliable balance between accuracy and computational efficiency,' yet no benchmarking against experimental diffusion coefficients, higher-level methods (e.g., hybrid functionals or MP2), or sensitivity tests is provided. Given that GGA functionals can err by 20% or more on hydrogen bonding and ion solvation, this choice is load-bearing for the ion-specific claims.

    Authors: We acknowledge the load-bearing nature of the functional choice. The revised Methods section now cites prior benchmarking literature supporting revPBE-D3 for water structure and ion solvation, and we have added a brief sensitivity discussion comparing selected radial distribution functions to results obtained with other GGAs in the literature. Full re-training and benchmarking against hybrids or MP2 for the entire dataset is computationally prohibitive within the scope of this study; we have therefore added an explicit limitations paragraph and flagged this as an important direction for future work. revision: partial

  3. Referee: [Results] Results section (shell-decomposition analysis): the attribution of NaCl retardation to first- and second-shell Na+--water interactions and CsI acceleration to the diffuse I- shell is extracted from the trajectories, but without quantitative decomposition of diffusivity contributions or error estimates on the PMFs and time-dependent diffusivities, the mechanistic picture remains qualitative.

    Authors: We accept that the original analysis was largely qualitative. We have now performed a quantitative shell-resolved decomposition of the water diffusivity anomaly, reporting the fractional contribution (with bootstrap-derived standard errors) from the first and second shells of each ion. Error bars obtained via block bootstrapping have been added to all PMF and time-dependent diffusivity plots. These results are presented in a new figure and accompanying text that make the mechanistic attribution substantially more quantitative. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper trains a MACE MLFF on energies/forces/stresses from independent DFT calculations using revPBE-D3, then performs MD simulations to compute water diffusion coefficients as a function of concentration for NaCl and CsI. These computed diffusivities are compared to experiment and prior DeePMD results on the same reference theory; they are not used as training targets or fitted parameters. Mechanistic explanations (shell retardation, anion-driven acceleration) are extracted via post-hoc analysis of trajectories and PMFs. No step reduces by construction to its inputs, no self-definitional relations appear, and no load-bearing self-citations are invoked to justify the central reproduction claim.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard assumptions of DFT accuracy for water and ions plus the transferability of the trained MLFF; no new free parameters or postulated entities are introduced beyond the neural-network weights.

axioms (2)
  • domain assumption revPBE-D3 DFT provides reliable energies, forces, and stresses for aqueous electrolyte systems at ambient conditions.
    Explicitly stated as the source of training data for the MACE model.
  • domain assumption Classical molecular dynamics with the resulting MLFF faithfully reproduces real-time water diffusion coefficients in the studied concentration range.
    Required for the claim that simulations reproduce the experimental anomaly.

pith-pipeline@v0.9.0 · 5610 in / 1364 out tokens · 35030 ms · 2026-05-10T12:38:01.955068+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.