REVIEW 3 major objections 4 minor 3 references
Comparison of water models for structure prediction
T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read This paper argues that a simple rigid four-site model of water, TIP4P/2005, reproduces experimental liquid structure more accurately than more complex flexible, polarizable, or five-to-seven-site models over the full 254–366 K range.
desk verdict A useful, well-run benchmark of 44 water models; the D2O-for-H2O neutron reference is the main caveat but does not sink the central conclusion. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing comparison is the total scattering structure factor $S(Q)$, computed from simulated partial radial distribution functions $g_{ij}(r)$ by a weighted Fourier transform, combined with the goodness-of-fit measure $R$ normalized per data set to a relative R-factor $R_{rel} = R/R_{best}$ and summed into $R_{tot}$. This combined metric is what lifts four-site models to the top: because the X-ray and neutron weights emphasize different partials, fitting both data types simultaneously is a stricter test than fitting either alone.
What would settle it
Recompute the combined ranking using experimental light-water neutron total scattering structure factors, such as those from H/D isotopic substitution measurements, instead of heavy-water data at 295 K; if the top-17 ordering changes by more than the current 10% spread, the paper's conclusion about four-site dominance would be overturned.
Extended reading notes
Core claim
Forty-four classical pairwise-additive water models were simulated under an identical protocol; trajectories produced partial radial distribution functions, and X-ray and neutron weighted total scattering structure factors were compared with experimental data. The paper's central conclusion is that on the combined relative R-factor $R_{tot}$ (average X-ray relative R-factor plus average neutron relative R-factor), TIP4P/2005 ranks highest over the full temperature range, with TIP4P/ε, TIP4Q, TIP4P/2005f, TIP4P-FB, and other four-site models in close succession. The top 17 models span less than 10% in $R_{tot}$, so they are statistically comparable to one another. More complex models, including five-, six-, and seven-site, flexible, polarizable, and Buckingham-potential models, do not show a significant structural advantage; the worst performers include several polarizable models and two models with poor density. The paper further shows that the OPC model, though best for X-ray data alone, fails neutron data largely because its intramolecular geometry differs from gas-phase water geometry.
Load-bearing premise
The neutron-diffraction leg of the ranking uses heavy-water (D2O) experimental data as a stand-in for simulated light-water (H2O) models, without an isotope correction or sensitivity analysis.
Editorial extensions
If this is right
- Four-site rigid models are sufficient for pure-liquid-water structure, so simulations needing structural accuracy can use TIP4P/2005 without paying the computational cost of polarizability or flexibility.
- The spread of less than 10% among the top 17 models means many cheap models give statistically indistinguishable structural fits, allowing cost to guide selection within that group.
- Recent three-site models such as OPC3 and OPTI-3T are nearly as accurate as the best four-site models, providing an even cheaper alternative for large-scale simulations.
- Model developers should validate against both neutron and X-ray structure factors rather than only the oxygen-oxygen distance, because the OPC case shows that intramolecular geometry can spoil neutron agreement.
- For pure water at ambient pressure, adding interaction sites or polarizable terms is not justified by structural accuracy alone.
Reading between the lines
- If heavy-water and light-water structures differ by more than the roughly 10% spread separating the top 17 models, the neutron leg of the ranking could shift; using light-water neutron data from H/D isotopic substitution would test this directly.
- The temperature-shift behavior noted for TIP3P and OPC3 suggests that part of the apparent model error is a shifted temperature scale, so aligning models by effective temperature could change rankings at the edges of the 254–366 K window.
- The same $R_{tot}$ protocol could benchmark machine-learned and other advanced water potentials against the same experimental data, quantifying whether their added cost buys structural improvement.
- The paper's conclusion is property-specific: for thermodynamic, dynamic, or other non-structural properties, more complex models may still be needed.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript reports molecular dynamics simulations of 44 classical water models at seven temperatures (254–366 K). It computes partial radial distribution functions and total scattering structure factors, compares them against X-ray diffraction data for H2O (Skinner et al.) and neutron diffraction data for D2O (Soper; Ohtomo et al.), and introduces relative R-factors to rank the models. The central claims are that models with more than four interaction sites, and flexible or polarizable models, do not provide significant structural advantages, and that TIP4P-type four-site models, especially TIP4P/2005, give the best overall agreement, with recent three-site models (OPC3, OPTI-1T, OPTI-3T) being competitive.
Significance. If the central claim holds, the paper provides a useful benchmark for model selection in classical simulations of liquid water. Its strengths are a broad and uniform simulation protocol across 44 models, direct comparison with experimental TSSFs rather than derived PRDFs, cross-checks of density and self-diffusion against literature values, and full tabulation of R-factors and structural data. The work fits no new parameters: all model parameters come from prior literature, and the experimental data are external benchmarks. However, the main ranking rests on a combined XRD/ND comparison whose ND leg uses D2O reference data for H2O simulations, and the reported R-factors carry no statistical uncertainties; both issues need to be addressed before the ranking claims can be considered robust.
major comments (3)
- [Section 2.2, Eq. (2), and Section 4.3.2] The neutron-diffraction leg of the comparison uses experimental D2O structure factors (Soper 283/295 K and Ohtomo et al. 298–368 K) as the reference for simulated H2O models, with deuterium scattering lengths in the weighting of Eq. (2). No isotope correction or sensitivity analysis is provided. D2O is more strongly ordered than H2O, so the reference curve is systematically shifted relative to the H2O target, and because the top 17 models in Rtot (Section 4.3.3, Table S6) are separated by less than 10%, a few-percent isotope offset could reorder the top cluster. Please either simulate D2O with the leading models to quantify the isotope effect, or add a sensitivity test that bounds the D2O/H2O contribution to the ND R-factors.
- [Section 4.3.3 and Tables 3, 4, S6] R-factors are reported without statistical uncertainties, yet the text states that differences among the top 17 models are 'not significant' and that any of them can be reliably used for structural analysis. Adjacent ranks in Table S6 (e.g., TIP4P/2005 with Rtot 2.50 and TIP4P/ε with 2.52) differ by less than 1%, and no estimate of the noise floor is given. The authors should add error bars (for example, block averaging over independent trajectory segments) or otherwise show that the ranking is stable under plausible simulation and experimental uncertainty. Without this, the 'no significant advantage' claim for models beyond four sites is not quantitatively supported.
- [Section 4.3.3] The claim that 'using different ND dataset combinations from the three publications does not significantly affect model rankings' is not demonstrated. Table 4 shows that the best model changes with dataset (TIP4P-BG for the 284 K Soper data, TIP6P-Ew for the 295 K Soper data, and TIP4P/2005f for the Ohtomo data), so an explicit table of Rtot under alternative ND dataset combinations is needed to support the statement. This is load-bearing because Rtot determines the paper's central ranking.
minor comments (4)
- [Section 1] In the first paragraph of the introduction, 'such a model that that simultaneously reflects' contains a duplicated 'that'; please correct.
- [Table S1] The TIP6P-Ew entry lists dOH = 0.98000 nm; this is an order of magnitude larger than the TIP6P value (0.09800 nm) and is almost certainly a typo for 0.098000 nm. Please verify the parameter file and correct the table, since other groups may use these parameters.
- [Figure 12 caption] The caption for the 'mix' model says it combines the intramolecular part of TIP4P/2005 with the intramolecular part of OPC, whereas Section 4.3.3 states that the mixed PRDFs use the intermolecular part of OPC and the intramolecular part of TIP4P/2005; the caption should be corrected to match the text.
- [Section 4.3.3] The term 'PRDSs' appears to be a typo for 'PRDFs'.
Circularity Check
No significant circularity: all 44 model parameter sets are taken from prior literature, the experimental XRD and ND data are external benchmarks, and the combined R-factor ranking is a comparative metric rather than a fitted prediction.
full rationale
The paper's derivation chain is self-contained against external benchmarks. The 44 water models are taken from prior literature with no parameters fitted in this work, and the experimental X-ray and neutron diffraction curves are independent published data sets. The R-factor in Eq. (5) is a standard goodness-of-fit measure, and the relative R-factor Rrel = R/Rbest is explicitly defined as a normalization, not as a prediction derived from the data being compared. The combined ranking Rtot is a sum of such relative R-factors, so the statement that TIP4P/2005 ranks highest is a direct consequence of the external comparisons, not of any fitted quantity. The paper explicitly discloses that a few models (e.g., OPTI-1T, OPTI-3T, TIP4P-Buck) were originally optimized against structural or PRDF data, but this does not drive the central conclusion because the top performers include TIP4P/2005, which was not fitted to structure, and the authors do not present these fitted models as independent predictions. The use of D2O neutron data as a proxy for H2O is a correctness/robustness concern about reference choice, but it is not circularity: the reference is external to the model parameters and to the ranking metric. No self-citation chain, imported uniqueness theorem, or ansatz-by-citation is load-bearing in this paper. Therefore the circularity score is 0.
Assumptions & free parameters
assumptions (5)
- standard math The Fourier transform relation between partial radial distribution functions and total scattering structure factors (Eq. 1) is valid.
- domain assumption Classical effective pair-additive potentials with fixed point charges and Lennard-Jones or Buckingham repulsion-dispersion can capture the experimentally measured liquid-water structure.
- domain assumption The published experimental X-ray and neutron diffraction data used as reference are accurate on the scale of the R-factor differences being ranked.
- domain assumption Heavy-water neutron diffraction data can be compared directly with simulated light-water models without isotope correction.
- domain assumption A single simulation protocol applied uniformly to all models does not bias the structural comparison.
Cite this review
Pith. "Pith review of Comparison of water models for structure prediction." pith.science (2026). https://pith.science/paper/5EGYMSGB
@misc{pith2026250523446,
author = {Pith},
title = {Pith review of: Comparison of water models for structure prediction},
year = {2026},
howpublished = {\url{https://pith.science/paper/5EGYMSGB}},
note = {Machine review of arXiv:2505.23446}
}
read the original abstract
Describing the interactions of water molecules is one of the most common, yet critical, tasks in molecular dynamics simulations. Because of its unique properties, hundreds of attempts have been made to construct an ideal interaction potential model for water. In various studies, the models have been evaluated based on their ability to reproduce different properties of water. This work focuses on the atomic-scale structure in the liquid phase of water. Forty-four classical water potential models are compared to identify those that can accurately describe the structure in alignment with experimental results. In addition to some older models that are still popular today, new or re-parametrized classical models using effective pair-additive potentials that have appeared in recent years are examined. Molecular dynamics simulations were performed over a wide range of temperatures and the resulting trajectories were used to calculate the partial radial distribution functions. The total scattering structure factors were compared with data from neutron and X-ray diffraction experiments. Our analysis indicates that models with more than four interaction sites, as well as flexible or polarizable models with higher computational requirements, do not provide a significant advantage in accurately describing the structure. On the other hand, recent three-site models have made considerable progress in this area, although the best agreement with experimental data over the entire temperature range was achieved with four-site, TIP4P-type models.
Reference graph
Works this paper leans on
-
[1]
G.S. Kell, Density, thermal expansivity, and compressibility of liquid water from 0° to 150°C: Correlations and tables for atmospheric pressure and saturation reviewed and expressed on 1968 temperature scale, J. Chem. Eng. Data 20 (1975) 97. doi : 10.1021/je60064a005
-
[2]
D.E. Hare, C.M. Sorensen, The density of supercooled water. II. Bulk samples cooled to the homogeneous nucleation limit, J. Chem. Phys. 87 (1987) 4840 –4845. doi:10.1063/1.453710
-
[3]
M. Holz, S.R. Heil, A. Sacco, Temperature-dependent self-diffusion coefficients of water and six selected molecular liquids for calibration in accurate 1H NMR PFG measurements, Phys. Chem. Chem. Phys. 2 (2000) 4740–4742. doi:10.1039/b005319h
doi:10.1039/b005319h 2000
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.