Molecular Dynamics Simulations of $\gamma$-Belite(010)-Water Interfaces with High-Dimensional Neural Network Potentials

Bernadeta Prus; J\"org Behler

REVIEW 3 major objections 4 minor 1 cited by

Molecular Dynamics Simulations of $\gamma$-Belite(010)-Water Interfaces with High-Dimensional Neural Network Potentials

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper argues that at the γ-belite (010)–water interface, water dissociation is governed by the protonation of accessible SiO4 oxygen atoms, and that water specifically stabilizes a calcium-dimer surface defect that does not exist in va

desk verdict Solid first MLP-MD study of γ-belite-water with a convincing protonation-dissociation picture; the type II 'only in water' claim is plausible but under-supported by a single short run. read the letter →

arxiv 2512.09702 v1 pith:6D3PWCZR submitted 2025-12-10 cond-mat.mtrl-sci physics.chem-ph

Bernadeta Prus , Jörg Behler This is my paper · ORCID

classification cond-mat.mtrl-sciphysics.chem-ph

keywords belitedicalciumsilicategamma-belite(010)surfacesolid-liquidinterfacewaterdissociationneuralnetworkpotentialmoleculardynamicsdefects

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Belite, the dicalcium silicate that makes low-carbon cement possible, is well known for reacting slowly with water. This paper uses a neural network potential trained to density-functional-theory data to run nanosecond molecular dynamics of the most stable γ-belite (010) face in contact with liquid water. The authors find that the extent of water dissociation is set by protonation of exposed SiO4 oxygen atoms: essentially every proton produced by water splitting lands on a surface oxygen, while most hydroxide ions adsorb at surface calcium atoms. They also identify two structural defects on the second-most-stable T2 termination—a calcium displacement plus tetrahedron rotation that is stable even in vacuum, and a calcium dimer that forms and persists only in the presence of water. The broader claim is that water can stabilize a variety of surface structures on belite, which matters for understanding and eventually enhancing its hydration reactivity.

What carries the argument

The load-bearing tool is a high-dimensional neural network potential (HDNNP)—a machine-learned model that represents the system's energy as a sum of environment-dependent atomic energies—trained on about 24,000 density-functional-theory structures, enabling nanosecond, several-thousand-atom molecular dynamics of the solid-liquid interface. The mechanistic driver is the protonation of accessible SiO4 oxygen atoms: these oxygen sites accept the protons released by water dissociation, and the resulting hydroxide ions are captured by undercoordinated calcium atoms. The analysis mechanics include species labeling from MD trajectories (adsorbed versus solvated versus surface-bound water, hydroxide

What would settle it

Run unbiased density-functional-theory molecular dynamics, or use a neural-network potential that explicitly includes T5 and type-II-dimer configurations in its training data, on the fully dimer-covered T2 surface and on the T5 termination in water. If the type II dimers dissociate within picoseconds, or if the T5 full/empty rows remain intact instead of breaking, the paper's central stability assignments are wrong. A complementary observational check would be to look for the predicted 3.75 Å Ca-Ca distance and the bridging hydroxide-water pattern at the wet (010) surface by surface-sensitive

Watch

Extended reading notes

Core claim

The paper's central finding is that chemistry at the γ-belite (010)-water interface follows a simple proton-transfer rule: accessible oxygen atoms of surface SiO4 tetrahedra become protonated, and the degree of water dissociation is determined by how many such oxygens are exposed. On the T2 termination, essentially all 32 SiO4 tetrahedra carry a proton (32.5 on average), and about 31 of the resulting hydroxide ions adsorb at surface and subsurface calcium sites, completing their octahedral coordination; the T3 termination, with less accessible oxygens, dissociates far less water. Independently, two surface defects appear on T2: type I, a calcium displacement coupled to a SiO4 rotation, is st

Load-bearing premise

The load-bearing premise is that the neural network potential, trained on T2 and T3 interfaces, remains accurate for the structures it did not train on—especially the T5 full/empty-row termination and the water-stabilized type II calcium dimer—so the reported defect statistics and stabilities come from the true energy surface.

Editorial extensions

If this is right

The most stable T3 termination dissociates little water, while the only slightly less stable T2 termination dissociates much more because its SiO4 oxygens are accessible; termination-specific exposure, not bulk chemistry, controls initial reactivity.
The type I defect—a calcium displacement plus SiO4 rotation—is stable in vacuum and, when seeded, reconstructs the entire T2 surface and lowers its cleavage energy.
The type II defect—a 3.75 Å calcium dimer—is unstable in vacuum, becomes stable when four water molecules (two dissociated) bridge the two calcium atoms, and a fully dimer-covered T2 surface persists in aqueous molecular dynamics.
The full/empty-row T5 termination has a vacuum cleavage energy close to that of T2, but its calcium rows do not remain intact in water, suggesting that water reshapes even near-degenerate surface structures.
The persistence of all observed surface features in long simulations indicates substantial barriers to surface rearrangement, so multiple surface motifs can coexist at the belite-water interface.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same protonation rule should apply to other belite polymorphs and calcium silicates: any termination exposing non-bridging SiO4 oxygens should dissociate water more readily, offering a route to tune hydration reactivity by surface modification rather than bulk doping.
The water-stabilized type II dimer implies that surface calcium mobility on belite in water may be higher than vacuum models suggest; if so, dissolution and early hydration products could nucleate preferentially at such defect sites.
The predicted 3.75 Å Ca-Ca dimer distance and bridging hydroxide pattern are concrete fingerprints that surface-sensitive experiments—grazing-incidence diffraction or vibrational spectroscopy of wet belite—could search for.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The paper is worth engaging with. It is the first MLP-based MD study of γ-belite(010)-water interfaces, and it delivers a clear mechanistic picture: water dissociation at the T2 termination is driven by protonation of accessible SiO4 oxygens, with most hydroxides staying at surface calcium sites, while T3 is far less reactive. That core claim is supported by direct DFT checks and by a well-documented HDNNP with respectable errors (energy RMSE ~0.6 meV/atom, forces ~36 meV/Å). The type I defect study is the strongest part — a single seed defect converts the whole surface in vacuum, with DFT slab calculations at two thicknesses, and a lower cleavage energy. That is a solid, reproducible result.

The soft spots are real but not fatal. The type II Ca dimer is claimed to be stable only in water. The evidence is a DFT vacuum optimization that relaxes the dimer away, a small-cell DFT check with four waters that stabilizes it, and one 100 ps HDNNP run of a pre-formed dimer surface in bulk water. That shows persistence, not equilibrium stability. There is no free-energy profile, no repeated seeding runs, and the DFT check is a 2×1 cell rather than a realistic interface. So the 'only in water' conclusion is plausible but under-supported; the paper should soften it to 'stabilized by water in our simulations' or add a free-energy or dissolution test.

The T5 full/empty-row section is explicitly exploratory — the authors admit the HDNNP was not trained on T5-like structures. Their conclusion that the rows break up in water is tentative, and they say so. That is acceptable but should be identified as a prediction to check with a retrained potential. The bigger issue is reproducibility: neither the trained potential nor the dataset is released, which matters when several conclusions depend on extrapolation.

Minor: Table III averages lack error bars, though the SI shows trajectory statistics for the equilibration phase, so this is easily fixed. The density of 2.82 vs 2.96 g/cm3 experimental is a bit off but within the usual GGA range.

Overall, this is a serious, carefully written simulation paper with a real new mechanistic result. The type II claim needs tempering or stronger evidence, not a desk rejection. I would send it to peer review and ask the authors to either provide a free-energy/umbrella-sampling estimate for dimer formation or rephrase the claim to match the evidence, and to release the potential/data.

Bring to reading group if you're interested in mineral-water interfaces or MLP practice.

Referee Report

3 major / 4 minor

Summary. The paper trains a high-dimensional neural network potential (HDNNP) on RPBE-D3 data and uses it in molecular dynamics simulations of several terminations of the γ-belite (010) surface in contact with liquid water. The authors find that water adsorbs both molecularly and dissociatively, that dissociation is controlled mainly by protonation of accessible SiO4 surface oxygens, and that the resulting hydroxide ions are mostly adsorbed at surface Ca atoms. The most stable T3 termination is shown to be only weakly reactive, while the slightly less stable T2 termination develops two types of Ca defects: a type I defect (Ca displacement plus SiO4 rotation) that is stable in vacuum, and a type II defect (a Ca dimer) that is claimed to be stable only in water. A full/empty-row T5 termination is also explored as a side case. The central claim is that a variety of surface structures exist at this interface and are stabilized by water.

Significance. If the main claims hold, this is a valuable contribution to the atomistic understanding of γ-belite hydration, relevant to low-carbon cement chemistry. The paper's strengths include an actively learned DFT reference dataset, bulk validation of the HDNNP against DFT and experiment, direct DFT geometry optimizations supporting the vacuum stability of type I defects and the stabilizing role of four water molecules for type II, and carefully documented computational settings. I do not regard the study as circular: the MLP is trained to DFT data and used for interpolation, and the key defect findings are checked by explicit DFT calculations. However, the thermodynamic stability of the type II dimer in water is the load-bearing claim that currently rests on a single 100 ps trajectory; this and the lack of statistical error bars for the principal quantitative results prevent me from recommending acceptance at this stage.

major comments (3)

[Section IV.C.3, Type II (Figs. 10–12)] The claim that the type II Ca dimer is "only stable in water" is supported by (i) a DFT vacuum optimization that relaxes an inserted dimer toward a type I-like pattern, (ii) DFT stabilization by four explicit water molecules in a small (2×1) vacuum cell, and (iii) one 100 ps HDNNP NPT trajectory of a fully dimerized (4×4) T2 surface in bulk water, with the surface frozen for the first 10 ps. Item (iii) demonstrates kinetic persistence, not thermodynamic stability. There is no free-energy profile, no independent repeat, and no DFT energy check of the hydrated dimer-row state. Because the HDNNP was not trained on this fully dimerized motif, a systematic overstabilization of the Ca–Ca dimer cannot be excluded. This is the central link in the 'only in water' conclusion and the broader 'structural variety stabilized by water' message. I ask for either free-energy sampling along the Ca displac
[Section IV.C.3, Full/empty-row reconstruction (Fig. 14)] The authors state that the HDNNP "has not been trained to T5-like structures" and then use it to conclude that the full/empty T5 rows "do not remain intact" in water. This extrapolation is acknowledged, but the claim as written exceeds the evidence: a single 100 ps MD trajectory of an initially constructed T5 surface cannot establish instability if the potential is untrained for that motif. This is a side exploration rather than the central claim, but it is part of the paper's general suggestion of structural variety. The conclusion should be rephrased as a tentative observation, or additionally supported by DFT calculations or a refined potential.
[Table III and Section IV.C.2] The principal quantitative statements—essentially all protons adsorbed at surface oxygens, 96% vs 24% Ca-bound hydroxide for T2/T3, and the time-averaged species counts (32.5 OsH−, 30.7 O*H−, 8.2, 7.6)—are reported without statistical uncertainty. The SI shows standard deviations over 8 trajectories only for the NPT equilibration endpoints (Figs. S8/S9), not for the 1 ns NVT production used for Table III and Figure 3. Without block averages or independent production runs, the differences between T2 and T3 cannot be distinguished from sampling noise. I request error bars for all time-averaged quantities, or additional independent production trajectories.

minor comments (4)

[Section III.C] Species classification relies on a 2.9 Å Ca–O adsorption threshold derived from the bulk RDF decay. Since this threshold directly defines H2O*/O*H− counts in Table III and the density profiles in Fig. 3, a sensitivity analysis (e.g., 2.8 and 3.0 Å) would strengthen the claim that the qualitative conclusions are unaffected by this choice.
[Section IV.C.3, Type I] For the thick slab, the type-I reconstruction converts only the top surface, but the reported cleavage energy in Eq. (2) averages over both surfaces. The statement that the "effective stabilization" of the transformed surface is higher is plausible but not directly demonstrated. Please report the per-surface energy of the reconstructed side separately, or explain the decomposition.
[Section IV.C.2, Vertical density profiles] The density profiles in Fig. 3 are very informative, but the text sometimes refers to peaks without indicating whether they are statistically converged. A statement on the block-length or time window used for the averages would help.
[Introduction and Acknowledgements] There are a few language issues, e.g., "limiations" in the Introduction and minor grammatical slips elsewhere. These do not affect the science but should be corrected in a revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MLP-MD results are validated against independent DFT and experimental benchmarks; self-citations are methodological, not load-bearing.

full rationale

The central derivation chain is: train a high-dimensional neural network potential to DFT reference data, run MD simulations of the T2 and T3 interfaces with water, and analyze the resulting trajectories. The MD observables (protonation counts, hydroxide adsorption, defect formation) are simulation outcomes, not parameters fitted to those same observables. The potential is a surrogate for DFT, so the MD results inherit DFT quality, but this is standard interpolation/extrapolation with an ML potential, not circular reasoning. Key structural claims are independently checked by explicit DFT calculations: the type I defect is optimized by DFT in vacuum and shown to lower the cleavage energy; the type II dimer is optimized by DFT in vacuum (where it relaxes to a type I-like geometry) and with four water molecules (where it is stabilized). The paper explicitly acknowledges that the HDNNP was not trained to T5-like structures and that refinement would be needed for that surface, which is a stated limitation rather than a hidden circular assumption. Self-citations to the HDNNP method (Behler-Parrinello), RuNNer, and n2p2 are methodological references and are not used to justify the physical conclusions. Bulk lattice parameters, cohesive energy, density, and cleavage energies are compared against independent experimental and earlier DFT values, providing external checks. No equation or fitted parameter is shown to reduce by construction to the claimed predictions. Therefore no circular step is identified and the score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central result rests on the fitted MLP and on standard domain assumptions about DFT accuracy, simulation length, and species assignment. No new physical entities are introduced; the defects are geometric rearrangements of existing atoms.

free parameters (4)

HDNNP weight parameters = not listed (trained on 24,384 structures)
The potential's network weights are optimized to reproduce DFT energies and forces; all MD conclusions depend on this fit. Test RMSEs are 0.654 meV/atom and 36.2 meV/Bohr.
ACSF cutoff radius and symmetry-function parameters = 6 Å; η values e.g. 0, 0.014, 0.062, 0.486 bohr^-2
Chosen by hand in SI Section E; these define the local atomic environment representation used by the potential.
Ca-O adsorption distance threshold = 2.9 Å
Defines adsorbed vs non-adsorbed species; based on the first peak in the bulk γ-belite RDF (SI Fig. S1).
MD simulation time scales = 1 ns equilibration + 1 ns production
Simulation length chosen for cost; the authors note high barriers and no further Ca displacements in NVT, implying possible under-sampling of slow surface changes.

assumptions (4)

domain assumption RPBE-D3 DFT provides accurate reference energies and forces for bulk water, γ-belite, and their interfaces.
Used for all training data and DFT validations; the accuracy of the MLP and conclusions inherits the DFT accuracy.
domain assumption HDNNP interpolation within its training data is accurate in unconstrained MD.
Relies on the reported RMSEs; extrapolation to T5 is acknowledged as untested by the authors.
domain assumption 1 ns equilibration is sufficient for the surface to reach the observed rearrangements and that these are representative.
Authors state equilibrium is reached after about 1 ns, but they also note that high barriers may prevent further transformations.
domain assumption Assignment of water species by closest-O H atoms is correct.
Section III.C describes this classification; it directly affects the reported hydroxide-to-water ratios.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Molecular Dynamics Simulations of $\gamma$-Belite(010)-Water Interfaces with High-Dimensional Neural Network Potentials." pith.science (2026). https://pith.science/paper/6D3PWCZR

@misc{pith2026251209702,
  author       = {Pith},
  title        = {Pith review of: Molecular Dynamics Simulations of $\gamma$-Belite(010)-Water Interfaces with High-Dimensional Neural Network Potentials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6D3PWCZR}},
  note         = {Machine review of arXiv:2512.09702}
}

abstract

Belite -- dicalcium silicate Ca$_2$SiO$_4$ -- is a main constituent of low-carbon cement. In this work, we study several terminations of the (010) surface of $\gamma$-belite, its most stable polymorph, by molecular dynamics simulations. The energies and forces are provided by a high-dimensional neural network potential trained to density functional theory data. Water can interact in molecular form as well as dissociatively with the investigated interfaces, and the degree of dissociation is determined primarily by the protonation of SiO$_4$ groups accessible at the surface. A major part of the simultaneously formed hydroxide ions is adsorbed at surface calcium atoms, whose octahedral coordination spheres are completed by additional water molecules. The T3 termination, which is most stable in vacuum, shows only little reactivity in water. For the only slightly less stable T2 termination, however, two distinct types of surface defects are observed. The type I defect is even stable in vacuum and leads to a reconstruction of the entire surface, while the type II defect is only found in the presence of water. Overall, our results suggest that a variety of structures may be formed at the Ca$_2$SiO$_4$(010) surface, which are stabilized in the presence of water.

Figures

Figures reproduced from arXiv: 2512.09702 by the authors.

**Figure 2.** FIG. 2. DFT-optimized (2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** FIG. 3. Density profiles of different oxygen species (O*H [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (10 more)

**Figure 4.** Figure 4: FIG. 4. Top views of the relative probability distributions of the calcium atom positions during [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: FIG. 5. Lateral probability distributions of the oxygen atoms in the adsorbed hydroxide ions for the T2 (a,b) and T3 (c,d) [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: FIG. 6. Lateral probability distributions of oxygen atoms in adsorbed water molecules (with Ca-O distances smaller than 2.9 Å) [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 8.** Figure 8: FIG. 8. Coordination number distributions of the calcium [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

**Figure 7.** Figure 7: FIG. 7. Lateral distributions of the protonated surface oxygen [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

**Figure 11.** Figure 11: FIG. 11. Top view of the DFT-optimized T2 terminated [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]

**Figure 9.** Figure 9: FIG. 9. (a) Top views of the T2 terminated surfaces after [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

**Figure 10.** Figure 10: FIG. 10. Top view of the T2 terminated surface in vacuum [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

**Figure 13.** Figure 13: FIG. 13. (2 [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]

**Figure 14.** Figure 14: FIG. 14. Top view of the T5-water interface after a HDNNP [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

RuNNer 2.0: A Software Suite for High-Dimensional Neural Network Potentials
physics.chem-ph 2026-07 conditional novelty 6.0 of 10

RuNNer 2.0 accelerates fourth-generation neural network potentials by implementing a quasi-linear plane-wave charge-equilibration solver, making non-local charge transfer simulations of large systems practical.

Reference graph

Works this paper leans on

4 extracted references · cited by 1 Pith paper

[1]

Initial equilibration of the interfaces A key process for the equilibration of oxide-water interfaces is the saturation of undercoordinated inter- face atoms, which next to molecular adsorption also involves the dissociative adsorption of interfacial water molecules51,65. Specifically, there are two important pro- cesses atγ-belite-water interfaces, which...
[2]

The time- averaged numbers of these species in the equilibrated sys- tems obtained in 1 nsN V Tsimulations are compiled in Table III for both terminations

Characterization of the hydrated surface Surface adsorption During the initial equilibration of the interface, dif- ferent chemical species have been formed through wa- ter dissociation and proton transfer reactions. The time- averaged numbers of these species in the equilibrated sys- tems obtained in 1 nsN V Tsimulations are compiled in Table III for bot...
[3]

parallel

Surface defects and reconstruction of T2 As discussed above, for the T2 termination two types of calcium atom displacements have been observed dur- ing theN P Tequilibration of the solid-liquid interfaces, which indicate the possible existence of more favorable structures and the onset of surface reconstructions, which might be induced and stabilized by t...
[100]

lionanalysis,

and [001] directions. Both surfaces are equivalent, and the last row shows a bottom view. FIG. 14. Top view of the T5-water interface after a HDNNP- driven 100 ps MD simulation in theNPTensemble. The Ca atoms of the first layer are shown in dark green. The simulation box was built as a (4×4) supercell of the single unit, which correspond to (4×2) supercel...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Initial equilibration of the interfaces A key process for the equilibration of oxide-water interfaces is the saturation of undercoordinated inter- face atoms, which next to molecular adsorption also involves the dissociative adsorption of interfacial water molecules51,65. Specifically, there are two important pro- cesses atγ-belite-water interfaces, which...

[2] [2]

The time- averaged numbers of these species in the equilibrated sys- tems obtained in 1 nsN V Tsimulations are compiled in Table III for both terminations

Characterization of the hydrated surface Surface adsorption During the initial equilibration of the interface, dif- ferent chemical species have been formed through wa- ter dissociation and proton transfer reactions. The time- averaged numbers of these species in the equilibrated sys- tems obtained in 1 nsN V Tsimulations are compiled in Table III for bot...

[3] [3]

parallel

Surface defects and reconstruction of T2 As discussed above, for the T2 termination two types of calcium atom displacements have been observed dur- ing theN P Tequilibration of the solid-liquid interfaces, which indicate the possible existence of more favorable structures and the onset of surface reconstructions, which might be induced and stabilized by t...

[4] [100]

lionanalysis,

and [001] directions. Both surfaces are equivalent, and the last row shows a bottom view. FIG. 14. Top view of the T5-water interface after a HDNNP- driven 100 ps MD simulation in theNPTensemble. The Ca atoms of the first layer are shown in dark green. The simulation box was built as a (4×4) supercell of the single unit, which correspond to (4×2) supercel...