Pith. sign in

REVIEW 5 major objections 5 minor 60 references

A high-efficiency neuroevolution potential for tobermorite and calcium silicate hydrate systems with ab initio accuracy

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a neuroevolution machine learning potential trained on about 300 structures can match DFT accuracy for tobermorite and calcium silicate hydrate systems, while running fast enough on GPUs to simulate large…

desk verdict Useful, data-efficient NEP for tobermorite and C-S-H, but the 'ab initio accuracy' claim outruns the force errors and the amorphous C-S-H transfer rests on a single elastic modulus match. read the letter →

arxiv 2505.18993 v1 pith:6ODARD4M submitted 2025-05-25 cond-mat.mes-hall

Xiao Xu , Shijie Wang , Haifeng Qin , Zhiqiang Zhao , Zheyong Fan , Zhuhua Zhang , Hang Yin This is my paper · ORCID
classification cond-mat.mes-hall
keywords machinelearningpotentialneuroevolutiontobermoritecalciumsilicatehydratemoleculardynamicsactivemechanicalpropertiesthermalconductivity
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

The paper sets out to show that a neuroevolution machine learning potential (NEP) trained on just 302 density-functional-theory-labeled structures can predict energies, forces, and material properties of tobermorite and calcium silicate hydrates with accuracy comparable to DFT. This matters because cement's main binding phase, C-S-H, has lacked an interatomic potential that is simultaneously accurate, fast, and scalable. The authors demonstrate the potential on equations of state, elastic constants, phonon densities of states, thermal conductivity, and tensile tests, including a 115,303-atom amorphous C-S-H model. If correct, the work means a practical, data-efficient route to realistic molecular dynamics simulations of cementitious materials.

What carries the argument

The load-bearing object is the NEP model: a feedforward neural network with one hidden layer whose inputs are radial and angular descriptor components built from Chebyshev polynomial expansions (radial order 10, angular order 8) within a 4.5 Å cutoff, plus a Ziegler-Biersack-Littmark short-range repulsion term. The descriptor maps each atom's local environment to a site energy, and training minimizes a weighted loss over energy, force, and virial errors. Active learning via farthest-point sampling over descriptor distances iteratively adds the most diverse configurations from NEP-MD trajectories to the 302-structure training set, and GPU-accelerated molecular dynamics supplies the speed that makes large simulations practical.

What would settle it

Run NEP and DFT on the same set of water-rich C-S-H configurations with Ca/Si ratios above 1.7 or under pressures outside the ±10% volume training range and compare radial distribution function first peaks and atomic forces; if H-O peak heights or Ca-O distances deviate beyond the reported force RMSE, or if MD produces unphysical hydrogen clustering, the claimed ab initio transferability fails.

Watch

Extended reading notes

Core claim

Using the NEP framework with a Ziegler-Biersack-Littmark short-range repulsion hybrid and active-learning selection, the authors build a single-hidden-layer neural network potential for tobermorite 9 Å, 11 Å, and 14 Å and for disordered C-S-H. The model reaches test-set root-mean-square errors of 5.159 meV/atom in energy and 128.200 meV/Å in force, and reproduces lattice constants within 1.4% of experiment, with elastic-constant mean absolute errors of 4.96–8.97 GPa relative to DFT. It reproduces equations of state within 1 meV/atom, phonon densities of states without imaginary frequencies, and thermal conductivity trends consistent with earlier studies. On amorphous C-S-H, the predicted tensile modulus of 17.85 GPa is close to the experimental 17.35 GPa, although the paper states that predictions beyond roughly 10% strain are only qualitative because the training set did not include large plastic deformations.

Load-bearing premise

The model's near-DFT accuracy rests on the assumption that a 4.5 Å local descriptor with the chosen basis can represent hydrogen bonding and Ca–O ionic interactions well enough, and the paper's own radial distribution function deviations show this is the part most likely to fail under water-rich or high-Ca/Si conditions.

Editorial extensions

If this is right

  • DFT-level accuracy becomes accessible from roughly two orders of magnitude fewer training structures than earlier tobermorite machine learning potentials, cutting the cost of dataset generation.
  • Simulations can reach 115,303 atoms and thousands of atoms per GPU card, which means mechanical and thermal response of realistic cement microstructures can be probed directly.
  • Because the model reproduces lattice constants, elastic constants, phonon spectra, and thermal conductivity, it offers a single potential for structure–property studies across tobermorite polymorphs.
  • The same active-learning workflow can be re-run when new phases or deformation regimes are added, which the paper identifies as the route to large-strain and defect-rich C-S-H behavior.

Reading between the lines

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

  • The paper's own radial distribution function results show deviations in H-O first-peak height and Ca-O peak positions, so the claimed ab initio accuracy is most secure for local bonding environments represented in the training set; water-rich or high-Ca/Si compositions are a natural stress test beyond the paper.
  • Because the descriptor uses a 4.5 Å cutoff, longer-range electrostatics and hydrogen-bond networks are captured only implicitly; an extension that adds explicit electrostatics or a longer-range descriptor could improve transferability to hydrated gels.
  • A testable extension would be to retrain the same NEP architecture on mixed cement phases such as alite, belite, or aluminate to see whether the roughly 300-structure efficiency holds beyond tobermorite-like chemistries.
  • The observed degradation beyond 10% strain implies the current model should be read as a near-equilibrium potential; retraining with failure-trajectory snapshots is a concrete next step the paper itself suggests.
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.

Referee Report

5 major / 5 minor

Summary. The paper develops a neuroevolution machine learning potential (NEP) for tobermorite and calcium silicate hydrate (C-S-H) systems, trained on only 302 DFT-labeled structures via active learning. The authors report test-set RMSE values of 5.159 meV/atom for energy and 128.200 meV/Å for force, lattice constants within 1.4% of experiment, and elastic constant mean absolute errors of 4.96–8.97 GPa for the three tobermorite polymorphs. They further demonstrate GPU-accelerated MD efficiency, phonon DOS, thermal conductivity, and a 115,303-atom tensile simulation of amorphous C-S-H. The central claim is that NEP achieves DFT-level accuracy for tobermorite and C-S-H with drastically reduced training data and high computational efficiency.

Significance. If the central claim holds, the paper would provide a data-efficient, GPU-scalable machine learning potential for a technologically important material system, addressing a real gap in cement science. The strengths include the small training set (302 structures versus thousands for prior MLPs), the active-learning workflow, the quantitative EOS agreement (<1 meV/atom), the consistency of lattice constants with experiment, and the demonstration of GPU speedups over DP. The planned open release of the model (GitHub repository) is also a positive reproducibility measure. However, the significance is tempered by the acknowledged descriptor limitations for hydrogen bonding and Ca-O ionic interactions, the high force RMSE relative to the DFT convergence criterion, and the lack of direct DFT validation for the amorphous C-S-H application.

major comments (5)
  1. [§4.1, Fig. 3] The reported test-set force RMSE of 128.200 meV/Å (0.128 eV/Å) is about four times the DFT force convergence criterion of 0.03 eV/Å stated in §3. This ratio directly contradicts the claim that NEP achieves 'prediction accuracy comparable to DFT calculations'. The paper should discuss this discrepancy explicitly, report per-species or per-environment force errors, and compare with force RMSEs of other MLPs in this chemistry (e.g., DP, NequIP) to contextualize whether this level is acceptable for the intended MD applications.
  2. [Table 4, §4.2] Although the mean absolute errors of the elastic constant matrices (4.96–8.97 GPa) are reported as evidence of high accuracy, individual components show large deviations: C66 for the 11 Å and 14 Å structures has errors of 112.7% and 107.7% relative to DFT, and C33 for 14 Å deviates by 22.7%. These errors are load-bearing because elastic constants directly determine mechanical property predictions, and the claim that 'NEP delivers highly accurate predictions of elastic properties' is not supported by the full matrix. The paper should report and discuss the component-wise errors, especially for off-diagonal and shear constants, and explain why the MAE metric is appropriate when some components are off by more than a factor of two.
  3. [§4.2, Fig. 7] The authors acknowledge that the NEP model has 'limitations in accurately describing the short-range potential energy surface related to hydrogen bonding' and that the Ca-O bond's ionic character 'may not be fully captured by the NEP model'. These are precisely the interactions that dominate hydrated, water-rich C-S-H. Since the descriptor cutoff is set to 4.5 Å and the angular basis orders are limited (Table 2), the acknowledged deficiencies are not incidental but stem from the descriptor architecture. The paper should either provide additional validation in water-rich or high-Ca/Si environments (e.g., RDFs from DFT AIMD at the same state points) or explicitly restrict the 'ab initio accuracy' claim to the tobermorite-like environments that were directly validated.
  4. [§4.3, Fig. 10] The amorphous C-S-H tensile simulation (115,303 atoms, Ca/Si = 1.7, gel pores and water) is an extrapolation beyond the training distribution, which consists of tobermorite polymorphs and a small number of defect structures. The only quantitative validation for this system is the tensile modulus (17.85 GPa vs. an experimental 17.35 GPa), and the authors themselves state that predictions beyond ~0.1 strain are merely qualitative because the training data did not cover large plastic deformation. This does not substantiate the abstract's claim of 'ab initio accuracy' for C-S-H systems. The paper should either add DFT reference calculations on representative amorphous configurations (or at least on the water-containing defect structures) to confirm transferability, or revise the claim to state that the C-S-H application is a preliminary demonstration rather than a validated ab initio-level prediction.
  5. [Table 1, §3] The test set contains only 40 structures, sampled from the same tobermorite-based active-learning trajectories as the training set. This is a very small held-out set for assessing generalization to the diverse conditions (temperatures, pressures, deformation modes, and defective C-S-H structures) claimed in the paper. The generalization claim would be strengthened by reporting test errors on a larger, independently generated set, or by providing confidence intervals on the reported RMSE values.
minor comments (5)
  1. [§4.1] Typo: 'developd' should be 'developed'.
  2. [Table 4] The column header 'NequP' is an inconsistent abbreviation; elsewhere the model is called NequIP. Please harmonize the notation.
  3. [§4.2] The sentence contains a duplicated phrase: 'Inthisapproach,Inthisapproach,aseriesofsmalldeformations'. Please fix.
  4. [Throughout] Equations are numbered inconsistently (e.g., Eq. (1) appears as '#1)' and Eq. (10) is referenced after Eq. (9) in the text). Please ensure equation numbering is sequential and references match.
  5. [§2.1] The notation for the descriptor equations is garbled by PDF extraction; please check that subscripts, superscripts, and summation limits are typeset correctly, especially in Eqs. (2)–(5).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the NEP model is a supervised fit to DFT data, and its claimed predictions (test-set errors, EOS, elastic constants, RDF, PDOS, thermal conductivity) are evaluated against held-out DFT and experimental references, not against fitted values.

full rationale

The paper's central claim is that a NEP model trained on ~300 DFT-labeled structures achieves ab initio-level accuracy for tobermorite and C-S-H. This is a machine-learning fit, not a derivation, and the validation chain is self-contained: the 40-structure test set is held out from training, the reported RMSEs (5.159 meV/atom energy, 128.200 meV/Å force) are computed against DFT references on that test set, and the EOS, lattice constants, elastic constants, RDFs, PDOS, and thermal conductivity are emergent properties not included in the loss function. These are evaluated against independent DFT calculations and experimental values. The use of the NEP/GPUMD framework, with co-author Zheyong Fan as a developer, is a normal methodological citation to open-source, code-reproduced software and does not constitute load-bearing circularity. The C-S-H application in §4.3 uses the authors' own CSH modeling program [59] and compares the tensile modulus to an experimental value (17.85 GPa vs 17.35 GPa), which is an external benchmark, not an input. The acknowledged limitations—H-O and Ca-O RDF deviations and degraded accuracy beyond 10% strain—are explicitly stated and are correctness/transferability concerns, not circularity. No equation reduces to its own input, and no prediction is a renamed fit by construction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard supervised fitting plus several hand-chosen hyperparameters and the assumption that short-range descriptors can capture the relevant physics. The main risks are the 4.5 Å cutoff locality assumption and the representativeness of the 302-structure active-learning training set. No new particles or forces are postulated.

free parameters (6)
  • Descriptor cutoff radius = 4.5 Å
    Hand-chosen in §3; sets locality of radial and angular descriptors and directly affects accuracy for H-bond and ionic interactions.
  • Descriptor expansion orders = n_max=10, basis_size=10, l_max=8
    Hand-chosen (Table 2); controls expressivity of the local environment representation.
  • Hidden layer size = 50 neurons
    Hand-chosen (Table 2); capacity of the single-hidden-layer neural network.
  • Loss function weights = lambda_e=1, lambda_f=1, lambda_v=0.2, lambda_1=lambda_2=0.05
    Hand-chosen (Table 2); balance energy, force, virial, and regularization errors during SNES training.
  • ZBL hybrid switch setting = type 2
    Hand-chosen (Table 2); controls the empirical short-range repulsion added to prevent unphysical clustering.
  • NEP neural network weights and biases = Not reported individually; trained by SNES for 1e6 generations
    Fitted to DFT data; these are the primary fitted parameters of the potential.
assumptions (5)
  • domain assumption DFT-PBE-D3 calculations provide a sufficiently accurate ground truth for energies, forces, and virials of tobermorite and C-S-H.
    All training labels come from PBE-D3 (VASP); the claimed 'ab initio accuracy' is accuracy relative to this DFT functional, not to true chemical accuracy. Invoked throughout §3 and §4.
  • domain assumption Tobermorite 9 Å, 11 Å, and 14 Å structures serve as archetypal models for C-S-H.
    §3 states 'tobermorite is widely regarded as archetypal model for constructing C-S-H'; used to justify transfer from crystalline tobermorite to amorphous C-S-H.
  • domain assumption The local environment descriptor with 4.5 Å cutoff captures all interactions relevant for the target properties.
    §3 sets the cutoff; the paper's own RDF deviations for H-O and Ca-O indicate this is only partially satisfied.
  • domain assumption The 302 structures selected by active learning cover the configurational space sampled in MD simulations.
    §3 and §4.3; the paper acknowledges the training set does not cover large plastic deformation, so this assumption is only conditionally valid.
  • ad hoc to paper ZBL hybrid term correctly describes short-range repulsion and switching behavior.
    §3; the specific ZBL type 2 hybridization is introduced to stabilize MD, but its parameters are not derived from independent data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A high-efficiency neuroevolution potential for tobermorite and calcium silicate hydrate systems with ab initio accuracy." pith.science (2026). https://pith.science/paper/6ODARD4M

@misc{pith2026250518993,
  author       = {Pith},
  title        = {Pith review of: A high-efficiency neuroevolution potential for tobermorite and calcium silicate hydrate systems with ab initio accuracy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6ODARD4M}},
  note         = {Machine review of arXiv:2505.18993}
}
read the original abstract

Tobermorite and Calcium Silicate Hydrate (C-S-H) systems are indispensable cement materials but still lack a satisfactory interatomic potential with both high accuracy and high computational efficiency for better understanding their mechanical performance. Here, we develop a Neuroevolution Machine Learning Potential (NEP) with Ziegler-Biersack-Littmark hybrid framework for tobermorite and C-S-H systems, which conveys unprecedented efficiency in molecular dynamics simulations with substantially reduced training datasets. Our NEP model achieves prediction accuracy comparable to DFT calculations using just around 300 training structures, significantly fewer than other existing machine learning potentials trained for tobermorite. Critically, the GPU-accelerated NEP computations enable scalable simulations of large tobermorite systems, reaching several thousand atoms per GPU card with high efficiency. We demonstrate the NEP's versatility by accurately predicting mechanical properties, phonon density of states, and thermal conductivity of tobermorite. Furthermore, we extend the NEP application to large-scale simulations of amorphous C-S-H, highlighting its potential for comprehensive analysis of structural and mechanical behaviors under various realistic conditions.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 60 canonical work pages

  1. [1]

    409 (2023) 134137

    QLuo,YXiang,QYang,TLiang,YXie.Molecular simulationofcalcium-silicate-hydrate and its applications: A comprehensive review, Construction and Building Materials. 409 (2023) 134137

    work page 2023

  2. [2]

    RKMishra,AKMohamed,DGeissbühler,HManzano,TJamil,RShahsavari,AGKalinichev, SGalmarini,LTao,HHeinz,RPellenq,ACTvanDuin,SCParker,RJFlatt,PBowen.cemff: A force field database for cementitious materials including validations, applications and opportunities,CementandConcreteResearch.102(2017)68-89

    work page 2017

  3. [3]

    XMAretxabaleta,JLópez-Zorrilla,IEtxebarria,HManzano.Multi-stepnucleationpathwayof 24 C-S-Hduringcementhydrationfrom atomisticsimulations,NatureCommunications.14(2023) 7979

    work page 2023

  4. [4]

    Physical Origins of Thermal Properties of CementPaste,PhysicalReviewApplied.3(2015)064010

    M J Abdolhosseini Qomi, F-J Ulm, R J M Pellenq. Physical Origins of Thermal Properties of CementPaste,PhysicalReviewApplied.3(2015)064010

    work page 2015

  5. [5]

    Advances in atomistic modeling and understanding of drying shrinkage in cementitious materials, Cement andConcreteResearch.148(2021)106536

    M J Abdolhosseini Qomi, L Brochard, T Honorio, I Maruyama, M Vandamme. Advances in atomistic modeling and understanding of drying shrinkage in cementitious materials, Cement andConcreteResearch.148(2021)106536

    work page 2021

  6. [6]

    JXu,XChen,GYang, XNiu,F Chang,GLacidogna.Review ofresearchonmicromechanical properties of cement-based materials based on molecular dynamics simulation, Construction andBuildingMaterials.312(2021)125389

    work page 2021

  7. [7]

    Does nano basic building-block of C-S-H exist? – A review of direct morphologicalobservations,Materials&Design.238(2024)112699

    Y Yan, G Geng. Does nano basic building-block of C-S-H exist? – A review of direct morphologicalobservations,Materials&Design.238(2024)112699

    work page 2024

  8. [8]

    GRenaudin,JRussias,FLeroux,FFrizon,CCau-dit-Coumes.StructuralcharacterizationofC- S-HandC-A-S-Hsamples—PartI:Long-rangeorderinvestigatedbyRietveldanalyses,Journal ofSolidStateChemistry.182(2009)3312-3319

    work page 2009

Show all 60 references
  1. [9]

    CRößler,FSteiniger,HMLudwig.CharacterizationofC-S-HandC-A-S-Hphasesbyelectron microscopy imaging, diffraction, and energy dispersive X‐ray spectroscopy, Journal of the AmericanCeramicSociety.100(2017)1733-1742

    2017

  2. [10]

    Influences of cross-linking and Al incorporation on the intrinsic mechanical properties of tobermorite, Cement and Concrete Research.136(2020)106170

    J Li, W Zhang, K Garbev, G Beuchle, P J M Monteiro. Influences of cross-linking and Al incorporation on the intrinsic mechanical properties of tobermorite, Cement and Concrete Research.136(2020)106170

    2020

  3. [11]

    M Izadifar, F Königer, A Gerdes, C Wöll, P Thissen. Correlation between Composition and Mechanical Properties of Calcium Silicate Hydrates Identified by Infrared Spectroscopy and DensityFunctionalTheory,TheJournalofPhysicalChemistryC.123(2019)10868-10873

    2019

  4. [12]

    Understanding of bonding and mechanical characteristics of cementitious mineraltobermoritefromfirstprinciples,JComputChem.32(2011)306-314

    D Tunega, A Zaoui. Understanding of bonding and mechanical characteristics of cementitious mineraltobermoritefromfirstprinciples,JComputChem.32(2011)306-314

    2011

  5. [13]

    Nano-scale investigation of elastic properties of hydrated cement paste constituents using molecular dynamics simulations, Computational Materials Science

    S Hajilar, B Shafei. Nano-scale investigation of elastic properties of hydrated cement paste constituents using molecular dynamics simulations, Computational Materials Science. 101 (2015)216-226

    2015

  6. [14]

    Molecular dynamics study on the structure and mechanicalpropertiesoftobermorite,MaterialsScienceandEngineering:B.299(2024)116930

    X Wang, T Li, W Xie, LZhang, D Li, F Xing. Molecular dynamics study on the structure and mechanicalpropertiesoftobermorite,MaterialsScienceandEngineering:B.299(2024)116930

    2024

  7. [15]

    R Shahsavari, M J Buehler, R J M Pellenq, F J Ulm. First‐Principles Study of Elastic Constants and Interlayer Interactions of Complex Hydrated Oxides: Case Study ofTobermorite andJennite,JournaloftheAmericanCeramicSociety.92(2009)2323-2330

    2009

  8. [16]

    Aluminum-Induced Interfacial Strengthening in Calcium Silicate Hydrates: Structure, Bonding, and Mechanical Properties, ACS Sustainable Chemistry &Engineering.8(2020)2622-2631

    Q Zheng, J Jiang, J Yu, X Li, S Li. Aluminum-Induced Interfacial Strengthening in Calcium Silicate Hydrates: Structure, Bonding, and Mechanical Properties, ACS Sustainable Chemistry &Engineering.8(2020)2622-2631

    2020

  9. [17]

    J Fu, S Kamali-Bernard, F Bernard, M Cornen. Comparison of mechanical properties of C-S-H and portlandite between nano-indentation experiments and a modeling approach using various simulationtechniques,CompositesPartB:Engineering.151(2018)127-138

    2018

  10. [18]

    25 156(2022)106767

    YLi,HPan,ZLi.Abinitiometadynamicssimulationsontheformationofcalciumsilicateaqua complexes prior to the nuleation of calcium silicate hydrate, Cement and Concrete Research. 25 156(2022)106767

    2022

  11. [19]

    Structure, hydration, and chloride ingress in C-S-H: Insight from DFT calculations, Cement and Concrete Research

    I-H Svenum, I G Ringdalen, F LBleken, J Friis, D Höche, O Swang. Structure, hydration, and chloride ingress in C-S-H: Insight from DFT calculations, Cement and Concrete Research. 129 (2020)105965

    2020

  12. [20]

    145 (2016)170901

    J Behler.Perspective: Machine learning potentials for atomistic simulations, J Chem Phys. 145 (2016)170901

    2016

  13. [21]

    Molecular Models of Hydroxide, Oxyhydroxide, and Clay Phases and the Development of a General Force Field, J Phys Chem

    R T Cygan, J-J Liang, A G Kalinichev. Molecular Models of Hydroxide, Oxyhydroxide, and Clay Phases and the Development of a General Force Field, J Phys Chem. 108 (2004) 1255- 1266

    2004

  14. [22]

    Empirical force fields for complex hydrated calcio-silicate layeredmaterials,PhysChemChemPhys.13(2011)1002-1011

    R Shahsavari, R J Pellenq, F J Ulm. Empirical force fields for complex hydrated calcio-silicate layeredmaterials,PhysChemChemPhys.13(2011)1002-1011

    2011

  15. [23]

    DFan, SYang.Mechanical properties ofC-S-Hglobules and interfaces bymoleculardynamics simulation,ConstructionandBuildingMaterials.176(2018)573-582

    2018

  16. [24]

    Prediction and evaluation of thermal conductivity in nanomaterial- reinforcedcementitiouscomposites,CementandConcreteResearch.172(2023)107240

    Y Yang, Y Wang, J Cao. Prediction and evaluation of thermal conductivity in nanomaterial- reinforcedcementitiouscomposites,CementandConcreteResearch.172(2023)107240

    2023

  17. [25]

    Effect of porosity and temperature on thermal conductivity of jennite:A molecular dynamics study,Materials Chemistry and Physics

    S-N Hong, C-J Yu, U-S Hwang, C-H Kim, B-H Ri. Effect of porosity and temperature on thermal conductivity of jennite:A molecular dynamics study,Materials Chemistry and Physics. 250(2020)123146

    2020

  18. [26]

    Molecular dynamics study of the effect of moisture and porosity on thermal conductivity of tobermorite 14 Å, International Journal of ThermalSciences.159(2021)106537

    S-N Hong, C-J Yu, K-C Ri, J-M Han, B-H Ri. Molecular dynamics study of the effect of moisture and porosity on thermal conductivity of tobermorite 14 Å, International Journal of ThermalSciences.159(2021)106537

    2021

  19. [27]

    RKMishra,AKMohamed,DGeissbühler,HManzano,TJamil,RShahsavari,AGKalinichev, SGalmarini,LTao,HHeinz.cemff:Aforce fielddatabaseforcementitiousmaterialsincluding validations,applicationsandopportunities,CementandConcreteResearch.102(2017)68-89

    2017

  20. [28]

    Molecular dynamic simulations of cementitious systems using a newly developed force field suite ERICA FF, Cement and ConcreteResearch.154(2022)106712

    MValavi, Z Casar,AKunhi Mohamed, PBowen, S Galmarini. Molecular dynamic simulations of cementitious systems using a newly developed force field suite ERICA FF, Cement and ConcreteResearch.154(2022)106712

    2022

  21. [29]

    156(2022)106784

    EDuque-Redondo,PABonnaud,HManzano.AcomprehensivereviewofC-S-Hempiricaland computationalmodels,theirapplications,andpractical aspects,CementandConcreteResearch. 156(2022)106784

    2022

  22. [30]

    Z Qi, X Sun, Z Sun, Q Wang, D Zhang, K Liang, R Li, D Zou, L Li, G Wu, W Shen, S Liu. Interfacial Optimization for AlN/Diamond Heterostructures via Machine Learning Potential MolecularDynamicsInvestigationoftheMechanicalProperties,ACSApplMaterInterfaces.16 (2024)27998-28007

    2024

  23. [31]

    KZhu, Z Zhang.Equivariance isessential,local representation isa need:Acomprehensive and critical study of machine learning potentials for tobermorite phases, Computational Materials Science.246(2025)113363

    2025

  24. [32]

    Machine learning potentials for tobermorite minerals, Computational Materials Science

    K Kobayashi, H Nakamura, A Yamaguchi, M Itakura, M Machida, M Okumura. Machine learning potentials for tobermorite minerals, Computational Materials Science. 188 (2021) 110173

    2021

  25. [33]

    YZhou,HZheng,WLi,TMa,CMiao.Adeeplearningpotentialappliedintobermoritephases andextendedtocalciumsilicatehydrates,CementandConcreteResearch.152(2022)106685. 26

    2022

  26. [34]

    WLi,YZhou,LDing,PLv,YSu,RWang,CMiao.Adeeplearning-basedpotentialdeveloped for calcium silicate hydrates with both high accuracy and efficiency, Journal of Sustainable Cement-BasedMaterials.12(2023)1335-1346

    2023

  27. [35]

    W Li, C Xiong, Y Zhou, W Chen, Y Zheng, W Lin, J Xing. Insights on the mechanical properties and failure mechanisms of calcium silicate hydrates based ondeep-learning potential moleculardynamics,CementandConcreteResearch.186(2024)107690

    2024

  28. [36]

    HWang, LZhang,JHan,WE.DeePMD-kit:Adeep learningpackage formany-body potential energyrepresentationandmoleculardynamics,ComputerPhysicsCommunications.228(2018) 178-184

    2018

  29. [37]

    Deep Potential Molecular Dynamics:AScalable Model withtheAccuracyofQuantumMechanics,PhysRevLett.120(2018)143001

    LZhang, J Han, HWang, R Car,W E. Deep Potential Molecular Dynamics:AScalable Model withtheAccuracyofQuantumMechanics,PhysRevLett.120(2018)143001

    2018

  30. [38]

    Performance Comparisons of NequIP and DPMD Machine Learning Interatomic PotentialsforTobermorites,ComputationalMaterialsScience.244(2024)113212

    K Zhu. Performance Comparisons of NequIP and DPMD Machine Learning Interatomic PotentialsforTobermorites,ComputationalMaterialsScience.244(2024)113212

    2024

  31. [39]

    E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials,NatCommun.13(2022)2453

    S Batzner,AMusaelian, LSun, M Geiger,J PMailoa, M Kornbluth, N Molinari,TE Smidt, B Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials,NatCommun.13(2022)2453

    2022

  32. [40]

    GPUMD: A package for constructing accurate machine-learned potentials and performing highlyefficientatomisticsimulations,TheJournalofChemicalPhysics.157(2022)114801

    Z Fan, Y Wang, P Ying, K Song, J Wang, Y Wang, Z Zeng, K Xu, E Lindgren, J M Rahm. GPUMD: A package for constructing accurate machine-learned potentials and performing highlyefficientatomisticsimulations,TheJournalofChemicalPhysics.157(2022)114801

    2022

  33. [41]

    Neuroevolution machine learning potentials: Combining high accuracy and low cost in atomistic simulations andapplicationtoheattransport,PhysicalReviewB.104(2021)104309

    Z Fan, Z Zeng, C Zhang, Y Wang, K Song, H Dong, Y Chen, T Ala-Nissila. Neuroevolution machine learning potentials: Combining high accuracy and low cost in atomistic simulations andapplicationtoheattransport,PhysicalReviewB.104(2021)104309

    2021

  34. [42]

    Advances in modeling complex materials:Theriseofneuroevolutionpotentials,arXivpreprintarXiv:250111191.(2025)

    P Ying, C Qian, R Zhao, Y Wang, F Ding, S Chen, Z Fan. Advances in modeling complex materials:Theriseofneuroevolutionpotentials,arXivpreprintarXiv:250111191.(2025)

    2025

  35. [43]

    Quantum-corrected thickness-dependent thermal conductivity in amorphous silicon predicted by machine learning molecular dynamics simulations,PhysicalReviewB.107(2023)054303

    Y Wang, Z Fan, P Qian, M A Caro, T Ala-Nissila. Quantum-corrected thickness-dependent thermal conductivity in amorphous silicon predicted by machine learning molecular dynamics simulations,PhysicalReviewB.107(2023)054303

    2023

  36. [44]

    Z Li, JWang, H Dong,YZhou, LLiu, J-YYang. Mechanistic insights into water filling effects on thermal transport of carbon nanotubes from machine learning molecular dynamics, InternationalJournalofHeatandMassTransfer.235(2024)126152

    2024

  37. [45]

    Interfacial optimization for AlN/diamond heterostructures via machine learning potential molecular dynamics investigation of the mechanical properties, ACS Appl Mater Interfaces

    Z Qi, X Sun, Z Sun, Q Wang, D Zhang, K Liang, R Li, D Zou, L Li, G Wu. Interfacial optimization for AlN/diamond heterostructures via machine learning potential molecular dynamics investigation of the mechanical properties, ACS Appl Mater Interfaces. 16 (2024) 27998-28007

    2024

  38. [46]

    K Xu,YHao, T Liang, PYing, J Xu, J Wu, Z Fan.Accurate prediction of heat conductivity of waterbyaneuroevolutionpotential,TheJournalofChemicalPhysics.158(2023)

    2023

  39. [47]

    General-purpose neural network potential forTi-Al-Nb alloys towardslarge-scalemoleculardynamicswithabinitioaccuracy,Physical ReviewB.110(2024) 184115

    Z Zhao, MYi,W Guo, Z Zhang. General-purpose neural network potential forTi-Al-Nb alloys towardslarge-scalemoleculardynamicswithabinitioaccuracy,Physical ReviewB.110(2024) 184115

    2024

  40. [48]

    Development of a neuroevolution machinelearningpotentialofPd-Cu-Ni-Palloys,Materials&Design.231(2023)112012

    R Zhao, S Wang, Z Kong, Y Xu, K Fu, P Peng, C Wu. Development of a neuroevolution machinelearningpotentialofPd-Cu-Ni-Palloys,Materials&Design.231(2023)112012

    2023

  41. [49]

    VASPKIT: A user-friendly interface facilitating 27 high-throughputcomputingandanalysisusingVASPcode,ComputerPhysicsCommunications

    V Wang, N Xu, J-C Liu, G Tang, W-T Geng. VASPKIT: A user-friendly interface facilitating 27 high-throughputcomputingandanalysisusingVASPcode,ComputerPhysicsCommunications. 267(2021)108033

    2021

  42. [50]

    The atomic simulation environment-a Python library for workingwithatoms,JournalofPhysics:CondensedMatter.29(2017)273002

    A H Larsen, J J Mortensen, J Blomqvist, I E Castelli, R Christensen, M Dułak, J Friis, M N Groves, B Hammer, C Hargus. The atomic simulation environment-a Python library for workingwithatoms,JournalofPhysics:CondensedMatter.29(2017)273002

    2017

  43. [51]

    JZiegler.JPBiersackandU.Littmark,Thestoppingandrangeofionsinsolids.1(1985)

    1985

  44. [52]

    https://github.com/HYinSD/CSH-NEP

  45. [53]

    Principal component analysis,NatureReviewsMethodsPrimers.2(2022)100

    M Greenacre, PJ Groenen, T Hastie,AI d’Enza,AMarkos, E Tuzhilina. Principal component analysis,NatureReviewsMethodsPrimers.2(2022)100

    2022

  46. [54]

    EBonaccorsi,SMerlino,ARKampf.TheCrystal StructureofTobermorite14Å(Plombierite), aC-S-HPhase,JournaloftheAmericanCeramicSociety.88(2005)505-512

    2005

  47. [55]

    SMerlino, EBonaccorsi,TArmbruster.The real structures ofclinotobermorite andtobermorite 9 Å: OD character, polytypes, and structural relationships, European Journal of Mineralogy.12 (2000)411-429

    2000

  48. [56]

    The real structure of tobermorite 11A: normal and anomalous forms, OD character and polytypic modifications, European Journal of Mineralogy

    S Merlino, E Bonaccorsi, T Armbruster. The real structure of tobermorite 11A: normal and anomalous forms, OD character and polytypic modifications, European Journal of Mineralogy. 13(2001)577-590

    2001

  49. [57]

    Elastic properties of the main species present in Portland cementpastes,ActaMaterialia.57(2009)1666-1674

    H Manzano, J S Dolado, AAyuela. Elastic properties of the main species present in Portland cementpastes,ActaMaterialia.57(2009)1666-1674

    2009

  50. [58]

    Molecular Simulation of Calcium Silicate Composites:Structure,Dynamics,andMechanicalProperties,Journal oftheAmericanCeramic Society.98(2014)758-769

    D Hou, T Zhao, Z Jin, H Ma, Z Li, L Q Chen. Molecular Simulation of Calcium Silicate Composites:Structure,Dynamics,andMechanicalProperties,Journal oftheAmericanCeramic Society.98(2014)758-769

    2014

  51. [59]

    SWang, F Ren,YSong, GPapadakis,YYang, HYin.Issuesof standardizingC-S-H molecular models: Random defect distribution and its effects on material performance, Construction and BuildingMaterials.470(2025)140527

    2025

  52. [60]

    AZhou,JKang,RQin,HHao,TLiu,ZYu.Weavingthenext-levelstructureofcalciumsilicate hydrate at the submicron scale via a remapping algorithm from coarse-grained to all-atom model,CementandConcreteResearch.180(2024)107501

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.