{"id":"6caefbfd-7c5f-4cf2-84bf-97d8d28584b8","arxiv_id":"2505.18993","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A neuroevolution machine learning potential trained on 302 DFT structures predicts tobermorite and C-S-H properties accurately and enables 100k+ atom GPU molecular dynamics.","lead":"Researchers built a machine learning potential for tobermorite and calcium-silicate-hydrate cement phases, trained on only 302 density-functional-theory structures, and showed it reproduces structural, mechanical, and thermal properties well. Because the potential runs on GPUs, it can simulate amorphous C-S-H models with over 100,000 atoms, making large-scale molecular dynamics of concrete materials practical.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"NEP's central claim of ab initio accuracy for C-S-H is not directly validated: the 115k-atom water-rich C-S-H simulation (§4.3) lacks a DFT reference, and the descriptor's known H-O and Ca-O deficiencies (Fig. 7) make this transfer unsafe.","rationale":"The reader correctly identifies the descriptor cutoff/basis as the physical weak point; Figure 7 confirms the model struggles with exactly the H-O and Ca-O interactions that characterize hydrated C-S-H. What makes this load-bearing is that the paper's most novel application, a 115,303-atom amorphous C-S-H tensile simulation, is validated only against a single experimental modulus and not against DFT. A single integrated property can match by coincidence, especially with tunable pore-water placement. If the NEP's force errors are substantially larger on these out-of-distribution structures than the 128 meV/Å test-set RMSE, the 'ab initio accuracy' and transferability claims would be unsupported. The proposed DFT spot-check directly tests the transfer and also probes whether the Figure 7 RDF deviations worsen in water-rich environments. This does not change the reader's conditional verdict; it sharpens the condition under which the paper could be accepted, namely that the C-S-H transfer needs a DFT benchmark. I therefore agree with the reader's weakest assumption and recommend keeping the verdict unchanged.","tokens_in":17557,"tokens_out":7309,"duration_ms":45422,"concrete_test":"Perform DFT single-point energy and force calculations on a representative subset of configurations from the 115,303-atom C-S-H simulation used in §4.3, such as 50 snapshots at tensile strains of 0, 0.05, and 0.1, each reduced to a ~500-atom periodic cell (or equivalently on a smaller C-S-H cell built by the same protocol). Compare NEP-predicted energies and forces to DFT. If the force RMSE on these out-of-distribution C-S-H configurations is substantially larger than the reported test-set value of 128 meV/Å (e.g., above 300 meV/Å) or the H-O and Ca-O RDF deviations worsen relative to Figure 7, the transferability claim fails and the verdict should be more restrictive. If the RMSE remains near the reported test-set value, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim includes 'ab initio accuracy' for C-S-H systems, but the trained NEP is only directly validated against DFT on tobermorite polymorphs and a small held-out set (40 structures) drawn from the same tobermorite-based active-learning trajectories. The amorphous C-S-H application in §4.3 (115,303 atoms, Ca/Si = 1.7, gel pores and water) is never compared against DFT: the only quantitative check is a macroscale tensile modulus (17.85 GPa vs a 17.35 GPa experimental value), which can be fortuitous given the stochastic pore placement and model construction choices. The descriptor's short-range fidelity is already known to be weak in the exact interactions that dominate hydrated C-S-H: §4.2 and Figure 7 report an H-O first-peak height deviation and shifted Ca-O first-peak positions, attributed by the authors to insufficient basis/angular representation of hydrogen bonding and Ca-O ionic character. If these errors persist in the water-rich, disordered C-S-H environments, which lie outside the tobermorite training distribution, the claimed 'ab initio accuracy' for the system's most relevant application is unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":17831,"tokens_out":3305,"duration_ms":31471,"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":[{"comment":"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.","section":"§4.1, Fig. 3"},{"comment":"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.","section":"Table 4, §4.2"},{"comment":"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.","section":"§4.2, Fig. 7"},{"comment":"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.","section":"§4.3, Fig. 10"},{"comment":"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.","section":"Table 1, §3"}],"minor_comments":[{"comment":"Typo: 'developd' should be 'developed'.","section":"§4.1"},{"comment":"The column header 'NequP' is an inconsistent abbreviation; elsewhere the model is called NequIP. Please harmonize the notation.","section":"Table 4"},{"comment":"The sentence contains a duplicated phrase: 'Inthisapproach,Inthisapproach,aseriesofsmalldeformations'. Please fix.","section":"§4.2"},{"comment":"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.","section":"Throughout"},{"comment":"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).","section":"§2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable engineering contribution, but the headline claim of 'ab initio accuracy' is broader than the evidence supports. The force RMSE being four times the DFT convergence criterion and the large elastic constant errors (C66, C33) should be directly addressed; otherwise the central claim will be vulnerable to scrutiny. The amorphous C-S-H section is the most important application but lacks any DFT anchor, despite the authors' own caution. I would like to see the authors either add targeted DFT validation for the amorphous/water-rich regime or explicitly downgrade the claim for that regime. The GitHub link states the model 'will be released'; reproducibility depends on the actual release, so the editor may wish to verify that the repository is public at the time of acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new things: this is the first NEP trained for tobermorite/C-S-H, the active-learning loop converges on 302 training structures (about 5x fewer than the NequIP dataset and 20x fewer than the DP set), and the ZBL repulsive hybrid prevents unphysical short-range collapse in MD. The validation suite is broad and honestly executed: EOS within 1 meV/atom, lattice constants within 1.4% of experiment, elastic constant MAEs of 5-9 GPa, PDOS with no imaginary modes, thermal conductivity reproducing the k22 > k11 > k33 anisotropy from earlier studies, and GPU acceleration at 665 timesteps/s on a single RTX 4090 for the 11 Å phase. None of this is manufactured, and the authors openly flag several limitations.\n\nThe soft spots are real, not fatal. The test-set force RMSE of 128.2 meV/Å is about four times the stated DFT force convergence of 0.03 eV/Å, so 'ab initio accuracy' is an overstatement for forces even though energies and EOS look fine. Several elastic constants miss badly: C66 for 11 Å and 14 Å is off by more than a factor of two, and C33 for 14 Å by 22%. The H-O and Ca-O RDF deviations are acknowledged, and they matter because those are exactly the interactions that dominate water-rich C-S-H. The 115k-atom amorphous C-S-H tensile test has no DFT cross-check; the only quantitative anchor is a tensile modulus of 17.85 GPa against an experimental 17.35 GPa, which could be fortuitous given the stochastic pore construction. The authors themselves say that beyond 10% strain the results are qualitative—a correct caveat, but it means the 'ab initio accuracy' claim for the system's most relevant application is unsubstantiated. Also, '1-2 orders of magnitude' data reduction is not uniform: against NequIP it is about 5x, not 20-100x. And the model files are promised at a GitHub repo but not yet released; that should be a condition of acceptance.\n\nBottom line: this is a serious, usable contribution for cement MLP practitioners. It deserves peer review and likely publication after the force-error framing is softened, the C-S-H transfer claim is qualified, the elastic outliers are discussed, and the model repo is actually populated. I'd cite it for the training set size and efficiency numbers, and I'd bring it to a reading group on MLIPs for hydrated materials. Engage with it; referee it; require the missing benchmarks and model release.","headline":"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.","tokens_in":18437,"tokens_out":3518,"would_cite":true,"duration_ms":30674,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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…","keywords":["machine learning potential","neuroevolution potential","tobermorite","calcium silicate hydrate","molecular dynamics","active learning","mechanical properties","thermal conductivity"],"falsifier":"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.","tokens_in":17334,"feed_emoji":"🧱","tokens_out":4650,"duration_ms":40547,"temperature":0.7,"pith_summary":"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.","feed_headline":"Trained on 302 structures, a neural potential matches DFT for cement","feed_subtitle":"GPU-accelerated molecular dynamics now reaches 115,000 atoms of hydrated calcium silicate.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the NEP framework and GPUMD software that form the methodological core of the paper.","marker":"[40]"},{"why":"Provides the earlier ANN-based tobermorite machine learning potential whose training data and accuracy are used for comparison.","marker":"[32]"},{"why":"Provides the deep potential model for tobermorite and C-S-H that serves as the main baseline for accuracy and speed comparisons.","marker":"[33]"},{"why":"Provides the NequIP tobermorite models and dataset sizes used as another comparison point.","marker":"[38]"},{"why":"Supplies the Ziegler-Biersack-Littmark repulsive potential used in the hybrid framework to prevent unphysical atomic clustering.","marker":"[51]"},{"why":"Supplies the DFT framework used to label the training and test structures.","marker":"[49]"},{"why":"Provides the experimental C-S-H tensile elastic modulus of 17.35 GPa that the model's 17.85 GPa prediction is compared against.","marker":"[17]"},{"why":"Provides experimental lattice parameters for tobermorite 14 Å used to benchmark the predicted lattice constants.","marker":"[54]"}],"fun_headline_variants":["Cement potential trained on 300 structures matches DFT accuracy","NEP force field: 300 structures, DFT-level accuracy for C-S-H","GPU-accelerated NEP simulates 115k atoms of cement hydrate","DFT-accurate cement potential from just 300 structures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Cement potential trained on 300 structures matches DFT accuracy","NEP force field: 300 structures, DFT-level accuracy for C-S-H","GPU-accelerated NEP simulates 115k atoms of cement hydrate","DFT-accurate cement potential from just 300 structures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001025,"raw_usage":{"total_tokens":4327,"prompt_tokens":958,"completion_tokens":3369,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":3292}},"tokens_in":574,"tokens_out":3369,"duration_ms":16346,"temperature":1.0,"reasoning_tokens":3292,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:22:48.203216+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"GPUMD: A package for constructing accurate machine-learned potentials and performing highlyefficientatomisticsimulations,TheJournalofChemicalPhysics.157(2022)114801","cited_arxiv_id":null,"evidence_quote":"Supplies the NEP framework and GPUMD software that form the methodological core of the paper."},{"cited_title":"Machine learning potentials for tobermorite minerals, Computational Materials Science","cited_arxiv_id":null,"evidence_quote":"Provides the earlier ANN-based tobermorite machine learning potential whose training data and accuracy are used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the deep potential model for tobermorite and C-S-H that serves as the main baseline for accuracy and speed comparisons."},{"cited_title":"Performance Comparisons of NequIP and DPMD Machine Learning Interatomic PotentialsforTobermorites,ComputationalMaterialsScience.244(2024)113212","cited_arxiv_id":null,"evidence_quote":"Provides the NequIP tobermorite models and dataset sizes used as another comparison point."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Ziegler-Biersack-Littmark repulsive potential used in the hybrid framework to prevent unphysical atomic clustering."},{"cited_title":"VASPKIT: A user-friendly interface facilitating 27 high-throughputcomputingandanalysisusingVASPcode,ComputerPhysicsCommunications","cited_arxiv_id":null,"evidence_quote":"Supplies the DFT framework used to label the training and test structures."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the experimental C-S-H tensile elastic modulus of 17.35 GPa that the model's 17.85 GPa prediction is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides experimental lattice parameters for tobermorite 14 Å used to benchmark the predicted lattice constants."}],"review_version":1}