{"id":"195d68cf-87b0-48d1-9941-4064488f6d63","arxiv_id":"2509.17862","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"OC25 is a large open dataset and baseline models for solid-liquid interfaces, but the claimed CO dimerization insights are absent from the manuscript body.","lead":"This paper introduces OC25, a dataset of over 7 million density functional theory calculations of catalyst surfaces in contact with solvents and ions, plus machine learning models trained on it. The abstract also claims new results on CO dimerization at copper electrodes, but those results do not appear in the submitted text.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's CO-dimerization result is entirely absent from the manuscript body, making the central claim unsupported.","rationale":"The reader's verdict of REJECT is well-founded. The strongest_claim—that the abstract's dimerization result is absent from the body—is exactly the most load-bearing concern. I agree with this. The weakest_assumption focuses on transferability of OC25-trained models and the magnitude of energy errors; that would be a valid concern if the dimerization simulations existed, but here the deeper issue is that no such simulations are presented at all. The body contains only dataset construction and baseline model evaluation; there is no methods section, no results, and no discussion of CO dimerization. The Outlook section's statement that reactivity predictions 'remain to be tested in future studies' (Section 4) explicitly undercuts the abstract's claims. This is an internal inconsistency between the paper's advertised contribution and its actual content, not a disagreement with scientific consensus. The dataset itself may be legitimate and useful, but the manuscript cannot be accepted as a demonstration of dimerization insights. The remedy would be either to remove the dimerization claims from the abstract and reframe the paper solely as a dataset release, or to add the missing simulations and results. Since the reader already recommends REJECT, my stress-test does not change that verdict.","tokens_in":15467,"tokens_out":5116,"duration_ms":45240,"concrete_test":"Run a systematic full-text and artifact search: (1) grep the manuscript PDF/LaTeX for 'dimer', 'free energy', 'umbrella', 'metadynamics', '7 ns', 'Cu(310)', 'surface charge' in a simulation context; (2) inspect the released GitHub/HuggingFace artifacts for any scripts, inputs, or outputs corresponding to CO dimerization simulations. If no quantitative free-energy results exist in either the manuscript or the released artifacts, the abstract's central claim must be removed or the paper expanded to include the study.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The manuscript's body is a dataset report: it introduces the OC25 dataset (Section 2), baseline model results (Section 3), and future outlook (Section 4). It contains no section, figure, table, or quantitative result on CO dimerization free-energy profiles, enhanced-sampling simulations, >800-atom cells, 7 ns trajectories, surface charge variation, cation identity effects, or Cu(310) facets. The abstract's central claim—'We find that dimerization is weakly sensitive to charge and cation identity, with appreciable stabilization only at the most negative charge densities, while extension to stepped Cu(310) reveals a more favorable pathway'—is therefore not derivable from any content in the paper. This is not a subtle methodological weakness; it is an absence of the advertised study. The Outlook explicitly states that model predictions of interfacial reactivity 'remain to be tested in future studies' (Section 4), directly contradicting the abstract's presentation of these as completed findings. The OC25 dataset may be a valuable community resource, but the paper as submitted cannot support the claimed dimerization insight; the title and abstract promise a demonstration that the body does not deliver.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces the Open Catalyst 2025 (OC25) dataset for solid-liquid interfaces, containing over 7.8 million single-point DFT calculations across 1.5 million explicit-solvent environments, and reports baseline machine-learned interatomic potential results. The body of the paper covers dataset construction, DFT settings and force-convergence filtering, train/validation/test splits including out-of-distribution solvent/ion splits, and baseline energy/force errors for eSEN and UMA models. The title and abstract, however, claim a concrete application: CO dimerization on Cu surfaces, with large (>800-atom) explicit-solvent cells, enhanced sampling up to 7 ns, free-energy profiles as functions of surface charge, cation identity, and facet, and specific findings about weak charge/cation sensitivity and more favorable pathways on Cu(310). This application and its results are entirely absent from the manuscript body, which instead concludes in Section 4 that interfacial reactivity predictions 'remain to be tested in future studies.'","tokens_in":15720,"tokens_out":2672,"duration_ms":27889,"significance":"If the OC25 dataset and baseline models are made openly available as described, the dataset could be a valuable community resource for training and benchmarking machine learning potentials at solid-liquid interfaces. The force-convergence analysis (Section 2.2.4, Figure 4) is a useful contribution, and the OOD splits are thoughtfully designed. However, the advertised central scientific result — the CO dimerization study — is not present, so the paper cannot currently support its title or abstract. The significance of the work as submitted is therefore limited to the dataset report, not to the claimed electrocatalytic insights.","major_comments":[{"comment":"The abstract's central claim — 'We find that dimerization is weakly sensitive to charge and cation identity, with appreciable stabilization only at the most negative charge densities, while extension to stepped Cu(310) reveals a more favorable pathway at modest reducing potentials' — is absent from the manuscript body. Sections 2 and 3 describe dataset construction and baseline model errors only; Section 4 explicitly states that 'these aspects remain to be tested in future studies.' No section, figure, table, or equation reports CO dimerization free-energy profiles, >800-atom simulations, enhanced sampling, surface charge variation, cation identity effects, or Cu(310) facet results. This is not a minor omission; it is the advertised central result.","section":"Abstract, Sections 2–4"},{"comment":"The transferability premise for the claimed application is unvalidated. OC25 structures are described as roughly 144-atom systems sampled from short high-temperature (1000 K) AIMD or 5-step relaxations. The abstract claims the models enable 'large cells (>800 atoms)' and 'enhanced sampling up to 7 ns' for CO dimerization, but no evidence is presented that models trained on these small off-equilibrium configurations remain accurate over long MD trajectories, at electrochemical conditions, or for system sizes five or more times larger. Baseline OOD errors in Table 2 are substantially larger than in-distribution errors, and no long-timescale stability or free-energy convergence test is reported.","section":"Section 2.1 and 2.2.3 vs. Abstract"},{"comment":"The Outlook itself contradicts the abstract. After stating that models 'may still be able to accurately predict interfacial properties and reactivity,' the text says 'these aspects remain to be tested in future studies.' This directly undermines the abstract's presentation of the CO dimerization results as completed findings. The paper should either include the missing study or be reframed and retitled as a dataset paper.","section":"Section 4"}],"minor_comments":[{"comment":"Typo: 'significantly lower than than the recently released' has a duplicated 'than'.","section":"Abstract"},{"comment":"Equation (2) in Appendix B.3 repeats Equation (1) verbatim with different notation; the pseudo-solvation energy is defined twice. Consolidate to avoid confusion.","section":"Section 2.3.1 and Appendix B.3"},{"comment":"The phrase 'surprisingly' / 'to our surprise' is used twice for the same observation about model robustness to label noise; one occurrence should be removed.","section":"Sections 3.2 and 4"},{"comment":"The text says structures with drift 'greater than 10 eV/Å' have much larger errors, then a threshold of '1 eV/Å' is selected. This is understandable but the relation between the two values should be stated explicitly to avoid apparent inconsistency.","section":"Figure 4"}],"recommendation":"reject","confidential_remarks":"The OC25 dataset construction is detailed and potentially useful, and the force-convergence study is a genuine methodological contribution. However, the title and abstract promise a CO dimerization study that is entirely missing from the body. This is not a local fixable issue; it requires either adding a large simulation campaign or substantially reframing the paper as a dataset release. As submitted, the manuscript's central claim is unsupported, so I cannot recommend publication in its current form. If the authors resubmit a revised version focused on the dataset without the dimerization claims, or with the missing study fully included, I would be willing to reconsider."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the OC25 dataset work is real and valuable, but the paper does not contain the CO dimerization study the abstract claims. The body is a dataset report—construction, splits, force-convergence analysis, baseline models. There is no section on free-energy profiles, no >800-atom cells, no 7 ns enhanced sampling, no charge or cation variation, no Cu(310). The Outlook even states that model predictions of interfacial reactivity 'remain to be tested in future studies,' which directly contradicts the abstract's 'We find...' statements.\n\nWhat's new and good: OC25 is the largest solid-liquid interface dataset I've seen, with 7.8M single-point DFT calculations over ~1.5M unique systems, 88 elements, multiple solvents and ions, and off-equilibrium sampling. The force-convergence analysis is careful: they quantify drift errors against tighter EDIFF, apply a 1 eV/A filter, and show models tolerate some label noise. The OOD splits for unseen solvents and ions are well designed. Baseline results are reported cleanly, with energy and force MAEs as low as 0.105 eV and 0.015 eV/A, and the released models and data are a real community resource.\n\nSoft spots: the abstract/body mismatch is not a minor blemish; it's load-bearing. The central claim is unsupported by any in-body evidence. Also, even if the dimerization study had been included, the transferability from ~144-atom off-equilibrium training configs at 1000 K to >800-atom explicit-solvent cells at electrochemical potentials would need justification; none is given. The dataset portion itself is solid, but the paper cannot be accepted as a demonstration of dimerization insights.\n\nWho it's for: people building or using ML interatomic potentials for solid-liquid interfaces will find the dataset and baselines useful. People reading for CO2 reduction mechanism insights will be disappointed. I'd like to see either the removal of all dimerization claims and a resubmission as a dataset paper, or the actual study added. As is, I'd reject, but the dataset deserves serious referee review, so I'd send it out—expecting major revision or rejection.\n\nI would cite the dataset if I needed interface data, but not the dimerization claim. For a reading group, it's a useful case study in matching abstract to content.","headline":"OC25 is a genuine dataset contribution, but the abstract's dimerization free-energy results are absent from the body—the paper is two different documents stapled together.","tokens_in":16222,"tokens_out":3270,"would_cite":true,"duration_ms":31438,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A new machine-learning dataset of 7.8 million solid-liquid interface calculations is used to claim that CO dimerization on copper is only weakly affected by surface charge and cation identity, except at very negative charges, and that stepp","keywords":["solid-liquid interfaces","machine learning interatomic potentials","CO dimerization","electrocatalysis","CO2 electroreduction","explicit solvation","graph neural networks","open dataset"],"falsifier":"Reproduce the claimed 7-ns enhanced-sampling simulation on Cu(100) and Cu(310) using the released OC25 checkpoint, and confirm that the reported free-energy ordering survives; then run conventional AIMD on one representative state to check whether the 0.1 eV model error flips the barrier difference. A simpler check: the body of the paper contains none of the dimerization free-energy profiles, so finding their data or scripts would be the first confirmation.","tokens_in":15397,"feed_emoji":"⚡","tokens_out":6342,"duration_ms":47091,"temperature":0.7,"pith_summary":"The paper aims to close a gap in open catalyst datasets: none had captured solid-liquid interfaces with explicit solvent and ions. It introduces OC25, a dataset of 7.8 million DFT single-point calculations across 1.5 million unique solvated surface configurations, and trains graph-neural-network potentials that reach 0.1 eV energy and 0.015 eV/Å force accuracy. As a demonstration, the abstract reports free-energy profiles of CO dimerization on copper from large explicit-solvent simulations, finding the reaction weakly sensitive to surface charge and cation identity except at strongly negative charge, and more favorable on stepped Cu(310). If these findings hold, they would guide how CO2 electroreduction catalysts should be designed and show that open ML potentials make such simulations routine.","feed_headline":"Charge and cations barely steer CO dimerization on copper","feed_subtitle":"A 7.8-million-calculation open dataset of electrified interfaces reveals that surface steps, not ions, may matter most in CO2 reduction.","key_machinery":"The central object is the OC25 dataset: 7,801,261 density-functional-theory energies and forces sampled from off-equilibrium configurations of 88 elements, eight solvents, nine ions, and 98 adsorbates at solid-liquid interfaces. The models trained on it are graph neural network potentials; their predicted energies and forces are accurate to about 0.1 eV and 0.015 eV/Å, and they are used to run explicit-solvent simulations with cells of more than 800 atoms for up to 7 ns. A drift-filtering criterion of 1 eV/Å and a pseudo-solvation-energy metric round out the methodology.","core_discovery":"The paper's central claim is that machine-learned potentials trained on the OC25 dataset become practical tools for explicit-solvent electrocatalysis, and that their application to CO dimerization on Cu surfaces shows weak sensitivity to surface charge and cation identity, with appreciable stabilization only at the most negative charge densities, while the stepped Cu(310) surface provides a more favorable pathway at modest reducing potentials.","pith_inferences":["A reader checking the paper will notice the body is devoted to dataset construction and benchmark tables; the CO dimerization free-energy profiles reported in the abstract do not appear in the manuscript, so reproducing them from the released models and dataset is the immediate next step.","If the weak cation dependence holds, the well-known alkali-cation effects on CO2 reduction selectivity would have to operate on later steps, such as protonation or desorption, rather than on the dimerization transition state.","The dataset's off-equilibrium sampling strategy could be reused to train potentials for other charged interfaces, such as battery electrode/electrolyte systems.","The claimed step-facet advantage, if robust, connects to experimental observations that roughened or defect-rich copper surfaces often shift product selectivity, making facet-resolved kinetic models a testable extension."],"forward_implications":["If the models' transferability holds, electrocatalytic transformations at solid-liquid interfaces can be simulated with explicit solvent at timescales orders of magnitude beyond ab initio methods.","The claimed weak charge and cation sensitivity of CO dimerization implies that cation promotion in CO2 reduction must act on other steps, not the initial C–C coupling.","The stepped-facet result suggests geometric surface engineering, not just electrolyte tuning, is a way to lower the dimerization barrier.","The force-drift analysis indicates that moderately loose DFT convergence can be used in training without degrading force accuracy, lowering the cost of future datasets.","Making OC25 publicly available lets the community benchmark and improve models on solid-liquid interfaces directly."],"fun_headline_variants":["Charge and cations barely steer CO dimerization on copper","Machine learning reveals Cu steps, not ions, favor CO coupling","7.8M calculations show CO dimerization blind to cation identity","Stepped Cu surfaces beat charge in CO dimerization simulations","Open Catalyst 2025: scale-up changes the CO coupling story"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire CO dimerization conclusion rests on the unstated assumption that ML potentials trained on roughly 144-atom off-equilibrium configurations, with energy errors around 0.1 eV, remain accurate enough in much larger explicit-solvent cells under electrochemical conditions to resolve the small free-energy differences between facets and charge states; the paper does not validate this transfer.","fun_headline_variants_meta":{"raw":{"variants":["Charge and cations barely steer CO dimerization on copper","Machine learning reveals Cu steps, not ions, favor CO coupling","7.8M calculations show CO dimerization blind to cation identity","Stepped Cu surfaces beat charge in CO dimerization simulations","Open Catalyst 2025: scale-up changes the CO coupling story"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1239,"prompt_tokens":651,"completion_tokens":588,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":395,"completion_tokens_details":{"reasoning_tokens":503}},"tokens_in":395,"tokens_out":588,"duration_ms":4632,"temperature":1.0,"reasoning_tokens":503,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T15:48:29.811586+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the claimed 7-ns enhanced-sampling simulation on Cu(100) and Cu(310) using the released OC25 checkpoint, and confirm that the reported free-energy ordering survives; then run conventional AIMD on one representative state to check whether the 0.1 eV model error flips the barrier difference. A simpler check: the body of the paper contains none of the dimerization free-energy profiles, so finding their data or scripts would be the first confirmation.","supporting_citations":[],"review_version":1}