{"id":"4a5a38ca-ed77-45f3-b99f-2c4c1b7374d5","arxiv_id":"2608.08848","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"AdsZeo stores over 1.2 billion saved particle coordinates from 62,075 GCMC simulations of methane in 4,775 sodium zeolite realisations, with validation showing internal consistency.","lead":"This paper releases AdsZeo, a coordinate-resolved dataset of methane adsorption in 4,775 sodium zeolite structures from 191 framework topologies, covering 62,075 simulations at 13 pressures. It is designed for machine learning models that need molecular configurations rather than only average adsorption values.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-run convergence and force-field transferability are the weakest links, but they are disclosed and do not invalidate the dataset claim.","rationale":"Reader and I converge on the same load-bearing assumption: single-seed GCMC convergence and force-field transferability. The paper's own block-check statistics quantify the tail (max 7.935 CH4/uc; 95th percentile relative 5.26%), so this is not a manufactured risk. However, the limitation is explicitly disclosed in Sections 2.3 and 4.1, the coordinate-derived loadings match RASPA scalars to 0.18 CH4/uc mean absolute deviation (Section 4.2), all isotherms are monotonic, and the three experimental comparisons are consistent within the acknowledged structural idealisation. A dataset intended for ML configuration generation can tolerate these approximate equilibrium labels, provided users are told to filter by block deviation; the usage notes already advise leakage-aware splits. The Figure 4 body/caption inconsistency is a real editorial defect but does not affect the deposited data or the central claim. I therefore see no reason to change the ACCEPT verdict, though an author response should reconcile the figure numbers and ideally add per-run convergence flags as metadata.","tokens_in":12032,"tokens_out":12690,"duration_ms":124140,"concrete_test":"Select the 50 simulations with the largest absolute block deviations from the deposited DuckDB and re-run each with 3 independent seeds and 1,000,000 production cycles. Compare each seed's mean loading and total energy to the published scalar records. If any re-run mean differs by more than the Section 4.2 RMS deviation (0.24 CH4/uc) plus the inter-seed standard error, those saved frames are not demonstrably equilibrated and the dataset should include a convergence flag or drop the equilibrium label for the affected runs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that AdsZeo stores equilibrium GCMC configurations. The least secure condition is that every one of the 62,075 single-seed runs (Section 2.3) has converged and that the force field assembled from Refs. 24-26 transfers to all 191 topologies, including Al-O-Al environments. Section 4.1's own block statistics show the tail is not tightly converged: the 95th percentile absolute relative block deviation is 5.26%, and the largest absolute block deviation is 7.935 CH4 per unit cell. For a run whose mean loading is near the 4 CH4/uc cutoff used for relative statistics, a 5% block drift means some saved frames are not demonstrably representative equilibrium samples. The external check (Section 4.5) covers only Na-ZSM-5, NaY, and NaX, so transferability to the other 188 topologies is plausible but not established by direct validation. None of this falsifies the dataset as a resource, because the limitations are disclosed and the coordinate-vs-scalar consistency check (Section 4.2) is strong. A separate internal inconsistency should be fixed: Section 4.1 text gives example means 32.3/50.5/60.0 CH4 while Figure 4 labels show 36.0/86.8/120.9; the figure and body need reconciliation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents AdsZeo, a coordinate-resolved dataset of grand-canonical Monte Carlo (GCMC) methane adsorption configurations in aluminium-substituted, sodium-exchanged zeolites. The authors generate 4,775 framework realisations from 191 IZA topologies, simulate each at 13 pressures between 0.1 and 100 bar at 298 K with RASPA, and store scalar adsorption records (isotherms, frame-level energy/counts) together with saved production-frame Cartesian coordinates of methane pseudo-atoms and Na+ cations in a DuckDB database. The paper describes the structure-generation workflow, the assembled force field, the GCMC protocol, the database schema, and four internal validation checks (block statistics, coordinate-derived loading consistency, pair-distance geometry, monotonicity of isotherms) plus a comparison with experimental methane isotherms for Na-ZSM-5, NaY, and NaX. The central claim is that the released coordinates represent equilibrated GCMC samples across a broad range of framework chemistry and pressure, suitable for training machine-learning models for configuration generation.","tokens_in":12308,"tokens_out":8382,"duration_ms":89022,"significance":"If the dataset meets its claims, it fills a genuine gap in the high-throughput adsorption literature, which has mostly archived scalar properties; the coordinate-level records would enable generative and density-estimation models for charged nanoporous hosts. The paper's internal validation is appropriately matched to the dataset claim: the coordinate-derived loadings reconstruct the RASPA scalar loadings with a mean absolute deviation of 0.18 CH4 per unit cell, no short-range coordinate artefacts are reported, all isotherms are monotonic, and the dataset and generation code are publicly available. The main risk to the equilibrium claim is the per-run convergence tail documented in Section 4.1, which is disclosed but not resolved at the level of individual runs. On balance, the resource is valuable and the weaknesses are largely addressable.","major_comments":[{"comment":"The paper labels all 62,075 runs as equilibrium samples, but each run used a single random seed and the block statistics in Section 4.1 show a non-negligible tail of runs with drift: the 95th percentile absolute relative block deviation is 5.26% and the largest absolute block deviation is 7.935 CH4 per unit cell. Because the central claim is that the saved coordinates are equilibrium configurations, the release should include a per-run convergence diagnostic (e.g., block-deviation values or a pass/fail flag) and the text should state the fraction of runs satisfying a defined convergence criterion; otherwise the 'equilibrium' label is not verifiable for the outlier runs.","section":"§2.3, §4.1"},{"comment":"The force field assembled from Refs. 24-26 is applied to all 191 topologies, including Al-O-Al environments, but the external validation in Section 4.5 covers only Na-ZSM-5, NaY, and NaX. The paper should either provide additional transferability evidence for representative topologies outside these three systems or explicitly state that the 'realistic methane adsorption behaviour' claim at the end of Section 4.5 is established only for the tested systems and is a transferability assumption elsewhere.","section":"§2.2, §4.5"}],"minor_comments":[{"comment":"The body text gives example mean methane counts of 32.3, 50.5, and 60.0 CH4 for the 1, 10, and 100 bar traces, while the Figure 4 labels show 36.0, 86.8, and 120.9 CH4; these numbers must be reconciled so that the primary production-stability figure is self-consistent.","section":"§4.1, Figure 4"},{"comment":"Section 2.1 states that Löwenstein's rule was not explicitly enforced, while Section 4.5 says Al substitutions were generated 'subject to the non-Loewenstein constraint'; please clarify whether this means the same thing, since the phrase 'non-Loewenstein constraint' could be read as enforcing absence of Al-O-Al rather than allowing it.","section":"§2.1, §4.5"},{"comment":"There is a typo in 'near-zere methan loading'; it should read 'near-zero methane loading'.","section":"§4.4"},{"comment":"The sentence 'do not show a appreciable, systematic first-to-last-block drift' contains a grammar error ('a appreciable'); it should read 'no appreciable systematic drift'.","section":"§4.1"},{"comment":"Figure 2 lists 'Structures / isotherms 4,775' while Table 3 reports 1,489,800 rows in the isotherms table; please clarify that the former counts distinct isotherms while the latter counts scalar records, to avoid an apparent inconsistency.","section":"Figure 2, Table 3"},{"comment":"The 13 pressures are only described as 'logarithmically spaced between 0.1 and 100 bar'; for reproducibility, please list the exact pressure values or point to where they are stored in the database schema.","section":"§2.3"}],"recommendation":"major_revision","confidential_remarks":"This is a solid dataset contribution with strong internal consistency checks and public code and data. My main request is to make the equilibrium claim verifiable per run by adding convergence diagnostics to the released database, and to reconcile the numerical inconsistency in the Section 4.1 figure/text. These are fixable within the manuscript's scope and do not require rerunning the production simulations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"AdsZeo is the first dataset I know that gives coordinate-level GCMC samples for methane in charged zeolites at this scale: 191 topologies, 4,775 Al/Na realisations, 62,075 runs, 200 saved frames per run, 1.25 billion coordinate rows. That is the genuinely new piece, and the paper makes a fair distinction against MOFX-DB and Crafted (scalar isotherms) and Open DAC (relaxed placements, not pressure-swept equilibrium frames). The internal validation is the strongest part. Recomputing loadings from saved coordinates matches RASPA scalars with mean absolute deviation 0.18 CH4 per unit cell; isotherms are monotonic; pair-distance checks show no wrapping or overlap artefacts; charge neutrality is checked. Data and code are deposited on Zenodo and GitHub, with a usage notebook, so the resource is immediately usable. This is a solid dataset paper. Where it is soft, in proportion: the equilibrium claim rests on a single GCMC run per state point. No repeated seeds, so the reported uncertainties in the isotherms table are intra-run statistics, not run-to-run variation. The block analysis shows stationarity on average, but the tail is not tight: largest block deviation 7.935 CH4 per unit cell and 95th percentile relative deviation 5.26% for runs above 4 CH4 per unit cell. For a run near that cutoff, a 5% drift means some saved frames are not demonstrably representative equilibrium samples. The force field is inherited from previously published parameters, including the group's own earlier work, which is legitimate here because those parameters are citable and independently checkable, but external validation covers only Na-ZSM-5, NaY, and NaX. Transfer to the remaining 188 topologies and Al-O-Al environments is plausible but not directly demonstrated. These limits are disclosed in the text and do not falsify the dataset as a resource, but users training generative models should condition on run-level quality and may want to subset the poorly converged tail. One small correctness issue: Section 4.1 says example traces have means 32.3, 50.5, and 60.0 CH4, while Figure 4 labels the same traces as 36.0, 86.8, and 120.9. That is an internal inconsistency that needs fixing. Also worth noting: zero-loading frames are kept, which is the right call for GCMC, but it means ML users need to handle variable particle counts. Overall, this deserves serious refereeing and, assuming the deposited files verify, publication. The coordinate-vs-scalar consistency check is exactly the evidence a dataset paper should provide. I would bring it to a reading group focused on ML for molecular simulation, and I would cite it if I were building generative models for adsorption.","headline":"Coordinate-resolved GCMC adsorption dataset that is genuinely new and internally validated; accept after fixing a text/figure mismatch and being clear about single-seed convergence limits.","tokens_in":771,"tokens_out":1661,"would_cite":true,"duration_ms":35575,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"AdsZeo records equilibrium methane configurations, not just averaged loadings, for 4,775 zeolite framework variants.","keywords":["adsorption","zeolites","grand canonical Monte Carlo","methane","molecular configurations","coordinate-resolved dataset","machine learning","sodium cations"],"falsifier":"Rerun the state point with the largest reported block deviation, 7.935 methane per unit cell, from new random seeds and compare the per-frame distribution of methane counts and energies; if the new runs do not fall within the reported block statistics, the equilibrium-equivalence claim for that framework realisation fails.","tokens_in":11856,"feed_emoji":"💎","tokens_out":10600,"duration_ms":103353,"temperature":0.7,"pith_summary":"AdsZeo is a coordinate-resolved dataset of equilibrium methane adsorption configurations in aluminium-substituted, sodium-containing zeolites. It saves the actual Monte Carlo coordinates of every methane pseudo-atom and mobile sodium cation at 13 pressures between 0.1 and 100 bar, alongside per-frame energy and loading statistics, rather than only ensemble-averaged isotherms. The paper argues that this configuration-level archive, spanning 191 zeolite topologies and 4,775 framework realisations from 62,075 grand canonical Monte Carlo simulations, fills a gap left by high-throughput adsorption databases that discard molecular configurations. If the saved frames are genuine equilibrium samples, the dataset provides a training substrate for machine-learning models that generate adsorption configurations or density fields in charged zeolite pores.","feed_headline":"1.25 billion coordinates map methane adsorption in zeolites","feed_subtitle":"AdsZeo pairs equilibrium frames with scalar records for 4,775 zeolite variants, built for ML configuration models.","key_machinery":"The load-bearing mechanism is a coordinate-preserving grand canonical Monte Carlo protocol: each framework realisation is equilibrated for 5,000 cycles and run for 200,000 production cycles, with one full configuration frame saved every 1,000 cycles. Each saved frame stores Cartesian coordinates of united-atom methane pseudo-atoms and mobile sodium cations together with per-frame counts and energy components, and all frames are linked by run and structure identifiers to scalar isotherm records in a processed database. This pairing of coordinates with scalar averages is what lets the dataset be checked for consistency and what makes it usable for configuration-level machine learning.","core_discovery":"The central claim is that AdsZeo provides coordinate-resolved adsorption data across variations in framework topology, aluminium content and distribution, sodium cation arrangement, pressure, and methane loading. The processed release contains 12,415,000 saved production frames and 1,245,376,215 saved particle-coordinate records. The paper validates these records by showing that loadings reconstructed from saved frames agree with the simulation software's scalar loadings to a mean absolute deviation of 0.1794 methane per unit cell, that block statistics show no systematic drift, that pair-distance checks show no unphysical overlaps, and that all isotherms increase monotonically with pressure. External comparisons with experimental methane isotherms for three sodium-exchanged zeolites reproduce the overall uptake trends, supporting the claim that the stored configurations are realistic equilibrium samples.","pith_inferences":["The saved mobile sodium coordinates make the database a potential resource for studying how cation positions respond to methane loading, a coupling the paper notes but does not itself analyse.","The force field is validated experimentally on only a few framework compositions, so a natural extension is targeted validation on additional aluminosilicate frameworks, especially those containing Al-O-Al environments.","Users modelling low-loading regimes should weigh the reported block-deviation statistics, with a largest absolute deviation of 7.935 methane per unit cell, when deciding which runs are converged enough for their task.","Coordinate-resolved datasets of this kind could be used to train conditional generative samplers that combine pressure, topology, and cation arrangement as inputs, potentially reducing the cost of future GCMC screening."],"forward_implications":["Machine-learning models can be trained to generate full adsorption configurations in charged zeolite pores, not just to predict scalar loadings.","Spatial statistics and density estimation of methane and sodium cation distributions become possible across 191 topologies and a 0.1 to 100 bar pressure range.","The saved frames can be used as equilibrium samples in benchmarks for configuration generation, provided data splits respect framework topology and run identity.","Zero-loading frames are retained as valid equilibrium samples, preserving the low-pressure onset of adsorption for modelling.","Because each topology has 25 aluminium-substitution realisations, the dataset supports studying how aluminium distribution and cation arrangement affect local adsorption geometry."],"supporting_citations":[{"why":"Supplies the united-atom methane model with the CH4-CH4 Lennard-Jones parameters used in every simulation.","marker":"[25]"},{"why":"Supplies the aluminosilicate framework and sodium cation parameters that set the adsorption energetics.","marker":"[26]"},{"why":"Provides the oxygen type for Al-O-Al environments and the non-Loewenstein force-field treatment the dataset follows.","marker":"[24]"},{"why":"Is the grand canonical Monte Carlo simulation engine whose production runs generate the saved frames and scalar records.","marker":"[5]"},{"why":"Provides the parent all-silica zeolite framework structures and the list of topologies selected for the dataset.","marker":"[22]"},{"why":"Converts the crystallographic unit cells into simulation-ready supercells with lattice vectors exceeding 24 angstroms.","marker":"[23]"},{"why":"Deposits the processed database and makes the dataset publicly available for reuse.","marker":"[27]"},{"why":"Provides experimental methane adsorption data for Na-ZSM-5 used in the external validation.","marker":"[28]"},{"why":"Provides experimental methane adsorption data for NaY used in the external validation.","marker":"[29]"},{"why":"Provides experimental methane adsorption data for NaX used in the external validation.","marker":"[30]"}],"fun_headline_variants":["AdsZeo: 1.25B methane coordinates across 4,775 zeolite variants","Coordinate-resolved adsorption data for zeolite ML","1.25B coordinates to train AI for zeolite gas adsorption","AdsZeo: 62k simulations, 1.25B coordinates to train zeolite models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The force field and the single-run GCMC protocol produce true equilibrium ensembles for every one of the 4,775 framework realisations, including those with aluminium distributions and Al-O-Al environments the parameters were not specifically fitted to.","fun_headline_variants_meta":{"raw":{"variants":["AdsZeo: 1.25B methane coordinates across 4,775 zeolite variants","Coordinate-resolved adsorption data for zeolite ML","1.25B coordinates to train AI for zeolite gas adsorption","AdsZeo: 62k simulations, 1.25B coordinates to train zeolite models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000889,"raw_usage":{"total_tokens":3845,"prompt_tokens":964,"completion_tokens":2881,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":2797}},"tokens_in":580,"tokens_out":2881,"duration_ms":22193,"temperature":1.0,"reasoning_tokens":2797,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:22:35.695406+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the state point with the largest reported block deviation, 7.935 methane per unit cell, from new random seeds and compare the per-frame distribution of methane counts and energies; if the new runs do not fall within the reported block statistics, the equilibrium-equivalence claim for that framework realisation fails.","supporting_citations":[{"cited_title":"Understanding the role of sodium during adsorption: a force field for alkanes in sodium-exchanged faujasites.Journal of the American Chemical Society, 126(36):11377–11386, 2004","cited_arxiv_id":null,"evidence_quote":"Supplies the united-atom methane model with the CH4-CH4 Lennard-Jones parameters used in every simulation."},{"cited_title":"Transferable force field for carbon dioxide adsorption in zeolites.The Journal of Physical Chemistry C, 113(20):8814–8820, 2009","cited_arxiv_id":null,"evidence_quote":"Supplies the aluminosilicate framework and sodium cation parameters that set the adsorption energetics."},{"cited_title":"Adsorption of carbon dioxide in non-l¨ owenstein zeolites.Chemistry of Materials, 35(13):5222–5231, 2023","cited_arxiv_id":null,"evidence_quote":"Provides the oxygen type for Al-O-Al environments and the non-Loewenstein force-field treatment the dataset follows."},{"cited_title":"RASPA: molecular simulation software for adsorption and diffusion in flexible nanoporous materials.Molecular Simulation, 42(2): 81–101, 2016","cited_arxiv_id":null,"evidence_quote":"Is the grand canonical Monte Carlo simulation engine whose production runs generate the saved frames and scalar records."},{"cited_title":"Elsevier, 2007","cited_arxiv_id":null,"evidence_quote":"Provides the parent all-silica zeolite framework structures and the list of topologies selected for the dataset."},{"cited_title":"marko-petkovic/porran: Porran, March 2025","cited_arxiv_id":null,"evidence_quote":"Converts the crystallographic unit cells into simulation-ready supercells with lattice vectors exceeding 24 angstroms."},{"cited_title":"AdsZeo: A dataset of equilibrium state configurations of adsorption in zeolite","cited_arxiv_id":null,"evidence_quote":"Deposits the processed database and makes the dataset publicly available for reuse."},{"cited_title":"Calorimetric heats of adsorption and adsorption isotherms","cited_arxiv_id":null,"evidence_quote":"Provides experimental methane adsorption data for Na-ZSM-5 used in the external validation."},{"cited_title":"Effect of cations on methane adsorption by nay, mgy, cay, sry, and bay zeolites.The Journal of Physical Chemistry, 97(49):12894–12898, 1993","cited_arxiv_id":null,"evidence_quote":"Provides experimental methane adsorption data for NaY used in the external validation."},{"cited_title":"Adsorption of methane on nax zeolite in the subcritical and supercritical regions.Bull","cited_arxiv_id":null,"evidence_quote":"Provides experimental methane adsorption data for NaX used in the external validation."}],"review_version":1}