{"id":"3ffd8211-b4ed-4ec6-96ee-5d765e47057b","arxiv_id":"2507.19307","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SSAGEN decouples crystal information generation from symmetry-constrained coordinate optimization, producing more stable structures than CDVAE and yielding 24 candidate photocatalysts for water splitting.","lead":"This paper introduces SSAGEN, a two-stage generative model that first proposes a crystal's composition, lattice, and space group, then relaxes atomic positions under symmetry constraints to favor stable structures. The authors apply it to photocatalytic water splitting, generating 200,000 candidate structures and reporting 3,318 DFT-validated candidates and 24 promising materials.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline stability improvements may be based on average formation energy rather than energy above the convex hull; the 148% and 180% numbers need a defined protocol and error bars.","rationale":"The reader's weakest assumption is that the second stage reliably finds true stable coordinates via universal ML potentials; I agree this is a risk. However, the more immediate problem is that the paper's own evidence for 'thermodynamic stability' appears to be average formation energy, which is a different quantity from thermodynamic stability. Even if MatterSim were perfect, the reported numbers would not establish stability. The PWS screening does define stability via ΔEhull and ΔGpbx, but those are GNN predictions, not DFT, so the 95.6% DFT validation does not rescue the stability claim. My recommendation remains CONDITIONAL: the method is plausible and the application is useful, but the headline stability numbers must be re-derived with a hull-based metric and a fully specified protocol before acceptance. This does not move the reader's verdict, so verdict_should_be is UNCHANGED.","tokens_in":13035,"tokens_out":7391,"duration_ms":74350,"concrete_test":"Recalculate the stability comparison in Table S3 using energy above the convex hull rather than average Ef. For the same generated structure sets (or a random subset of at least 200 per model), compute ΔEhull relative to the same reference phases used by OQMD/MP, and compare the fraction of structures below 0.1 eV/atom for SSAGEN vs CDVAE. Also report the exact derivation of 148% and 180%, including the number of structures, the ML potential used, and standard deviations. If the hull-based improvement is substantially smaller than the Ef-based improvement, the abstract should be revised to claim 'lower average formation energy' rather than 'improved thermodynamic stability.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing issue is that the paper's central quantitative claims—'148% and 180% improvements in thermodynamic and kinetic stability'—are not attached to a defined stability metric. In Section 2.1, the evidence for thermodynamic stability consists of average ML-predicted formation energies (e.g., -1.035 vs -0.412 eV/atom for SSAGEN vs CDVAE on OQMD, Table S3) and a DFT average Ef of -0.41 eV/atom for a sample. A lower average Ef is not thermodynamic stability: a structure can have a very negative formation energy and still lie above the convex hull, decomposing into a more stable mixture. The kinetic-stability evidence is 14 vs 4 phonon-stable structures out of 25 latent vectors, a sample far too small to support a 180% improvement or meaningful error bars. The related foundational premise in Section 1—that stable coordinates are 'uniquely determined' by crystal information—is only sparsely validated; MatterSim energies are not benchmarked against DFT hull distances for the 200,000 generated chemistries. Likewise, the PWS 'DFT validation' of 95.6% is based on calculated band gaps and edges, while the ΔEhull and ΔGpbx filters that define stability remain GNN predictions. If the SI defines 148%/180% differently (e.g., using ΔEhull), that protocol must be stated; as written, the headline overclaims.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces SSAGEN, a two-stage generative framework for crystal structure design. The first stage (CIVAE) generates crystallographic information—space group, lattice parameters, and composition—via a VAE with an autoregressive composition network, trained on OQMD. The second stage (SCO) enumerates Wyckoff positions and optimizes coordinates using universal machine learning potentials (MatterSim, M3GNet, CHGNet) with global and local optimization under symmetry constraints. The authors report that SSAGEN outperforms CDVAE in thermodynamic and kinetic stability (148% and 180% improvements), achieves high space-group fidelity, and reproduces known Si polymorphs and the CuAl convex hull. They then integrate SSAGEN into their MAGECS pipeline for photocatalytic water splitting (PWS), generating 200,000 structures, filtering by band gap, ΔEhull, and ΔGpbx, and obtaining 3,318 candidates. DFT calculations are used to validate band gaps and band edges, and 24 phonon-stable candidates are identified as optimal PWS photocatalysts. The paper claims a 95.6% DFT validation rate for PWS candidates and identifies 24 promising materials.","tokens_in":13313,"tokens_out":2938,"duration_ms":29628,"significance":"If the central claims hold, the paper makes a useful methodological contribution: decoupling crystal-information generation from coordinate optimization, with symmetry and Wyckoff constraints built into the optimization, is a sensible way to address the stability and symmetry limitations of generative models that directly generate coordinates. The integration with a property-optimization loop (BSA + three GNN predictors) and the identification of 24 phonon-stable, band-edge-aligned candidates is a concrete demonstration of the pipeline's practical value. The paper also provides a useful comparison of band-gap calculation methods (mBJ vs HSE vs GW) in the context of training ML models. The reconstruction fidelity results (91% space-group accuracy, 89% composition accuracy vs CDVAE) are clearly quantified and reproducible in principle. However, the headline quantitative claims are not yet supported with defined protocols or uncertainty estimates.","major_comments":[{"comment":"The headline improvements of '148% and 180%' in thermodynamic and kinetic stability are not attached to a defined metric. The evidence in Section 2.1 consists of average ML-predicted formation energies (Ef) and a phonon-stability count of 14 vs 4 out of 25 latent vectors. A lower average Ef is not thermodynamic stability: a structure with negative Ef can decompose into a more stable mixture and thus lie above the convex hull. The kinetic-stability sample of 25 latent vectors is far too small to support a 180% improvement or meaningful error bars. The manuscript must specify which metric defines each percentage (e.g., average Ef, ΔEhull, or phonon-stable fraction), report distributions with confidence intervals, and ideally report ΔEhull statistics for both SSAGEN and CDVAE.","section":"Abstract; Section 2.1; Conclusion"},{"comment":"The claim that 'DFT validation confirms 95.6% structures satisfy PWS requirements' is overstated. The DFT validation shown in Figure 5a-b addresses band gaps and band edges only. The stability filters ΔEhull ≤ 0.1 eV/atom and ΔGpbx ≤ 1.5 eV/atom remain GNN predictions for the 3,318 candidates. Thus, thermodynamic and aqueous stability are not independently confirmed for the full candidate set. The authors should report DFT-computed ΔEhull for a representative random subset (or all 3,318, if feasible) and clearly state what the 95.6% figure actually validates.","section":"Section 2.3; Figure 5; Table 1"},{"comment":"The load-bearing premise stated in the introduction—'Once the crystal information is defined, the most stable atomic coordinates and several metastable configurations are uniquely determined'—requires stronger support. The SCO relies on MatterSim and other universal ML potentials for energy and force evaluation, but the paper does not benchmark these potentials against DFT hull distances for the 200,000 generated chemical combinations. The DFT checks in Section 2.1 and 2.3 cover only a selected subset, mainly for band gaps and a few phonon spectra. Without a systematic comparison of UMLP-optimized structures against DFT-relaxed structures and DFT hull distances, the stability claims for the unvalidated majority of generated structures do not follow. At minimum, provide a random-sample benchmark of MatterSim energies against DFT for the composition and space-group distribution actually generated.","section":"Section 1; Section 2.2; Figure 3"},{"comment":"The band-edge screening and the reported DFT validation share a circular dependence. The empirical electronegativity formula (Eq. 1) computes ECB and EVB directly from Eg, so agreement between GNN-predicted and DFT-calculated ECB/EVB is partly forced by agreement on Eg; it is not an independent check of band-edge positions. To substantiate the band-edge alignment claim, the authors should compute absolute band alignments (e.g., with slab or vacuum-level approaches) for a subset of the 24 candidates and compare those directly with the empirical-formula values.","section":"Section 2.2, Eq. (1); Section 2.3, Figure 5"},{"comment":"The kinetic-stability comparison is based on phonon calculations for only 25 randomly selected latent vectors (14 stable for SSAGEN vs 4 for CDVAE). This sample is too small to support a 180% improvement or to draw quantitative conclusions about dynamical stability. The authors should either increase the sample size or report binomial confidence intervals and phrase the comparison qualitatively.","section":"Section 2.1, Figure 2f; Section 2.3"}],"minor_comments":[{"comment":"The manuscript reports MAE = 0.156 eV between predicted and DFT-calculated band gaps, but it does not report the number of structures used for this comparison. Please include the count and, if possible, a breakdown by composition type.","section":"Section 2.3, Figure 5a-b; Table 1"},{"comment":"There is a typo in the main text: 'colledcted' should be 'collected'. Also, the abbreviation 'Calu.' in Table 1 is nonstandard; use 'Calc.' for consistency.","section":"Section 2.1, Figure S10a; Methods"},{"comment":"The statement that BSA-identified structures 'such as SiTiO3, ZnO, TiO2' were 'previously validated experimentally' is ambiguous: please clarify whether these exact structures (with identical space groups and compositions) were generated and matched experimentally known phases, or whether they are merely chemically similar.","section":"Section 2.3, text near Figure S12"},{"comment":"The criterion 'eliminates structures with more than three species' is applied before DFT validation, but the rationale is given only as 'reduce the difficulty of experimental synthesis.' Please specify whether this threshold is standard in the PWS literature or an arbitrary choice.","section":"Section 2.3, selection criteria"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a potentially useful framework, but the two headline numbers (148%, 180%) and the 95.6% validation rate are not yet supported with defined protocols. The circularity between the empirical band-edge formula and the DFT 'validation' of band edges, and the reliance on unbenchmarked UMAP energies for the majority of generated structures, are the core technical concerns. I would encourage the editor to require the authors to either provide the missing metrics (ΔEhull distributions, DFT hull checks on a random subset, phonon sample with confidence intervals) or tone down the quantitative claims accordingly. The structural-generation ideas are publishable after these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is worth a look. Splitting generation into crystal information (VAE) and coordinate optimization under Wyckoff constraints with universal ML potentials is a real departure from end-to-end diffusion models like CDVAE. The reconstruction numbers are strong, and the PWS application—200,000 generated structures, 24 phonon-stable candidates with DFT band gaps—is a meaningful demonstration, even if many candidates are known compounds.\n\nThe soft spots are real. The headline 148% and 180% stability improvements come from average formation energies, not energy above the convex hull. A lower average Ef does not mean a structure is thermodynamically stable; it could still decompose into a more stable mixture. The kinetic-stability claim rests on 14 vs 4 phonon-stable structures out of 25 latent vectors—nice to see, but far too small for a percentage improvement. Neither number has a defined evaluation protocol or error bars. The paper should report hull distances and a larger, pre-registered phonon sample.\n\nThe premise that crystal information uniquely determines stable coordinates is load-bearing and only sparsely validated. MatterSim and other UMLPs are used to relax 200,000 compositions, but there is no benchmark of UMLP hull distances against DFT for those chemistries. DFT checks focus on band gaps; the ΔEhull and ΔGpbx filters that define stability remain GNN predictions. That is not fatal, but it means the stability claims for the unvalidated majority do not follow.\n\nThe band-edge validation is partly circular. ECB and EVB are computed from the empirical electronegativity formula using Eg; the screening objective uses the same formula, and the DFT validation reuses it for the band edges. So the 95.6% agreement is partly forced by Eg agreement. You need independent band-alignment calculations or a different formula to claim validation.\n\nAlso: no code or data release, and the paper mentions experimental validation without reporting any experiment. For a method paper, that is a significant gap.\n\nBottom line: the architecture is promising and the application is a serious effort, but the central quantitative claims are not yet supported as written. I would send it to peer review—the method deserves referee time—but I would not cite it yet. A major revision that defines the stability metrics, adds hull-distance and independent band-edge validation, and releases the pipeline would make this a solid contribution.\n\nBring it to reading group.","headline":"Two-stage generate-then-optimize pipeline is genuinely new, but headline stability numbers need a defined metric and the band-edge validation is partly circular.","tokens_in":13862,"tokens_out":2908,"would_cite":false,"duration_ms":28573,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["82D25","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"SSAGEN decouples crystal information from atomic coordinates, so stable, symmetry-correct structures can be generated without retraining, and applies this to find 24 DFT-validated photocatalytic water-splitting candidates.","keywords":["inverse design","crystal structure generation","generative model","crystal symmetry","Wyckoff positions","machine-learned potentials","photocatalytic water splitting","stability"],"falsifier":"Take a random sample of the generated structures that were not among the 3,318 DFT-checked candidates, and compute their formation energies relative to the convex hull and their phonon spectra with converged density-functional-theory calculations. If a substantial fraction of that sample falls above the hull or shows imaginary-frequency modes, the claim that the coordinate optimizer finds stable arrangements from crystal information alone would fail for the majority of generated structures.","tokens_in":12807,"feed_emoji":"⚛️","tokens_out":13581,"duration_ms":117843,"temperature":0.7,"pith_summary":"The paper claims that the hard part of inverse materials design—generating a new crystal that is actually stable and has the requested symmetry—can be split into two pieces: first generate only the crystal information (lattice parameters, composition, space group), then recover the atomic coordinates by constrained energy minimization. The generative model, called SSAGEN, learns the distribution of crystal information from a large database of computed crystal structures, while a second module enumerates symmetry-allowed Wyckoff positions and relaxes them with universal machine-learned interatomic potentials, so every output automatically carries the target space group and composition. Relative to a leading diffusion-based generative baseline, SSAGEN improves thermodynamic and kinetic stability of generated structures by 148% and 180%, and 94.9% of its structures have nontrivial symmetry. In the photocatalytic water-splitting application, 200,000 structures were generated; 3,318 passed all stability and band-gap filters; DFT validation confirmed 95.6% of those meet the requirements; and 24 phonon-stable candidates remain as optimal photocatalysts.","feed_headline":"Two-stage generation yields 24 water-splitting photocatalysts","feed_subtitle":"Decoupling crystal data from atomic coordinates lifts stability and screens 200,000 candidates down to 24","key_machinery":"The central machinery is a two-module pipeline. CIVAE (Crystallographic Information VAE) is a variational autoencoder whose decoder predicts the space group as a 230-way distribution, the six lattice parameters, and the composition as an autoregressive token sequence, all from a single latent vector; it is trained only to reconstruct crystal information, not coordinates. SCO (Stable Coordinates Optimizer) takes that output and enumerates the possible Wyckoff positions—the symmetry-distinguished sites atoms may occupy in the given space group—then performs global optimization with a bird-swarm search, local conjugate-gradient relaxation under fixed symmetry, primitive-cell reduction, and dynamic pruning of the search space, with energy and forces from a universal machine-learned interatomic potential. The identity doing the work is the crystallographic constraint structure: fixing space group, lattice, and composition restricts atomic degrees of freedom so sharply that only a few coordinates need optimization (e.g., two coordinates for ZnS), which is what lets the optimizer both preserve symmetry and reach low energies. Dynamic adjustment reduces the effective search space by 60–85% depending on symmetry.","core_discovery":"SSAGEN's central discovery is that stable coordinates do not have to be generated; they can be derived. The paper asserts that once lattice, composition, and space group are fixed, the most stable atomic arrangement—and a handful of metastable ones—is in effect determined, so the generator only needs to produce diverse but physically plausible crystal information. The coordinate step then becomes a constrained optimization problem: enumerate all Wyckoff position assignments compatible with the target space group, optimize them hierarchically with universal machine-learned potentials, reduce to primitive cells, and prune the search space dynamically. This yields structures whose symmetry is exact by construction, not enforced as a soft constraint, and whose energies are minimized rather than merely plausible. On the water-splitting target, the same loop with property predictors for band gap, energy above hull, and aqueous stability produced a 15-fold enrichment in viable candidates relative to the training set and 24 DFT- and phonon-validated photocatalysts.","pith_inferences":["If the core premise holds, then the rate-limiting step in inverse design shifts from generating coordinates to knowing the potential-energy surface: every improvement in universal interatomic potentials should translate into better crystals without retraining the generative model.","The paper's DFT validation covers a selected subset that already passed predictors; a natural stress test is to run converged DFT relaxation and phonons on a random draw of the generated structures that were not among the 3,318 checked candidates and compare pass rates.","The same information-then-optimize decomposition can in principle be reused for other target properties—mechanical, electronic, or magnetic—by keeping the crystal-information generator fixed and swapping the property predictors that steer the search.","The 24 candidates are computationally screened, not yet experimentally synthesized; actual photocatalytic performance under illumination and in aqueous electrolytes remains an open empirical question."],"forward_implications":["A single pretrained SSAGEN can produce crystals with a requested composition, space group, and lattice without fine-tuning, because the target information is fed directly into the coordinate optimizer rather than learned again.","Stability and symmetry become properties of the generation pipeline rather than properties to be checked after the fact, so screening resources can focus on function instead of filtering out invalid structures.","The water-splitting loop enriches the candidate pool roughly 15-fold relative to the training set and recovers known photocatalysts as well as 24 new phonon-stable candidates, suggesting the same pipeline can be retargeted to other functional properties.","Because 10 of 11 experimentally known silicon polymorphs were reproduced, the optimizer's ability to capture metastable structures may make SSAGEN useful not only for ground-state discovery but for exploring a material's structural landscape."],"supporting_citations":[{"why":"the diffusion-autoencoder generative model serving as the primary baseline for stability and symmetry comparisons.","marker":"[18]"},{"why":"supplies the training structures for the crystal-information variational autoencoder.","marker":"[31]"},{"why":"the bird-swarm algorithm used as the global optimizer for Wyckoff position search.","marker":"[32]"},{"why":"the universal machine-learned potential that supplies energies and forces during coordinate relaxation.","marker":"[35]"},{"why":"sets the band-edge and band-gap conditions a material must meet for photocatalytic water splitting.","marker":"[39]"},{"why":"provides the solid-aqueous equilibrium scheme behind the aqueous stability metric.","marker":"[41]"},{"why":"the structural and stability data used to train the property predictors in the inverse-design loop.","marker":"[45]"}],"fun_headline_variants":["Symmetry-assured AI whittles 200k down to 24 photocatalysts","Decoupled generation yields 24 stable water-splitting photocatalysts","Generative design ensures stability, finds 24 viable photocatalysts","Two-stage framework screens 200k to 24 proven photocatalysts","Inverse design with symmetry: 24 photocatalysts from 200k"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach rests on the assumption that once the lattice, chemical composition, and space group are set, the correct stable arrangement of atoms is essentially determined and can be found by the machine-learned potential, even for chemical combinations the potential has not been validated on.","fun_headline_variants_meta":{"raw":{"variants":["Symmetry-assured AI whittles 200k down to 24 photocatalysts","Decoupled generation yields 24 stable water-splitting photocatalysts","Generative design ensures stability, finds 24 viable photocatalysts","Two-stage framework screens 200k to 24 proven photocatalysts","Inverse design with symmetry: 24 photocatalysts from 200k"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000381,"raw_usage":{"total_tokens":2060,"prompt_tokens":1023,"completion_tokens":1037,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":937}},"tokens_in":639,"tokens_out":1037,"duration_ms":9447,"temperature":1.0,"reasoning_tokens":937,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:55:14.959979+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of the generated structures that were not among the 3,318 DFT-checked candidates, and compute their formation energies relative to the convex hull and their phonon spectra with converged density-functional-theory calculations. If a substantial fraction of that sample falls above the hull or shows imaginary-frequency modes, the claim that the coordinate optimizer finds stable arrangements from crystal information alone would fail for the majority of generated structures.","supporting_citations":[{"cited_title":"& Kim, J","cited_arxiv_id":null,"evidence_quote":"the diffusion-autoencoder generative model serving as the primary baseline for stability and symmetry comparisons."},{"cited_title":"& Günnemann, S","cited_arxiv_id":null,"evidence_quote":"supplies the training structures for the crystal-information variational autoencoder."},{"cited_title":"However, the search space of Wyckoff positions can be prohibitively large—for instance, Li ₂FeO₃ within space group 21 yields 10 ⁶ distinct Wyckoff configur ations","cited_arxiv_id":null,"evidence_quote":"the bird-swarm algorithm used as the global optimizer for Wyckoff position search."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"the universal machine-learned potential that supplies energies and forces during coordinate relaxation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"sets the band-edge and band-gap conditions a material must meet for photocatalytic water splitting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the solid-aqueous equilibrium scheme behind the aqueous stability metric."},{"cited_title":"A., Waldwick, B., Lazic, P","cited_arxiv_id":null,"evidence_quote":"the structural and stability data used to train the property predictors in the inverse-design loop."}],"review_version":1}