{"id":"16893e92-2c95-42d0-9b72-564c163c27d6","arxiv_id":"2504.16893","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CHGGen combines a predefined symmetrized host with diffusion inpainting and CHGNet relaxation, increasing the rate of symmetric crystal generation relative to unconditional diffusion on Zn-P-S and Li-Si chemistries.","lead":"Crystal Host-Guided Generation (CHGGen) fixes a pre-symmetrized atomic framework and uses diffusion inpainting to place the remaining atoms, yielding more symmetric crystals than fully unconditional generation. The authors demonstrate the approach on Zn-P-S and Li-Si systems using the CHGNet machine-learned potential for screening before DFT verification.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed symmetry gain may be an artifact of very loose spglib tolerances used to symmetrize the host and to score final structures; no unsymmetrized-host control or stricter-tolerance reporting exists.","rationale":"I agree with the reader's weakest assumption: the load-bearing risk is that tolerance inflation in the host symmetrization step, combined with a loose final symmetry metric, manufactures the reported improvement in symmetry success rate. The paper's own Section IV limitation statement explicitly concedes that the symmetry refinement is preliminary and mostly yields moderate monoclinic symmetries, which supports this concern. The concern is not about disagreement with the broader field or about the use of machine learning potentials; it is about whether the central empirical comparison measures what it claims. The proposed test isolates the contribution of host symmetrization and of the final tolerance threshold, and it adds the missing statistical reporting. Because the reader already assigned a conditional verdict pending controls and statistical reporting, my stress-test does not change that verdict; it sharpens the specific control that would settle the issue. If the control later shows that the advantage survives an un-symmetrized host and stricter tolerance, the central claim would be substantially strengthened. Until then, the present evidence is insufficient to take the symmetry improvement at face value.","tokens_in":16540,"tokens_out":5883,"duration_ms":60298,"concrete_test":"Re-run the Fig. 4f / Fig. 6 symmetry-success comparison on both chemical systems with three changes: (1) add an inpainting control that uses the raw, un-symmetrized P1 framework as the host; (2) determine final space groups at symprec = 1e-2 Å and angle tolerance = 5°, alongside the current 0.1 Å / 10° setting; (3) report the number of generated structures per composition and bootstrapped 95% confidence intervals for every success rate. If the inpainting advantage over unconditional generation is no longer significant under stricter final tolerance, or if it disappears when the host is not symmetrized, the headline claim is an artifact of the tolerance choices rather than a genuine gain in symmetric structure generation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on a single empirical comparison: inpainting into a symmetrized host produces a higher fraction of non-P1/P-1 structures than unconditional generation. The comparison is vulnerable because the host is constructed by removing the species with the broadest coordination distribution and then applying spglib with stol up to 2.0 Å and angle tolerance up to 30° until a non-P1 space group is found (Methods, Section III B). At these tolerances spglib will assign a symmetric space group to a nearly arbitrary arrangement, so the 'refined' host may be a symmetrized projection of a disordered framework rather than a chemically meaningful scaffold. Inpainting then locks the host atoms to that projected template through Eqs. 7-8, so any symmetry inherited by the generated structure is imposed by construction. The final symmetry assignment uses site tolerance 0.1 Å and angle tolerance 10°, which is still much looser than the 10^-3 to 10^-2 Å commonly used, and the metric counts every non-P1/P-1 space group, including low-symmetry C2/Cm. The authors themselves concede in Section IV that this refinement 'remains preliminary' and 'predominantly yields structures with moderate symmetry (e.g., C2, Cm)'. There is no control experiment with an un-symmetrized host, no report of host symmetrization displacement magnitudes, and no error bars or generation counts. If the host symmetrization or the loose scoring tolerance is what creates the apparent success, the headline comparison does not demonstrate a practical improvement in symmetric crystal structure prediction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CHGGen, a conditional crystal generation framework that combines score-based diffusion inpainting with a symmetrized host structure and the CHGNet foundation potential for relaxation and thermodynamic screening. The workflow removes atoms with broad coordination distributions (Zn or Li), symmetrizes the remaining framework with spglib using progressively looser tolerances, inpaints guest atoms into that framework, and relaxes the resulting structures with CHGNet before selected r2SCAN DFT verification. The method is demonstrated on the Zn-P-S and Li-Si chemical systems. The central claim is that inpainting into a symmetrized host yields a higher fraction of non-P1/non-P-1 crystal structures than unconditional generation, with a DFT-confirmed stable Li5Si2 polymorph in the Li-Si system.","tokens_in":16889,"tokens_out":3252,"duration_ms":31643,"significance":"If the reported symmetry gain is robust, CHGGen would be a practical modular tool for host-guided structure generation, particularly for intercalation chemistry and partial-occupancy systems, and the availability of code is a genuine strength. The paper also gives a clean demonstration of using a foundation potential for relaxation and screening in a generative pipeline. However, the central comparison is currently weakened by the construction of the host, the loose spglib tolerances used both to create and to score symmetric structures, and the lack of statistical grounding. The authors themselves state in Section IV that the symmetry refinement approach 'remains preliminary' and that it 'relies on spglib by simply increasing the tolerance threshold,' which directly bears on the headline claim.","major_comments":[{"comment":"The headline symmetry-success comparison in Figures 4f and 6 is partly built into the algorithm. The host framework is symmetrized by spglib with site tolerance up to 2.0 Å and angle tolerance up to 30°, and during every reverse step Eq. (7) re-noises the host positions as x_host_{t-1} = x_host_0 + σ_{t-1} z, followed by the mask combination in Eq. (8). The final host therefore retains the imposed symmetry by construction, and the guest atoms are the only free part. To support the claim that inpainting 'generates a higher fraction of symmetric structures than unconditional generation,' the authors need at least an unsymmetrized-host control, a report of the root-mean-squared displacement between the initial relaxed framework and its symmetrized projection, and a success-rate evaluation at strict spglib tolerances (for example, site tolerance no larger than 0.01 Å). Without these, the apparent improvement may be an artifact of tolerance inflation rather than a genuine crystallographic ordering effect.","section":"III.B and Methods; Eqs. (7) and (8)"},{"comment":"The success-rate metric counts every non-P1/non-P-1 space group as success, including C2 and Cm, and the authors acknowledge in Section IV that the method 'predominantly yields structures with moderate symmetry (e.g., C2, Cm).' The comparison also lacks sample counts, error bars, and any statistical test, so the statement that inpainting achieves 'significantly higher' success rates is not quantitatively supported. Please report the number of generated structures per condition, the distribution of assigned space groups, and a stricter definition of success (for example, only the crystal system or Laue class expected from the intended host framework). A bootstrap confidence interval on the success-rate difference would also be appropriate.","section":"Figure 4f, Figure 6, and Methods (success rate)"},{"comment":"The manuscript contains an explicit limitation statement: 'the current symmetry refinement approach remains preliminary as it relies on spglib by simply increasing the tolerance threshold.' This concession is in direct tension with the abstract's claim that the inpainting method 'generates a higher fraction of symmetric structures than unconditional generation.' As written, the central claim is defensible only under the assumption that loosened spglib tolerances produce chemically meaningful scaffolds, which is exactly what the skeptical reader will question. The authors should either provide evidence that the symmetrized frameworks correspond to genuine crystallographic scaffolds (for example, by comparing them to known structure prototypes or by reporting displacement distributions), or reframe the headline claim as a demonstration of conditional generation conditioned on a user-supplied host rather than as a general symmetry-success improvement.","section":"Section IV (Discussion)"}],"minor_comments":[{"comment":"The caption contains a typo: 'GHGGen' should be 'CHGGen'.","section":"Figure 6 caption"},{"comment":"The library name 'pytorch-lighting' should be 'pytorch-lightning'.","section":"Methods, Model architecture"},{"comment":"The phrase 'often relys' should be 'often relies'.","section":"Section IV, first paragraph of final discussion"},{"comment":"The resampling 'jump back' step applies z~N(0,1) and then sets x_t ← x_{t-1} + sqrt(σ_{t-1}^2 - σ_{t-2}^2) z, but the text and reference do not clarify whether this is intended as the forward transition under the VE-SDE; a sentence explaining the consistency of this step with the sampler would improve reproducibility.","section":"Algorithm, Inpainting Generation"},{"comment":"The Li5Si2 (R-3m) structure is acknowledged as having been identified previously by Tipton et al. and Morris et al., so the paper should present it as a validation of the pipeline rather than as a de novo discovery; the current phrasing 'successfully predicted' is acceptable but could be sharpened to avoid overclaiming novelty.","section":"Section III.D"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the open code availability is a strength. However, the headline symmetry-success claim is currently not separable from the tolerance-induced imposition of symmetry via the host construction and Eqs. (7) and (8). I recommend major revision: if the authors cannot demonstrate robustness to stricter tolerances or provide an unsymmetrized-host control, the claim should be reframed as a conditional-generation framework demonstration rather than an improved symmetry success rate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent engineering paper that does what it claims—host-guided inpainting plus CHGNet relaxation yields more non-P1/P-1 structures than their unconditional baseline—but the headline comparison is partly a tautology. The host is symmetrized at very loose spglib tolerances (stol up to 2.0 Å, angle up to 30°) and then locked into place at every reverse-diffusion step via Eqs. 7–8. So the final structure inherits the host symmetry by construction; the guest atoms are the only free part. The symmetry-success metric then counts structures that pass a still-loose final tolerance (stol 0.1 Å, angle 10°). The authors themselves concede in Section IV that the refinement \"remains preliminary.\" That doesn't kill the paper, but it means the central bar chart in Fig. 4f is less convincing than it looks.\n\nWhat is genuinely new: the modular pipeline integrating RePaint-style inpainting with a foundation potential, demonstrated on Zn-P-S and Li-Si. The locality-bias analysis—RDFs and coordination environments showing that unconditional generation makes amorphous \"mosaics\" at large cell sizes—is a nice, physically grounded motivation. The code is on GitHub, the DFT verification is real, and the paper is refreshingly honest about limitations, including the Li5Si2 rediscovery and the systematic softening of CHGNet energies.\n\nSoft spots, in proportion: the missing control baseline is the main one. Without an unsymmetrized-host or fixed-tolerance comparison, the claimed improvement in symmetry success rate is not established as a property of the inpainting itself. There are also no sample counts or error bars on the success-rate bars, and the metric excludes only P1/P-1 (and Pm for Li-Si), so \"moderate symmetry\" (C2/Cm) carries the result. Two case studies is thin for a framework paper, though acceptable for a proof of concept.\n\nBottom line: as a tool for placing intercalants or modifying known frameworks, this is likely useful, and the integration is clean. It is not a breakthrough in symmetric crystal generation; the symmetry is imposed, not discovered. The paper deserves a serious referee—send it to review, but ask for a control baseline, stricter-tolerance reporting, and generation counts. I'd accept after moderate revision.","headline":"Useful engineering for host-guided inpainting, but the headline symmetry gain is partly baked in by the symmetrized host and loose spglib tolerances.","tokens_in":17417,"tokens_out":2866,"would_cite":false,"duration_ms":24818,"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":"Conditional inpainting on a symmetrized host rescues symmetric crystal generation by diffusion models at large cell sizes, and the pipeline finds a new stable Li-Si polymorph.","keywords":["crystal structure prediction","diffusion models","inpainting generation","crystal symmetry","intercalation chemistry","machine learning interatomic potentials","Li-Si alloys","foundation potentials"],"falsifier":"Take a host framework obtained at the loosest tolerances (site tolerance 2.0, angle tolerance 30 degrees), relax it with DFT or a tight-tolerance force field, and re-determine its space group at standard tolerances; if the refined symmetry collapses to P1 or P-1, the scaffold that guides inpainting is an artifact. A complementary test is to run inpainting into a deliberately randomized, P1 version of the same host and compare symmetry success rates and decomposition energies with the symmetrized version.","tokens_in":16329,"feed_emoji":"💎","tokens_out":8102,"duration_ms":69010,"temperature":0.7,"pith_summary":"Diffusion-based generative models for crystals learn local bonding environments well but fail to assemble them into coherent, symmetric crystals once the unit cell grows, a failure the paper traces to the locality bias of graph neural networks. To fix this, the paper proposes conditional generation: a host framework is built by removing the atom type with the most flexible coordination (Zn or Li) and symmetrizing the remainder with loosened symmetry-matching tolerances, and the removed atoms are then placed back by diffusion inpainting. On the ZnS–P$_2$S$_5$ and Li–Si systems, the inpainting route yields a higher fraction of structures with nontrivial space groups than unconditional generation. After relaxation with a universal machine-learning interatomic potential, the workflow finds a Li$_5$Si$_2$ polymorph that sits below the known convex hull, along with low-energy metastable polymorphs. The practical payoff is that, if the scaffold is chemically meaningful, the same machinery offers a route to intercalation compounds, structural modification, and symmetry-preserving generation at larger cell sizes.","feed_headline":"Host-guided inpainting restores symmetry in AI-generated crystals","feed_subtitle":"Conditional diffusion on a symmetric framework finds ordered structures and a new stable Li5Si2 polymorph.","key_machinery":"Score-based diffusion on fractional coordinates with a variance-exploding SDE, where a custom SE(3)-equivariant graph neural network predicts the denoising score. The conditional component is the inpainting loop: a binary mask separates framework atoms from guest atoms, the framework is re-noised at each step to the correct noise level and blended back in, and resampling repeatedly revisits earlier diffusion steps so the guest positions harmonize with the host. The host itself is constructed by removing the flexible-coordination species and symmetrizing the remainder with a symmetry finder at tolerances up to site tolerance 2.0 and angle tolerance 30 degrees. A pretrained universal interatomic potential relaxes every generated structure and supplies decomposition energies for screening, with r2SCAN DFT used for final verification.","core_discovery":"The paper's central claim is that the failure of unconditional diffusion models to produce symmetric crystals at large cell sizes is not fundamental to diffusion itself but follows from the locality of the graph neural network that provides the score: the model reproduces correct short-range motifs yet outputs disordered mosaics because it never couples atomic positions to long-range crystallographic order. The remedy is to supply that order externally. CHGGen first generates and relaxes candidate structures, removes the species with the broadest coordination distribution, symmetrizes the remaining framework by repeatedly loosening the tolerance of a symmetry finder, and then runs masked inpainting so that only the guest atoms are diffused inside that fixed, high-symmetry host. In the two demonstrated systems, inpainting raises the symmetry success rate far above the unconditional baseline of under five percent, and the downstream relaxation-plus-DFT pipeline identifies Li$_5$Si$_2$ in space group $R\\bar{3}m$ as a thermodynamically stable phase below the known convex hull.","pith_inferences":["If the scaffold-symmetrization step is chemically sound rather than a tolerance artifact, the same host-guided conditional scheme should transfer to substitutional disorder, defective frameworks, and interfaces where the bulk stays fixed, territories where unconditional generation is currently unreliable.","A direct test of the method's ability to place guests correctly would be to inpaint lithium into a known intercalation host, such as a lithiated chloride or sulfide framework, and compare the generated lithium sites and ordering with experimentally determined ones; the paper demonstrates the workflow but does not quantify site-level accuracy.","The symmetry success metric counts any space group above P1 or P-1; future work could tighten the metric to require specific target space groups or Wyckoff positions, which would separate genuine prototype discovery from mere tolerance-assisted ordering.","Combining CHGGen with explicit symmetry-constrained or Wyckoff-position-based generation could push the recovered symmetries from the moderate monoclinic cases reported here toward more complex framework prototypes such as NASICON-type structures."],"forward_implications":["Inpainting on a symmetrized host should allow diffusion models to generate ordered crystals well beyond the roughly 20-atom scale where unconditional generation degrades.","Host-guest generation gives a direct route to intercalation compounds and partially occupied structures: fix the known framework and let diffusion place the mobile or interstitial species.","Because conditional and unconditional generation differ only in the mask applied during reverse diffusion, any future generative model can inherit the symmetry-improving strategy without retraining.","Foundation-potential relaxation tightens generated structures into local minima, but its systematic softening means screening thresholds for DFT follow-up should be set well below the usual 0.1 eV per atom; the paper suggests roughly 30 meV per atom.","The Li-Si case shows the pipeline can propose phases that fall below the known convex hull and are later confirmed by DFT, so generative inpainting can serve as a hypothesis generator for phase-diagram completion."],"supporting_citations":[{"why":"Supplies the pretrained universal interatomic potential used to relax generated structures and to compute decomposition energies for screening.","marker":"[12]"},{"why":"Defines the CDVAE generative approach that this work builds on and whose outputs are known to lack symmetry and stability.","marker":"[22]"},{"why":"Provides the MatterGen lattice-diffusion baseline against which CHGGen's symmetry success rates are compared in both chemical systems.","marker":"[28]"},{"why":"Establishes the score-based SDE framework from which the variance-exploding diffusion and ancestral sampling are taken.","marker":"[38]"},{"why":"Introduces the RePaint resampling algorithm that underlies the mask-based inpainting loop for harmonizing guest atoms with the host.","marker":"[43]"},{"why":"Supplies the earlier score-based inpainting formulation for crystal structures and the model and test evaluation used here.","marker":"[44]"},{"why":"Documents the inability of graph neural networks to capture crystal periodicity, the failure mode CHGGen is designed to circumvent.","marker":"[47]"},{"why":"Establishes the systematic softening of universal interatomic potentials, used to explain and correct the decomposition-energy screening threshold.","marker":"[50]"},{"why":"Earlier genetic-algorithm identification of the Li5Si2 R-3m phase that independently corroborates the CHGGen prediction.","marker":"[53]"},{"why":"Earlier random-structure-search identification of thermodynamically stable Li-Si phases that corroborates the generated Li5Si2 polymorph.","marker":"[54]"}],"fun_headline_variants":["Host-guided inpainting restores crystal symmetry","Inpainting inside symmetric frameworks improves AI crystals","New stable Li5Si2 polymorph found via guided inpainting","Diffusion with host inpainting boosts symmetric crystal yield","Inpainting into symmetric hosts yields better crystals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that deleting the atom type with the broadest coordination distribution and symmetrizing what remains at strongly loosened matching tolerances produces a real crystallographic scaffold rather than relabelling a disordered configuration as a higher-symmetry space group.","fun_headline_variants_meta":{"raw":{"variants":["Host-guided inpainting restores crystal symmetry","Inpainting inside symmetric frameworks improves AI crystals","New stable Li5Si2 polymorph found via guided inpainting","Diffusion with host inpainting boosts symmetric crystal yield","Inpainting into symmetric hosts yields better crystals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00057,"raw_usage":{"total_tokens":2674,"prompt_tokens":901,"completion_tokens":1773,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":1696}},"tokens_in":517,"tokens_out":1773,"duration_ms":13930,"temperature":1.0,"reasoning_tokens":1696,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:53:03.022928+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a host framework obtained at the loosest tolerances (site tolerance 2.0, angle tolerance 30 degrees), relax it with DFT or a tight-tolerance force field, and re-determine its space group at standard tolerances; if the refined symmetry collapses to P1 or P-1, the scaffold that guides inpainting is an artifact. A complementary test is to run inpainting into a deliberately randomized, P1 version of the same host and compare symmetry success rates and decomposition energies with the symmetrized version.","supporting_citations":[{"cited_title":"Crystal structure prediction with host-guided inpainting generation and foundation potentials","cited_arxiv_id":"2504.16893","evidence_quote":"Defines the CDVAE generative approach that this work builds on and whose outputs are known to lack symmetry and stability."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MatterGen lattice-diffusion baseline against which CHGGen's symmetry success rates are compared in both chemical systems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the score-based SDE framework from which the variance-exploding diffusion and ancestral sampling are taken."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the earlier score-based inpainting formulation for crystal structures and the model and test evaluation used here."},{"cited_title":"Waroquiers, X","cited_arxiv_id":null,"evidence_quote":"Documents the inability of graph neural networks to capture crystal periodicity, the failure mode CHGGen is designed to circumvent."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the systematic softening of universal interatomic potentials, used to explain and correct the decomposition-energy screening threshold."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier genetic-algorithm identification of the Li5Si2 R-3m phase that independently corroborates the CHGGen prediction."},{"cited_title":"Artrith, A","cited_arxiv_id":null,"evidence_quote":"Earlier random-structure-search identification of thermodynamically stable Li-Si phases that corroborates the generated Li5Si2 polymorph."}],"review_version":1}