{"id":"374f16dd-a58d-405f-af10-eb9d9a7b8aa9","arxiv_id":"2507.19799","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"CrysVCD generates valence-balanced compositions with an elemental language model and then constructs their crystal structures with a diffusion model, reporting improved stability and functional property targeting.","lead":"This paper introduces CrysVCD, a two-stage generative model that first writes charge-balanced chemical formulas and then draws crystal structures for them with a diffusion model. It reports higher fractions of thermodynamically and phonon-stable generated crystals, plus candidate high thermal conductivity and high dielectric materials.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MatterSim is both the fine-tuning label source and the evaluation oracle for the 85%/68% stability claims; without DFT cross-checks on generated crystals, those headline rates remain unverified.","rationale":"The reader's weakest_assumption identifies exactly the vulnerability I find most load-bearing: the headline stability percentages are generated and evaluated with the same MLIP surrogate, with no independent DFT validation on the generated structures. This matters because the paper's novelty and its abstract emphasize those percentages, and the active-learning loop can exploit surrogate-specific biases rather than physical stability. The concern is not an indictment of MLIP surrogates or of the authors; it is a standard validation gap, and it is partially acknowledged in SI §3.3, which shows that surrogate error grows on biased subsets. I also note that the 'guaranteed valence-balanced by construction' claim is composition-level: Stage II receives atom types and counts, not the oxidation-state assignments, so structural valence validity is not separately guaranteed by construction. The architecture contribution remains interesting and the released code is a real asset, but the stability claim should not be taken at face value until DFT rates on generated crystals are reported. Since the reader already conditioned acceptance on this point, my stress-test does not change the verdict.","tokens_in":16103,"tokens_out":4067,"duration_ms":49596,"concrete_test":"Draw 150 generated structures from the stability-fine-tuned CrysVCD model, relax them with MatterSim, then recompute Ehull with DFT (PBE, Materials Project hull) and phonon spectra with DFT force constants (VASP + phonopy), applying the same thresholds (Ehull < 0.1 eV; no imaginary frequency > 0.01 THz). Compare DFT rates to 85%/68% and compute the MatterSim-vs-DFT confusion matrix on these generated, out-of-distribution structures, not MP-20 entries. If rates drop materially or confusion is poor, rephrase headline claims as surrogate-optimized. Supplementary: also run this on 50 structures from unconditional CrysVCD to confirm the incremental gain is not a surrogate artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claims—85% thermodynamic stability and 68% phonon stability—are produced by a closed loop. In 'MLIP and DFT calculations', Ehull is evaluated with MatterSim-v1.0.0-1M and phonon stability with MatterSim-v1.0.0-5M; in 'Stability guided generation and fine-tuning', those same surrogate evaluations provide the positive/negative labels used to fine-tune the conditional model. Evaluation and label generation therefore share the same error model. The SI's only validation of MatterSim Ehull is on MP-20 entries (Fig. S10), which are in-distribution and likely part of MatterSim's training data; it contains no DFT benchmark for phonon stability on generated structures. The paper itself shows in SI §3.3 that on a biased high-kph subset, surrogate error rises (κRMSE 0.86 vs 0.58 reported for MatterSim's general benchmark), confirming that out-of-distribution error can be substantial. Consequently, the 85%/68% figures may measure the model's ability to satisfy MatterSim's preferences rather than DFT stability. This is not a dispute about using MLIPs in general; it is a missing independent check on the specific claim that generated crystals are stable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CrysVCD, a two-stage generative model for crystal structures. Stage I uses a transformer-based elemental language model to autoregressively generate valence-balanced chemical compositions, with token embeddings initialized from electronic configurations; Stage II uses a DiffCSP-style diffusion model to generate crystal structures conditioned on the composition. The authors show that valence-constrained generation lowers the energy-above-hull distribution relative to plain DiffCSP, and that a concurrent learning loop, in which unconditional generations are labeled by stability metrics from the MatterSim machine-learning interatomic potential and then used to fine-tune a conditional model, raises the fraction of generated crystals classified as thermodynamically stable (roughly 85% with Ehull < 0.1 eV) and as phonon-stable (68%). They also demonstrate conditional generation of high-thermal-conductivity materials (e.g., hexagonal GeC with DF T-verified kph = 183 W/m/K) and high-dielectric-constant materials (e.g., CoSnO2 with εDFT,zz = 70.5). The framework is proposed as a modular plug-in that can be integrated into existing generative pipelines.","tokens_in":16450,"tokens_out":4119,"duration_ms":46683,"significance":"If the stability statistics are robust under independent evaluation, this is a useful contribution: the two-stage design explicitly enforces charge neutrality before structure generation, the embeddings encode periodic chemical trends, and the closed-loop fine-tuning scheme is a practical way to steer generation toward stability. The authors also commit to releasing code and data, and they provide DFT-level verification for specific functional candidates (GeC, CoSnO2). The principal uncertainty is that the headline 85% and 68% stability rates are measured with MatterSim, the same surrogate that provides the fine-tuning labels, and the provided surrogate validation (SI Fig. S10) is on in-distribution MP-20 compounds rather than on the generated crystals. The paper itself shows in SI §3.3 that surrogate error grows on a biased high-κ subset (κRMSE 0.86 vs 0.58), so an independent DFT cross-check of the generated crystals is needed to support the central quantitative claims.","major_comments":[{"comment":"The reported 85% thermodynamic and 68% phonon stability rates are computed with MatterSim (MatterSim-v1.0.0-1M for Ehull and MatterSim-v1.0.0-5M for phonons), which is the same MLIP that provides the positive/negative labels used to fine-tune the conditional model. This closed loop means the rates partly measure how well the model satisfies the surrogate's preferences, not necessarily how stable the crystals are at the DFT level. The SI validation (Fig. S10) only benchmarks MatterSim on MP-20 entries that are already DFT-stable and likely part of MatterSim's training distribution; it does not validate on generated hypothetical crystals. To make the headline claim load-bearing, the authors should re-evaluate a random sample (e.g., 100–200) of the generated crystals with DFT (PBE) for both Ehull and phonon stability, and report the resulting stability percentages together with a MatterSim-vs-DFT comparison for those structures.","section":"Stability guided generation and fine-tuning; MLIP and DFT calculations"},{"comment":"The abstract states that fine-tuned CrysVCD achieves \"85% thermodynamic stability,\" whereas the Results text states that the model \"achieves > 85% stability rate\" and does not give an exact percentage or the number of crystals evaluated. Please reconcile these numbers, specify the denominator and the exact Ehull threshold used (the text mentions 0.1 eV, while Fig. 3b refers to thresholds of 0 and -0.05 eV), and state the number of generated crystals over which the 85% and 68% rates are computed.","section":"Abstract and Results (Stability guided generation and fine-tuning)"},{"comment":"The high-κ dielectric generation uses the same GNNOPT surrogate both to fine-tune the conditional model and to screen/evaluate the generated candidates, with DFT verification reported for only one material (CoSnO2). For the claim that the ε distribution shifts toward values above 50, the authors should report how many generated candidates were DFT-verified, or at least list the DFT-computed ε for a handful of top-ranked candidates, so that the reader can gauge the surrogate bias on out-of-distribution generated structures.","section":"Generating high-performance functional materials with CrysVCD (Fig. 4d-f)"}],"minor_comments":[{"comment":"The abstract contains apparent typos: \"Y et\" and \"V ery\" should read \"Yet\" and \"Very,\" respectively.","section":"Abstract"},{"comment":"The text contains the typo \"CrysCVD\" (first paragraph of the benchmarking discussion) where the model name should be \"CrysVCD.\"","section":"Results, Stability guided generation and fine-tuning"},{"comment":"The caption mentions fine-tuning with Ehull thresholds of 0 and -0.05 eV, but these threshold values are not defined or explained in the main text; please clarify what these thresholds mean for the fine-tuning procedure.","section":"Fig. 3(b) caption"},{"comment":"The comproportionation constraint states that no element may simultaneously exhibit both positive and negative oxidation states; consider clarifying that this applies to different atoms of the same element within one structure, which appears to be the intended meaning.","section":"Methods, Chemical formula generation"},{"comment":"The statement that \"all generated crystals are guaranteed to be valence-balanced by construction\" should specify that the charge-balance filter is applied to ionic compounds (as described in Methods), while alloys are balanced by construction through zero-valence tokens; this nuance is present but could be stated more explicitly where the claim is made.","section":"Results, Model architecture and Methods"}],"recommendation":"major_revision","confidential_remarks":"The main technical concern is the circularity between MatterSim as the fine-tuning label source and as the evaluation oracle for the headline stability statistics. Adding a DFT-based validation set of generated crystals is the most direct way to resolve this, and it should be feasible within the scope of the paper. I also suggest that the authors check whether the MP-20 training split overlaps with MatterSim's training data; if it does, the SI's in-distribution validation (Fig. S10) is even weaker than it appears."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I've read the CrysVCD paper. Here's my take: the two-stage architecture is the real contribution, and it's a good one. Autoregressive generation of valence-balanced formulas with electronic-configuration embeddings, then a diffusion CSP step, is a clean, modular idea that could be plugged into any structure generator. The authors show it works with DiffCSP, and the valence decomposition handles mixed-valence compounds (Fe3O4). The efficiency gain in screening (orders of magnitude fewer valence checks) is plausible and well demonstrated. I also give credit for their functional demonstrations: the GeC thermal conductivity and CoSnO2 dielectric candidates are checked against DFT, not just the surrogate.\n\nThe soft spot is the stability headline. The 85% thermodynamic and 68% phonon stability rates are computed with MatterSim, which is the same model used to generate the fine-tuning labels. That's a closed loop. The SI's validation of MatterSim on MP-20 is in-distribution and says nothing about phonons on generated structures; the paper itself shows surrogate error grows on a biased subset (κRMSE 0.86 vs 0.58). So the absolute numbers likely overstate DFT-level stability. The relative improvement (49→68% phonon stable after fine-tuning) is still meaningful as surrogate optimization, but the abstract presents these as thermodynamic and phonon stability without caveat. That's a load-bearing issue for the headline claims, but it's fixable: give DFT Ehull and phonon checks on a random sample of generated crystals, or reword the claims as surrogate-evaluated rates.\n\nI don't think the paper is unserious. The authors are honest about other limitations (excluded elements, DFT-derived dielectric labels). The architecture is the contribution; the stability numbers are an overreach. With a revision that either validates or softens those numbers, it would be a solid addition to the generative materials literature.\n\nI'd send it to peer review and lean conditional acceptance with major revision.","headline":"A smart modular two-stage generative model with a real evaluation weakness: stability claims rest on a closed-loop surrogate.","tokens_in":16840,"tokens_out":3981,"would_cite":true,"duration_ms":43267,"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":"CrysVCD claims that generating a charge-balanced chemical formula before generating the crystal structure produces chemically valid crystals, and that fine-tuning on machine-learned stability scores raises thermodynamically stable output…","keywords":["valence balance","composition-first generation","crystal diffusion model","thermodynamic stability","phonon stability","conditional generation","thermal conductivity","dielectric constant"],"falsifier":"Take a few hundred CrysVCD-generated structures that the MLIP labels stable and recompute their energy above hull and phonon imaginary frequencies with density functional theory; if the stable fraction falls well below 85% thermodynamic or 68% phonon stability, the headline stability numbers are artifacts of the surrogate evaluator.","tokens_in":1788,"feed_emoji":"🔬","tokens_out":1889,"duration_ms":106964,"temperature":0.7,"pith_summary":"This paper proposes CrysVCD, a generative model for crystals that refuses to let the network guess charge balance. Instead, a transformer first writes a chemical formula as a sequence of element-oxidation-state tokens whose charges sum to zero, and a diffusion model then builds the atomic structure for that formula. The claim is that putting the valence constraint before structure generation removes a large class of chemically invalid outputs and makes stability screening far cheaper, since checking a formula's total charge is nearly instantaneous compared with denoising a structure. When the model is fine-tuned on stability scores from a machine-learned interatomic potential, the paper reports 85% of generated crystals are thermodynamically stable and 68% are phonon-stable, and it demonstrates targeted generation of a high-thermal-conductivity semiconductor and a high-dielectric compound. A sympathetic reader would take the central insight to be that chemical rules can be made part of the generative grammar rather than a post-hoc filter.","feed_headline":"Valence-first design yields 85% stable crystals","feed_subtitle":"A two-stage model writes charge-balanced formulas before building structures, lifting phonon stability to 68%.","key_machinery":"The load-bearing object is the valence-constrained elemental language model in Stage I. It tokenizes a formula as valence-labeled element-count pairs $(a^v, c)$, embeds each token by the electronic shell occupancies of that ion following the Aufbau principle, and trains an autoregressive transformer to predict the next token as a joint logit over element-valence and count. A filter enforces charge neutrality $\\sum c_i v_i = 0$ before any structure is generated, so every final crystal is valence-balanced by construction. Stage II is an E(3)-equivariant diffusion model performing crystal structure prediction conditioned on the formula, with property embeddings added at both stages and classifier-free guidance at weight $w=2$ for conditional generation.","core_discovery":"The central discovery is that decomposing crystal generation into composition first, structure second, with an explicit charge-neutrality constraint, produces chemically valid crystals and improves downstream stability. The composition generator treats each atom in a specific oxidation state, e.g. Fe$^{2+}$ or O$^{2-}$, as a token and autoregressively builds the formula under the constraint that the sum of charge times atom count equals zero, $\\sum c_i v_i = 0$; mixed-valence cases such as Fe$_3$O$_4$ are handled by allowing multiple valence tokens for one element. The structure generator then performs crystal structure prediction conditioned on that formula, so the only structures ever denoised are ones whose chemistry is already valid. Fine-tuning the pair on energy-above-hull and phonon-stability labels from a machine-learned interatomic potential shifts the generated distribution toward stable materials, raising the phonon-stable fraction from 49% to 68% and reaching over 85% thermodynamic stability; the same conditioning machinery steers generation toward high thermal conductivity (a GeC candidate with 183 W $m^{{-1}}$ $K^{{-1}}$) and high dielectric constant (CoSnO$_2$, with surrogate-predicted $\\varepsilon \\approx 62$ and a DFT-verified component of 70.5).","pith_inferences":["If the machine-learned potential is biased toward its own training distribution, the 85% and 68% stability rates would likely drop under full DFT re-evaluation; the paper's own DFT checks cover only a few selected functional candidates, not a statistical sample of the stability claims.","The composition-first idea transfers naturally to other hard chemical constraints, such as electronegativity ordering, magnetic valence, or site-occupancy rules, where the same tokenization-plus-filter trick could prune invalid candidates before expensive structure generation.","The concurrent learning loop of generate, evaluate with a fast surrogate, and fine-tune on the resulting labels is a general recipe for scarce property data; its ceiling for dielectric design is set by surrogate accuracy, so the high-kappa candidates should be read as surrogate-guided hypotheses until experimental or higher-level validation.","A direct test of the plugin claim would be to swap the Stage II diffusion model for a different structure predictor and measure whether the valence-constrained compositions still lower the energy-above-hull distribution relative to unconstrained generation; the paper shows this effect for one backbone only."],"forward_implications":["Every crystal CrysVCD outputs is charge-balanced by construction, so valence screening no longer needs a post-hoc search over oxidation-state assignments; the paper reports the composition step checks total valence orders of magnitude faster than the diffusion step.","Fine-tuning on stability labels from a machine-learned interatomic potential shifts the generated distribution toward lower energy above hull and fewer imaginary phonon modes, raising phonon-stable output from 49% to 68%.","The same property-conditioning mechanism can target functional targets beyond stability: high thermal conductivity and high dielectric constant, yielding candidates (hexagonal GeC, CoSnO$_2$) that are absent from the training data.","Because Stage II only needs a formula-conditioned structure generator, the valence-constrained composition model can be bolted onto other generative backbones rather than only the diffusion model demonstrated here."],"supporting_citations":[{"why":"Supplies the diffusion-based crystal structure prediction architecture used as Stage II and as the baseline for stability comparison.","marker":"[8]"},{"why":"Supplies the machine-learned interatomic potential used to evaluate energy above hull and phonon stability, which also provides the fine-tuning labels.","marker":"[9]"},{"why":"Defines classifier-free guidance, the conditioning mechanism used to steer both stages toward target properties.","marker":"[13]"},{"why":"Inspires the electronic-configuration embedding that initializes element-valence tokens in the composition transformer.","marker":"[19]"},{"why":"Provides the equivariant graph neural network surrogate used to predict dielectric constants and guide high-kappa generation.","marker":"[27]"},{"why":"Provides the training compositions and the convex-hull reference data used for stability evaluation.","marker":"[30]"},{"why":"Provides the autoregressive transformer architecture on which the elemental language model is built.","marker":"[31]"},{"why":"Provides the anharmonic phonon workflow used to compute thermal conductivity for functional candidate validation.","marker":"[37]"}],"fun_headline_variants":["Valence constraint boosts crystal stability to 85%","CrysVCD: charge-balanced generation yields stable crystals","From formulas to structures: valence-checked diffusion hits 85%","Chemical rules in diffusion: 85% stability, 68% phonon-safe"],"cache_read_input_tokens":19072,"weakest_assumption_plain":"The 85% and 68% stability rates assume the machine-learned interatomic potential's energy-above-hull and phonon predictions are accurate enough to serve both as fine-tuning labels and as the evaluation metric; only a handful of functional candidates are checked with density functional theory.","fun_headline_variants_meta":{"raw":{"variants":["Valence constraint boosts crystal stability to 85%","CrysVCD: charge-balanced generation yields stable crystals","From formulas to structures: valence-checked diffusion hits 85%","Chemical rules in diffusion: 85% stability, 68% phonon-safe"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000745,"raw_usage":{"total_tokens":3344,"prompt_tokens":993,"completion_tokens":2351,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":2278}},"tokens_in":609,"tokens_out":2351,"duration_ms":18677,"temperature":1.0,"reasoning_tokens":2278,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:59:46.195851+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a few hundred CrysVCD-generated structures that the MLIP labels stable and recompute their energy above hull and phonon imaginary frequencies with density functional theory; if the stable fraction falls well below 85% thermodynamic or 68% phonon stability, the headline stability numbers are artifacts of the surrogate evaluator.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion-based crystal structure prediction architecture used as Stage II and as the baseline for stability comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the machine-learned interatomic potential used to evaluate energy above hull and phonon stability, which also provides the fine-tuning labels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Inspires the electronic-configuration embedding that initializes element-valence tokens in the composition transformer."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the equivariant graph neural network surrogate used to predict dielectric constants and guide high-kappa generation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the training compositions and the convex-hull reference data used for stability evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the autoregressive transformer architecture on which the elemental language model is built."}],"review_version":1}