{"id":"0820ecc1-9100-4259-8009-3e8720db6a16","arxiv_id":"2607.11712","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Contrastive alignment of UMA slab and bulk embeddings retrieves true parent bulks at R@1>91% and R@3>98%, enabling an adsorption-energy-targeted bulk discovery pipeline for generative catalysis.","lead":"CatRetriever maps catalyst surface slabs and bulk crystals into one embedding space so a generated surface can retrieve likely parent bulks. That link lets generative surface models check bulk stability, symmetry, and adsorption behavior they otherwise cannot assess.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Holdout R@1/R@3 is measured only on FAIRChem-enumerated slabs from MP bulks, not on the CatGPT-generated queries that drive the discovery claim.","rationale":"The reader's weakest_assumption correctly isolates the transfer gap between frozen-UMA + InfoNCE embeddings trained on MP-enumerated pairs and the CatGPT queries used downstream. The present stress-test sharpens that gap: the published R@1/R@3 figures are distributionally matched to training and therefore cannot be read as evidence that the same ranking remains faithful for generative slabs. No independent non-learned baseline is needed to see the issue; the missing metric is simply the same recall numbers on the actual application queries. Because the modular pipeline is still a useful demonstration and the enumerated holdout itself is clean, the verdict remains CONDITIONAL rather than REJECT. A single targeted re-evaluation on CatGPT slabs would settle whether the central retrieval claim extends to the setting the paper ultimately cares about.","tokens_in":18387,"tokens_out":566,"duration_ms":4726,"concrete_test":"For a stratified sample of ~100 CatGPT query slabs that have an MP bulk of matching composition, recompute top-1/top-3 retrieval against the full 38 901-bulk pool and report R@1, R@3, and the fraction exceeding the 0.84 threshold. If R@1 falls below ~70% or the high-score fraction collapses relative to the enumerated holdout, the transfer of the abstract claim to the generative pipeline is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract and Parent-Bulk Retrieval Performance section report R@1 > 91% and R@3 > 98% on both in-distribution and holdout sets. Those sets are constructed exclusively from FAIRChem/Open Catalyst Project slabs generated by Slab.from_bulk_get_specific_millers() on Materials Project bulks (Construction of Bulk and Slab Database; Supplementary Note S1). The holdout therefore tests generalization to unseen MP bulks under the same enumeration distribution used for training, not to the conditionally generated CatGPT slabs that are the actual queries in the Application section. CatGPT slabs need not match any pre-enumerated termination or local geometry (Introduction; Supplementary Note S3). The paper never reports R@k, positive/negative score separation, or margin statistics for those 581 target-window CatGPT queries against known or DFT-validated parents. Consequently the headline retrieval numbers do not directly underwrite the claim that CatRetriever supplies physically usable parents for generative-model outputs; the 0.84 score threshold (Supplementary Note S4) is itself calibrated only on the enumerated evaluation set.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript introduces CatRetriever, a contrastive model that maps UMA-derived slab and bulk embeddings into a shared latent space so that a slab query can rank parent bulk candidates by cosine similarity. On FAIRChem-enumerated slabs from Materials Project bulks, it reports R@1 > 91% and R@3 > 98% for both in-distribution and bulk-holdout splits against a full pool of ~39k candidates. The authors then assemble a discovery pipeline that (i) generates N*-adsorbed slabs with conditional CatGPT at a target adsorption energy, (ii) retrieves or MatterGen-expands parent bulks above a score threshold of 0.84, and (iii) validates candidates by multi-site N* adsorption energy distributions and bulk thermodynamics, illustrated for ammonia decomposition.","tokens_in":18722,"tokens_out":1282,"duration_ms":9731,"significance":"Closing the slab-to-bulk gap is a genuine bottleneck for surface generative models in catalysis: without a parent bulk one cannot assess formation energy, surface energy, symmetry, or synthesizability. Framing the problem as full-pool contrastive retrieval rather than exhaustive structural matching is a practical and modular contribution, and the bulk-level holdout plus full-pool InfoNCE design is methodologically sound for the enumerated setting. The end-to-end pipeline (CatGPT → CatRetriever → MatterGen expansion → UMA adsorption screening) is a coherent demonstration that existing surface generators can be connected to bulk-verified candidates. Code is released, which supports reproducibility. If the retrieval signal is shown to transfer to generative slabs, the work would be a useful post-generation bridge for the community.","major_comments":[{"comment":"The headline R@1/R@3 numbers (abstract; Parent-Bulk Retrieval Performance; Fig. 2a) are measured exclusively on FAIRChem/Open Catalyst Project slabs produced by Slab.from_bulk_get_specific_millers() from MP bulks (Construction of Bulk and Slab Database; Supp. Note S1). The holdout therefore tests generalization to unseen MP bulks under the same enumeration distribution used for training, not to the conditionally generated CatGPT slabs that drive the discovery claim (Application section; Supp. Note S3). CatGPT slabs need not match any pre-enumerated termination. No R@k, positive/negative score separation, or margin statistics are reported for the 581 target-window CatGPT queries against known or DFT-validated parents. The central claim that CatRetriever supplies physically usable parents for generative-model outputs is therefore only indirectly supported.","section":null},{"comment":"The operational retrieval-score threshold of 0.84 (Supp. Note S4; Fig. S4) is calibrated solely on the enumerated evaluation set (precision/recall vs. true parent labels). Its transfer to CatGPT queries is assumed rather than demonstrated. Because the discovery pipeline accepts or rejects candidates on this threshold before adsorption validation, the manuscript should either (a) provide an independent calibration or proxy validation on generative slabs (e.g., StructureMatcher/DFT parent recovery on a subset, or score-vs-adsorption-overlap correlation), or (b) clearly reframe the abstract/results so that the quantitative R@k claims are restricted to the enumerated setting and the generative application is presented as a qualitative demonstration.","section":null},{"comment":"In the application, structural compatibility is equated with high cosine score, then used to justify bulk thermodynamic and multi-facet adsorption evaluation (Application of CatRetriever-Assisted Bulk Discovery Framework; Fig. 4). The paper itself notes that high score does not guarantee target adsorption environments, yet the pipeline still treats score > 0.84 as sufficient to shortlist for expensive multi-site screening. A quantitative link between retrieval score and downstream catalytic relevance (e.g., fraction of high-score vs. low-score candidates that land in the target window) would strengthen the load-bearing assumption that the learned ranking is a faithful filter for generative slabs.","section":null}],"minor_comments":[{"comment":"Figure numbering in the application section is inconsistent: the text refers to 'Figure 3a' for adsorption-energy distribution types and later to 'Figures 3c and 3d' for NbS/Ca3Cd, while the caption of Figure 4 covers those panels. Align figure labels and cross-references.","section":null},{"comment":"Supplementary Note S2: the InfoNCE formulation uses the full candidate pool as negatives, which is a strength; state the pool size M explicitly in the main text when describing training, not only in the SI.","section":null},{"comment":"The abstract states R@1 > 91% and R@3 > 98% without distinguishing enumerated vs. generative queries; a short clarifying phrase would prevent over-reading of the discovery claim.","section":null},{"comment":"Table S1/S2 and the MatterGen redundancy analysis (Supp. Note S5.2) are useful; consider moving a one-sentence summary of the low MP-overlap (mean 1.16%) into the main text so readers see that expansion is not mere regeneration.","section":null},{"comment":"Typos/notation: 'spos and sneg' density rescaling (Fig. 2b caption) could be stated more clearly; 'Ca3Cd' vs. 'Ca₃Cd' is inconsistent across text and SI.","section":null}],"recommendation":"major_revision","confidential_remarks":"The technical core (contrastive full-pool retrieval on enumerated slabs) is solid and publishable. The main risk is overclaiming transfer to generative slabs without direct metrics. If the authors either add a modest generative-slab validation or carefully scope the abstract/claims, this is a clear accept after revision; if they insist on the current abstract wording without new evidence, the gap remains load-bearing. Scope is appropriate for a methods/applications journal in ML for materials/catalysis."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful thing here is a clean, modular fix for a real gap: surface generators spit out slabs with no parent bulk, so you cannot check formation energy, hull distance, or multi-facet adsorption. CatRetriever trains small projection heads on frozen UMA scalar embeddings with full-pool InfoNCE and gets R@1 ~91.5–91.9 % and R@3 ~98.8–98.9 % on both in-distribution and bulk-holdout sets. That is honest evaluation (bulk-level split, full 38k pool, score separation and margin analysis of failures). Code is promised on GitHub. The pipeline that then ranks MP candidates, expands with MatterGen when scores are low, and screens N* adsorption distributions for ammonia decomposition is a coherent demonstration, not vaporware.\n\nWhat is actually new is the formulation plus the measured retrieval numbers and the end-to-end workflow. Contrastive dual-encoder alignment itself is not new; they cite the right prior work. The application is illustrative case studies (NbS, Ca3Cd) rather than a closed discovery loop with experimental validation.\n\nThe soft spot that matters is the one the stress-test flags. The headline R@k numbers are measured only on FAIRChem-enumerated slabs from MP bulks—the same distribution used for training. The 581 CatGPT queries that drive the discovery claim are never given the same R@k, score-separation, or margin statistics against known parents. The 0.84 threshold is calibrated on the enumerated set. So the abstract’s retrieval claim does not directly underwrite “physically usable parents for generative-model outputs.” That is a real but proportionate gap, not a collapse of the method; the authors already note chemical-space limits and the post-generation (not joint) design. Missing non-learned baselines (composition match, structure fingerprint) is a secondary omission.\n\nThis is for people building or using surface generative models who need a practical bulk bridge. Math and data look solid for what they claim; citations are appropriate. I would send it to peer review. A referee should demand the CatGPT-query retrieval diagnostics and a couple of simple baselines; those are fixable. Worth reading if you work in this niche.","headline":"Solid modular retrieval bridge with real holdout numbers on enumerated slabs; the generative-application claim is only weakly underwritten by those numbers.","tokens_in":19298,"tokens_out":534,"would_cite":true,"duration_ms":5369,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"CatRetriever recovers parent bulk crystals from catalyst surface slabs with >91% top-1 accuracy, closing the gap left by surface generative models.","keywords":["slab-to-bulk retrieval","contrastive learning","catalyst discovery","adsorption energy screening","inverse design","surface generative models","bulk crystal representation"],"falsifier":"Take a large set of CatGPT-generated slabs whose true parent bulks are known by construction, run CatRetriever against the full bulk pool, and check whether R@1 falls well below the reported 91% or whether high-scoring retrieved bulks systematically fail to reproduce the target adsorption-energy window on enumerated surfaces.","tokens_in":19297,"feed_emoji":"🧪","tokens_out":625,"duration_ms":4993,"temperature":0.7,"pith_summary":"Surface generative models for catalysts produce slab–adsorbate structures but leave the parent bulk unknown, so bulk properties such as formation energy, surface energy, and synthesizability cannot be checked. CatRetriever treats the missing link as a retrieval problem: it maps slabs and bulk crystals into one shared embedding space with contrastive learning and ranks candidate bulks by similarity. On both in-distribution and held-out Materials Project slabs the correct bulk appears first more than 91% of the time and inside the top three more than 98% of the time. The same model is then wired into a discovery pipeline that expands the bulk search space with a generative model when database hits are weak and validates candidates by whether their surfaces can realize a target adsorption energy. The result is a modular post-generation bridge that turns surface-only designs into bulk-verified catalyst candidates.","feed_headline":"Surface-to-bulk retrieval hits 91% top-1 for catalyst design","feed_subtitle":"Contrastive embeddings recover parent crystals so generative surface models can be bulk-validated","key_machinery":"Shared latent space of L2-normalized slab and bulk vectors (Z_s, Z_b) obtained by projecting UMA scalar-channel embeddings; retrieval score is their cosine similarity, optimized with full-pool InfoNCE so that matched pairs rank above all other bulks.","core_discovery":"CatRetriever, a contrastive model that aligns frozen foundation-model embeddings of slabs and bulks through small projection heads, retrieves the true parent bulk from a Materials Project-scale pool with R@1 above 91% and R@3 above 98% on both in-distribution and completely held-out bulk structures, thereby supplying the missing bulk identity for surfaces produced by catalyst generative models.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["CatRetriever matches slabs to parent bulks at 91% top-1","Contrastive slab-bulk retrieval reaches 91% R@1 accuracy","Slab queries recover true parent bulks above 91% top-1","Aligning frozen embeddings yields 91% top-1 bulk retrieval","From generative slabs to parent bulks: 91% top-1 recovery"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method assumes that embeddings learned on database-enumerated slab–bulk pairs still give a reliable ranking for conditionally generated slabs that need not match any pre-enumerated termination, so a high cosine score can be trusted as a usable parent bulk.","fun_headline_variants_meta":{"raw":{"variants":["CatRetriever matches slabs to parent bulks at 91% top-1","Contrastive slab-bulk retrieval reaches 91% R@1 accuracy","Slab queries recover true parent bulks above 91% top-1","Aligning frozen embeddings yields 91% top-1 bulk retrieval","From generative slabs to parent bulks: 91% top-1 recovery"]},"model":"grok-4.5","effort":"low","cost_usd":0.00438,"raw_usage":{"total_tokens":1288,"prompt_tokens":789,"num_sources_used":0,"completion_tokens":99,"cost_in_usd_ticks":43800000,"prompt_tokens_details":{"text_tokens":789,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":400,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":789,"tokens_out":99,"duration_ms":4144,"temperature":1.0,"reasoning_tokens":400,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T03:41:22.111171+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Take a large set of CatGPT-generated slabs whose true parent bulks are known by construction, run CatRetriever against the full bulk pool, and check whether R@1 falls well below the reported 91% or whether high-scoring retrieved bulks systematically fail to reproduce the target adsorption-energy window on enumerated surfaces.","supporting_citations":[],"review_version":1}