Pith. sign in

REVIEW 3 major objections 5 minor 14 references

CatRetriever: Contrastive Representation Learning for Slab-to-Bulk Retrieval in Generative Catalyst Discovery

T0 review · 3 major / 5 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read CatRetriever recovers parent bulk crystals from catalyst surface slabs with >91% top-1 accuracy, closing the gap left by surface generative models.

desk verdict Solid modular retrieval bridge with real holdout numbers on enumerated slabs; the generative-application claim is only weakly underwritten by those numbers. read the letter →

arxiv 2607.11712 v1 pith:ND2WHMG6 submitted 2026-07-13 cs.LG cond-mat.mtrl-sci

Jungho Oh , Woosung Kim , Dong Hyeon Mok , Jonggeol Na , Seoin Back This is my paper · ORCID
classification cs.LGcond-mat.mtrl-sci
keywords slab-to-bulkretrievalcontrastivelearningcatalystdiscoveryadsorptionenergyscreeninginversedesignsurfacegenerativemodelsbulkcrystalrepresentation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

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.

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 (3)
  1. 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.
  2. 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.
  3. 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.
minor comments (5)
  1. 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.
  2. 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.
  3. 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.
  4. 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.
  5. Typos/notation: 'spos and sneg' density rescaling (Fig. 2b caption) could be stated more clearly; 'Ca3Cd' vs. 'Ca₃Cd' is inconsistent across text and SI.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: R@1/R@3 are post-hoc metrics on held-out MP-enumerated slabs; InfoNCE training does not force the reported recalls by construction.

full rationale

CatRetriever trains only the projection heads via single-positive InfoNCE on UMA (l=0) embeddings of FAIRChem-enumerated slab–bulk pairs drawn from an 80 % bulk split; the 20 % bulk-holdout slabs (and their parents) are excluded from training and validation. R@1 and R@3 are then computed by ranking the full 38 901-bulk pool after training, so the headline numbers (in-distribution 91.9 %/98.8 %, holdout 91.5 %/98.9 %) are genuine out-of-sample retrieval metrics, not quantities baked into the loss or fitted parameters. The 0.84 operating threshold is likewise an ordinary precision–recall choice on the same enumerated evaluation set and is not re-used as a “prediction.” Self-citations to the authors’ CatGPT generator appear only in the modular application pipeline; they do not underwrite the retrieval numbers themselves, which rest on independent Materials Project data. No self-definitional identity, fitted-input-as-prediction, uniqueness theorem, or ansatz smuggling is present in the derivation chain.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claim rests on standard contrastive learning plus domain assumptions that MP-stable binary/ternary bulks and FAIRChem-enumerated low-index slabs define a meaningful parent relation, and that frozen UMA embeddings are informative enough that a small projection head can rank parents. Free parameters are ordinary training and operating-point choices (temperature, threshold, architecture). No new physical particles or forces are postulated; CatRetriever is a method, not a physical entity.

free parameters (4)
  • InfoNCE temperature τ = 0.07
    Fixed at 0.07 for contrastive training; scales the softmax over the full bulk pool and affects ranking sharpness.
  • Top-1 retrieval-score threshold = 0.84
    Operational cutoff (0.84) chosen so that ~90% of correct top-1 retrievals are retained; gates which candidates enter adsorption validation.
  • Projection-head architecture and training hyperparameters = hidden 256; lr 0.001; batch 64; best epoch 22
    Two-layer MLP 128→256→128, AdamW lr=0.001, batch size 64, early stopping on validation R@1; chosen by validation rather than derived.
  • Target N* adsorption energy window = −0.90 eV target; −1.00 to −0.80 eV window
    Application target −0.90 eV with acceptance window −1.00 to −0.80 eV taken from ammonia-decomposition literature and used to filter CatGPT outputs.
assumptions (5)
  • standard math InfoNCE / contrastive dual-encoder training aligns matched pairs in a shared embedding space usable for retrieval (standard ML).
    Invoked throughout Learning a Joint Slab-Bulk Representation and Supplementary Note S2; loss formula is the standard single-positive InfoNCE over the full bulk pool.
  • domain assumption Binary/ternary MP bulks with E_form < 0 eV/atom and E_hull < 0.1 eV/atom, plus FAIRChem low-index slab enumeration, define the ground-truth parent relation for training and evaluation.
    Construction of Bulk and Slab Database; mirrors OC20-style criteria. Wrong or incomplete parent labels would inflate or deflate R@k.
  • domain assumption Frozen UMA-s-1p1 node features, restricted to the l=0 scalar channel and mean-pooled, are sufficiently informative structure descriptors for slab–bulk correspondence.
    Supplementary Note S2.1; backbone never fine-tuned. If UMA misses termination- or composition-critical cues, retrieval is limited by the backbone.
  • ad hoc to paper High cosine retrieval score implies structural compatibility adequate to justify bulk-level thermodynamic and multi-facet adsorption evaluation of generative slabs.
    Application section and Supplementary Note S4 treat score ≥ 0.84 as ‘structurally plausible’; catalytic relevance is then checked separately via UMA adsorption distributions.
  • domain assumption UMA-evaluated adsorption energies are a faithful enough surrogate for DFT adsorption energies for screening purposes.
    Used for CatGPT filtering and candidate validation (Supplementary Notes S3, S6, S9); pipeline conclusions inherit UMA task error.
invented entities (1)
  • CatRetriever shared slab–bulk latent space (projection heads + retrieval score r(s,b)=Z_s^T Z_b) independent evidence
    purpose: Provides a fast similarity ranking from slab queries to bulk candidates without exhaustive structural comparison.
    Methodological construct trained on MP pairs; independent evidence is the holdout R@k and score-separation plots, not an external physical observable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CatRetriever: Contrastive Representation Learning for Slab-to-Bulk Retrieval in Generative Catalyst Discovery." pith.science (2026). https://pith.science/paper/ND2WHMG6

@misc{pith2026260711712,
  author       = {Pith},
  title        = {Pith review of: CatRetriever: Contrastive Representation Learning for Slab-to-Bulk Retrieval in Generative Catalyst Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ND2WHMG6}},
  note         = {Machine review of arXiv:2607.11712}
}
read the original abstract

Inverse design is an emerging data-driven paradigm for efficiently navigating vast chemical spaces to discover new materials with targeted properties, and in the context of heterogeneous catalysis, surface generative models have recently advanced this goal by directly generating catalyst surface-adsorbate structures. However, these models typically operate at the slab level and do not provide the corresponding parent bulk structure, making it difficult to assess bulk-dependent properties such as formation energy, surface energy, crystallographic symmetry, and synthesizability. Here, we address this missing slab-to-bulk connection as a retrieval problem and introduce CatRetriever, a contrastive representation learning model that aligns slab and bulk crystal representations in a shared latent space. From a slab query, CatRetriever accurately retrieves plausible parent bulk candidates with R@1 > 91% and R@3 > 98% on both the in-distribution and holdout evaluation sets. We further extend the CatRetriever framework into an adsorption energy targeted bulk discovery pipeline that combines bulk retrieval, generative search space expansion, and adsorption energy distribution analysis. This workflow evaluates candidates by both structural compatibility with the query slab and their ability to access the target adsorption energy range across diverse surface environments. CatRetriever therefore provides a scalable route for connecting catalyst generative models with physically plausible and adsorption energy compatible bulk catalyst discovery.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 3 canonical work pages

  1. [1]

    P.; Hautier, G.; Chen, W.; Richards, W

    Jain, A.; Ong, S. P.; Hautier, G.; Chen, W.; Richards, W. D.; Dacek, S.; Cholia, S.; Gunter, D.; Skinner, D.; Ceder, G.; Persson, K. A. Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. APL Mater. 2013, 1, 011002. DOI: 10.1063/1.4812323

    doi:10.1063/1.4812323 2013

  2. [2]

    P.; Richards, W

    Ong, S. P.; Richards, W. D.; Jain, A.; Hautier, G.; Kocher, M.; Cholia, S.; Gunter, D.; Chevrier, V. L.; Persson, K. A.; Ceder, G. Python Materials Genomics (pymatgen): A robust, open-source Python library for materials analysis. Comput. Mater. Sci. 2013, 68, 314–319. DOI: 10.1016/j.commatsci.2012.10.028

    doi:10.1016/j.commatsci.2012.10.028 2013

  3. [3]

    H.; Mortensen, J

    Larsen, A. H.; Mortensen, J. J.; Blomqvist, J.; Castelli, I. E.; Christensen, R.; Dułak, M.; Friis, J.; Groves, M. N.; Hammer, B.; Hargus, C.; et al. The Atomic Simulation Environment—a Python library for working with atoms. J. Phys.: Condens. Matter 2017, 29, 273002. DOI: 10.1088/1361-648X/aa680e

    doi:10.1088/1361-648x/aa680e 2017

  4. [4]

    Open Catalyst 2020 (OC20) dataset and community challenges

    Chanussot, L.; Das, A.; Goyal, S.; Lavril, T.; Shuaibi, M.; Riviere, M.; Tran, K.; Heras-Domingo, J.; Ho, C.; Hu, W.; et al. Open Catalyst 2020 (OC20) dataset and community challenges. ACS Catal. 2021, 11, 6059–6072. DOI: 10.1021/acscatal.0c04525

    doi:10.1021/acscatal.0c04525 2020

  5. [5]

    M.; Dzamba, M.; Fu, X.; Gao, M.; Shuaibi, M.; Barroso-Luque, L.; Abdelmaqsoud, K.; Gharakhanyan, V .; Kitchin, J

    Wood, B. M.; Dzamba, M.; Fu, X.; Gao, M.; Shuaibi, M.; Barroso-Luque, L.; Abdelmaqsoud, K.; Gharakhanyan, V .; Kitchin, J. R.; Levine, D. S.; et al. UMA: A family of universal models for atoms. arXiv 2025, arXiv:2506.23971. DOI: 10.48550/arXiv.2506.23971

    doi:10.48550/arxiv.2506.23971 2025

  6. [6]

    Representation learning with contrastive predictive coding

    van den Oord, A.; Li, Y .; Vinyals, O. Representation learning with contrastive predictive coding. arXiv 2018, arXiv:1807.03748. DOI: 10.48550/arXiv.1807.03748

    arXiv doi:10.48550/arxiv.1807.03748 2018

  7. [7]

    Decoupled weight decay regularization

    Loshchilov, I.; Hutter, F. Decoupled weight decay regularization. In International Conference on Learning Representations; 2019

    2019

  8. [8]

    Toward Controllable Catalyst Inverse Design via Large-Scale Autoregressive Pretraining

    Mok, D. H.; Na, J.; Back, S. Toward controllable catalyst inverse design via large- scale autoregressive pretraining. arXiv 2026, arXiv:2606.17445. DOI: 10.48550/arXiv.2606.17445

    work page Pith review arXiv doi:10.48550/arxiv.2606.17445 2026

Show all 14 references
  1. [9]

    Diffusion model-guided inverse design of bimetallic catalysts for ammonia decomposition

    Yang, J.; Ye, K.; Xie, S.; Li, Q.; Milhans, C.; Liu, F.; Che, F. Diffusion model-guided inverse design of bimetallic catalysts for ammonia decomposition. J. Am. Chem. Soc. 2026, 148, 537–546. DOI: 10.1021/jacs.5c14652

    2026 doi

  2. [10]

    A generative model for inorganic materials design

    Zeni, C.; Pinsler, R.; Zügner, D.; Fowler, A.; Horton, M.; Fu, X.; Wang, Z.; Shysheya, A.; Crabbé, J.; Ueda, S.; et al. A generative model for inorganic materials design. Nature 2025, 639, 624–632. DOI: 10.1038/s41586-025-08628-5

    2025 doi

  3. [11]

    Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set

    Kresse, G.; Furthmüller, J. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev. B 1996, 54, 11169–11186. DOI: 10.1103/PhysRevB.54.11169

    1996 doi

  4. [12]

    P.; Burke, K.; Ernzerhof, M

    Perdew, J. P.; Burke, K.; Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett. 1996, 77, 3865–3868. DOI: 10.1103/PhysRevLett.77.3865

    1996 doi

  5. [13]

    From ultrasoft pseudopotentials to the projector augmented- wave method

    Kresse, G.; Joubert, D. From ultrasoft pseudopotentials to the projector augmented- wave method. Phys. Rev. B 1999, 59, 1758–1775. DOI: 10.1103/PhysRevB.59.1758

    1999 doi

  6. [14]

    Graph theory approach to determine configurations of multidentate and high coverage adsorbates for heterogeneous catalysis

    Deshpande, S.; Maxson, T.; Greeley, J. Graph theory approach to determine configurations of multidentate and high coverage adsorbates for heterogeneous catalysis. npj Comput. Mater. 2020, 6, 79. DOI: 10.1038/s41524-020-0345-2

    2020 doi

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.