acceptodds
Under review as a conference paper at ICLR 2027

Diffusion Sampling of Adsorbate Configurations on Catalyst Surfaces

Abstract

Reliable adsorption-energy estimation in heterogeneous catalysis requires sampling diverse low-energy adsorbate configurations on a slab, but existing data-driven methods rely on dense per-system placement labels that are computationally expensive to obtain and typically constrain the search to global rigid-body translations and rotations. We instead cast adsorbate placement as conditional Boltzmann sampling under an energy induced by a pretrained ML interatomic potential, which removes the need for dense supervision and exposes the full adsorbate coordinate space, including internal conformer degrees of freedom. We introduce AdsorbSample, an energy-trained conditional diffusion sampler whose source, controller, and differentiable restraint potentials are tailored to adsorbate–surface geometry, so that physically implausible configurations are suppressed in the proposal. On the OC20-Dense benchmarks, AdsorbSample achieves the strongest MLIP Pass@\(k\) across evaluated budgets and substantially improves DFT-verified success at low sampling budgets, while remaining competitive on raw structural validity and diversity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
adsorbate, catalyst, diffusion sampling, neural sampler, ML interatomic potentials

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_hbwvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AdsorbSample treats adsorbate placement as conditional sampling from a restrained Boltzmann target defined by a pretrained ML interatomic potential. A surface-conditioned source and typed equivariant controller generate full adsorbate Cartesian coordinates, allowing internal conformational changes. OC20 IS2RE pretraining is followed by energy/force-based Flow Sampling without dense OC20-Dense placement targets. Candidates undergo bond-restrained UMA relaxation, validity filtering, and unrestrained-energy ranking. On 44 ID and 50 OOD systems, the method improves MLIP Pass@k and generally improves low-budget DFT single-point success over AdsorbML/AdsorbDiff. Additional analyses study source/pretraining choices, controller architecture, joint surface sampling, and single-system motif replay and flexible-adsorbate examples.

Critical strengths
  1. Learning from an available MLIP energy landscape addresses a meaningful supervision bottleneck in catalyst configuration search. Full-coordinate generation, explicit bond topology, surface conditioning, and separate interaction branches form a well-motivated application of energy-trained sampling. The distinction between sparse geometric pretraining and the absence of dense placement supervision during energy training is stated clearly. This is a useful contribution even without introducing a new general-purpose sampling objective.
  2. The evaluation separates structural validity, configurational diversity, MLIP discovery, and DFT single-point verification. Applying bond-restrained relaxation to all main comparators is a valuable fairness control, and Appendix C.3 also reports unconstrained relaxation. Low-budget improvements are substantial: DFT SR@1 rises from 13.6% to 25.0% on ID and from 18.0% to 28.0% on OOD relative to AdsorbML. The architecture and modeling ablations, including the unsuccessful joint-surface variant, help identify which design choices matter. These results support weak acceptance as a promising scientific-ML method with useful empirical evidence.
Critical weaknesses
  1. The uncertainty and physical scope of the DFT evidence need stronger qualification. Results use only 44 ID and 50 OOD systems; DFT SR@k selects from the first k samples in seed/frame order, without reported uncertainty over sample ordering or model training. Several differences represent only a few systems, and DFT superiority is not uniform: at k=10 the ID result is 56.8% versus AdsorbML's 63.6%, with an OOD tie at 52.0%. Moreover, Appendix C.3 retains bond restraints without a final unconstrained relaxation, and C.4 performs DFT single points without ionic relaxation. This supports candidate-energy discovery, but does not yet establish the same ranking for fully relaxed DFT adsorption minima.
  2. The strongest evidence for flexible-adsorbate search and mode-collapse recovery comes from separately trained single-system models. The GGE experiment additionally changes temperature, adds system-specific geometric restraints, and performs endpoint refinement before replay (Appendix D.4). It demonstrates feasibility, but does not isolate the benefit of full-coordinate sampling or establish amortized generalization to larger flexible adsorbates. This distinction matters because internal conformational search is a central motivation; the current broad claims should remain tied to the tested settings.
Questions for authors
  1. How were the 44 ID and 50 OOD systems selected from OC20-Dense? Can paired system-level uncertainty and the sensitivity of DFT SR@k to candidate ordering be reported using existing evaluations?
  2. For a representative subset, do low-energy discoveries and method rankings persist after releasing the bond restraint and performing DFT ionic relaxation? If such results are unavailable, please delimit the adsorption-minimum claim accordingly and report residual physical forces where available.
  3. In the GGE example, can the rigid-placement baseline receive the same applicable restraints and endpoint-refinement budget, and how much training/MLIP-query compute does the single-system sampler require? This would clarify the incremental conformational-search benefit under a comparable resource budget.
3 comments

We thank the reviewer for their response. References are to the revised PDF.


W1 and Q1. Uncertainty.

The 44 ID and 50 OOD systems are AdsorbDiff’s evaluation subsets of the Open Catalyst Challenge 2023 subset of OC20-Dense, obtained by an 80/20 split of the ID systems and a random sample of 50 of the 200 OOD systems, and we made no further selection (§4). DFT SR@k is now the mean ± std over four candidate orderings (Table 2), and paired bootstrap 95% intervals over systems quantify the uncertainty from the small evaluation set (App. D.2, L1174–1194; Table 10). Pooled over both splits, the DFT gain over AdsorbML is +9.3 [+3.7, +14.9] points at k=1 and +7.4 [+0.5, +14.4] at k=2, and the two methods are comparable at k=5 and 10 (§5.1). For example, the ID SR@10 gap the reviewer noted (56.8% against 63.6%) came from a single ordering, and over four orderings it becomes 61.4 ± 5.2 against 59.7 ± 3.4.


W1 and Q2. Restraint release and DFT relaxation.

The ranking of methods persists in both tests, and DFT relaxation confirms the below-reference discoveries (App. D.1, L1092–1162; Tables 6–9). After releasing the restraint on all candidates, about 95% of structures of every method are UMA minima (Table 6), AdsorbSample keeps the highest Pass@k at every budget (Table 7), and its DFT SR@k changes by less than 3 points (Table 8). Under DFT ionic relaxation, every SR@1 single-point success persists for all three methods tested (L1126–1131), so single-point SR@1 is, if anything, conservative. Of the 22 basins that AdsorbSample keeps at least 0.1 eV below the reference after release, DFT relaxation confirms 20, against 6 of 8 for AdsorbML (Table 9). Residual forces are reported in Table 6.


W2 and Q3. Flexible-adsorbate evidence.

We agree, and now present these as single-system case studies (§5.3, L451–513), with the extension to amortized models listed as future work (conclusion, L527–529). As a control, we ran constrained minima hopping (CMH) with UMA and the same GGE restraints under two budgets (§5.3; App. E.4, Table 16). Both flexible methods recover the two-ring geometry and rigid placement does not. The sampler trains in 11 A100 GPU-hours with about 60k UMA evaluations. At matched GPU time, where CMH uses about six times as many evaluations, CMH returns 18 passing minima in 7 basins and AdsorbSample returns 5 in 5. At matched UMA evaluations (2 GPU-hours of CMH), CMH returns 2 passing minima in 1 basin. AdsorbSample therefore finds more basins per energy evaluation, and CMH finds more per GPU-hour once training cost is counted.

The case study thus shows that an energy-trained diffusion sampler scales to a 43-atom adsorbate with seven rotatable bonds and reaches the same low-energy mode as a dedicated global optimizer, with fewer energy evaluations. Unlike CMH, whose cost is paid in full for each new system, the sampler's cost is mostly training, which a conditional model could in principle amortize across systems. We have revised §5.3 to state this comparison explicitly (L507–513).

Thank you for the clarifications and additional experiments. The analysis across candidate orderings and the new DFT relaxations address my main concerns about the robustness of the results. The minima-hopping comparison is helpful, though the evidence for flexible adsorbates still comes from a single system. I will maintain my rating.

Thank you for reading the revision and for the follow-up. We are glad the new analyses address your main concerns and we appreciate your time and constructive suggestions.

Reviewer_FKiuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper reframes adsorbate placement on catalyst slabs as conditional sampling from a Boltzmann-like target induced by UMA and geometric restraints, rather than supervised learning from densely enumerated placements. AdsorbSample instantiates this with Flow Sampling: a surface-conditioned harmonic source, a PaiNN-style controller with three typed message-passing branches, specialized flow-matching pretraining, and a damping regularizer. It samples full Cartesian adsorbate coordinates, so internal conformers are explored during generation. At inference, candidates are refined with bond-restrained UMA relaxation, validity-filtered, and ranked by unrestrained UMA energy. On extensive ID and 50 OOD OC20-Dense validation systems, it reports the best MLIP Pass@k at all budgets and better DFT SR@k at small k.

Critical strengths
  1. The problem is valid and sound. The asymmetry between cheap MLIP energy/force queries and expensive dense placement labels is real, and casting placement as energy-based sampling is a natural way to exploit it.
  2. The method is a well-motivated adaptation rather than a naive borrowing from the existing works: the typed-branch controller, slab-conditioned source, and the replay-buffer treatment are collectively specialized to this science problem. Meanwhile, the architecture ablation (Table 4) supports the claim that this inductive bias matters more than capacity (a 100M-parameter DiT does worst). This is an interesting comparison of architecture design for such specialized scientific task.
Critical weaknesses
  1. The evaluation set is very small with no uncertainty estimates. 44 ID and 50 OOD systems out of 1000 per split. Many of the reported differences in DFT SR@k are 1–3 systems (e.g. ID SR@2: 36.4 vs 27.3 = 4 systems). No confidence intervals or even seed variance are reported. The selection criterion for the subset is also not stated, which raises a selection-bias question.

  2. There is no compute-cost comparison. Flow Sampling requires MLIP gradient queries throughout training, and inference uses 250 Euler–Maruyama steps per candidate before relaxation. Meanwhile, AdsorbML needs only relaxations. The paper never reports training cost, number of MLIP calls, or even a wall-clock time per system.

  3. There are missing baselines. CatFlow is the only prior full-coordinate generator and is discussed in the intro but not compared. AdsorbFlow is also cited and not compared.

  4. “Without dense supervision” is exaggerated. The model is pretrained on ~460k DFT-relaxed OC20 IS2RE structures, and removing this costs ~11 points of ID Pass@1 and ~20 points of Pass@10. This is already supervision.

Questions for authors

See above weaknesses.

2 comments

We thank the reviewer for their response. References are to the revised PDF.


W1. Uncertainty and selection.

We now report DFT SR@k as the mean ± std over four candidate orderings (Table 2) and paired bootstrap 95% intervals over systems for every comparison (App. D.2, Table 10). The reviewer’s example, ID SR@2, becomes 36.9 ± 4.7 against 23.3 ± 7.0 for AdsorbML, a larger gap than in the initial submission. Pooled over both splits, the gain over AdsorbML excludes zero for MLIP Pass@k at every budget and for DFT SR@k at k=1 (+9.3 [+3.7, +14.9]) and k=2 (+7.4 [+0.5, +14.4]). Against AdsorbDiff and CatFlow, the Pass@k gains exclude zero at every budget. The evaluation subsets are exactly those of AdsorbDiff, with no further selection (§4).


W2. Compute.

Added as App. C.5 (Table 5). Training is a one-time cost of 626 A100 GPU-hours, of which only about 5% goes to the 1.88M UMA evaluations, and the trained model applies to unseen systems without retraining, as in our OOD evaluation. At inference, AdsorbSample is cheaper per candidate than AdsorbML. Generation takes 0.29 s and calls no MLIP, which is 1.9% of the relaxation time, and AdsorbSample candidates relax in 14.8% fewer UMA steps, so the total cost per candidate is 13.2% lower. At low budgets, AdsorbSample also reaches the same DFT success with fewer candidates. Its pooled SR@1 (27.6%) is close to AdsorbML’s pooled SR@2 (28.2%), so one AdsorbSample candidate (15.7 s) reaches about the same success rate as two AdsorbML candidates (36.2 s).


W3. Baselines.

Both are added under our evaluation pipeline (Tables 1–2; §4, L314–319; App. C.2, L958–996). We trained AdsorbFlow with its released code and gave CatFlow the ground-truth slab, which removes the slab-generation part of its task. AdsorbSample achieves higher MLIP Pass@k and DFT SR@k than both at every budget on both splits, with Pass@10 of 66.0% against 53.3% and 28.2% on ID and 70.9% against 58.0% and 20.8% on OOD. This holds even though AdsorbFlow trains on the dense placement labels that AdsorbSample does not use.


W4. Supervision.

Thank you for this comment. We now write "dense per-system placement labels" (abstract, L013 and L017). The new pretrained-only row (Tables 1–2; §5.2, L417–421) quantifies what IS2RE pretraining contributes on its own: Pass@1 of 10.6% on ID and 12.2% on OOD, below AdsorbML. Energy training, which uses no structural labels, raises Pass@1 to 27.2% and 23.8%, so the performance that distinguishes AdsorbSample comes from energy training.

As the discussion period closes today, we wanted to follow up on our response above. The revision addresses each of the four weaknesses with new results: uncertainty estimates over candidate orderings and systems (W1), a compute-cost comparison (W2), the AdsorbFlow and CatFlow baselines (W3), and revised wording on supervision (W4).

If any concern remains, or if any part of the presentation is still unclear, we would be glad to address it in the remaining time. Thank you again for your review.

Reviewer_917cOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AdsorbSample uses a pretrained machine-learned interatomic potential to train a conditional diffusion model for generating adsorbate configurations on a catalyst surface. It combines OC20 structural pretraining with Flow Sampling, a surface-conditioned source distribution, a specialized equivariant controller, and geometric restraints, then evaluates generated candidates after MLIP relaxation and DFT single-point calculations. The model generates full adsorbate coordinates, allowing internal conformational changes in addition to rigid placement. The reported OC20-Dense results show improved MLIP discovery metrics and encouraging DFT success at small candidate budgets, while the independent contributions of energy training and flexible generation, and the physical interpretation of the final structures, remain unresolved.

Critical strengths
  1. The approach addresses a concrete data bottleneck. Using MLIP energies and gradients for the adaptation stage reduces dependence on dense placement labels for each adsorbate–surface system. The paper describes a specific route for applying an existing energy-based training method to this setting.
  2. The evaluation includes an energy reference beyond the training potential. At DFT SR@1, AdsorbSample obtains 25.0% versus 13.6% for AdsorbML on ID systems and 28.0% versus 18.0% on OOD systems. The paper also reports candidates with DFT single-point energies below the OC20-Dense reference minima. These results support further investigation of low-budget discovery.
  3. The relaxation control is applied across methods. The main comparisons use bond-restrained relaxation for the competing methods, and the relaxation ablation examines restrained and unrestrained relaxation for all three. This is an informative control for a consequential component of the pipeline.
  4. The negative and system-specific results expose practical limitations. The movable-surface variant performs poorly, and the mode-collapse example shows why motif-stratified replay can help in a particular system. The paper acknowledges that this replay modification is not yet part of the main amortized model.
Critical weaknesses

W1. The pretrained-only baseline is missing. The ablation removes structural pretraining but omits direct evaluation of the pretrained flow-matching model, leaving the incremental benefit of energy training unclear. Please evaluate that checkpoint with the same candidate budget, filtering, relaxation, and ranking, including a focused DFT comparison.

W2. The contribution of flexible generation could be separated more clearly from the additional search components. The GGE/Ni(111) experiment adds system-specific training, restraints, temperature changes, stratified replay, and local refinement, while the rigid-placement conformer budget is underspecified. Please compare full-coordinate and rigid-body variants under matched settings, and include a flexible torsion-search control or a constrained-minima-hopping baseline [1] with the same energy/force-call budget.

W3. The stability of the restrained outputs remains unresolved. The pipeline omits final unrestrained relaxation, so restraint forces may balance nonzero physical forces at the reported geometries. Please report residual unrestrained forces and test a representative subset after releasing restraints, with limited DFT relaxation for the claimed adsorption minima.

W4. Basin comparisons would benefit from consistent relaxation conventions. Generated candidates are relaxed with a restrained MLIP, whereas references come from DFT relaxation; the strict 0.05 Å A-RMSD and 0.02 eV thresholds can label these endpoint differences as unmatched basins. Please compare consistently processed structures, examine threshold sensitivity, and supplement clustering with adsorption-site or motif labels.

W5. The periodic-distance formula needs clarification for nonorthogonal cells. Component-wise wrapping of fractional coordinates can return a nonminimal Cartesian displacement in a nonorthogonal cell. Please clarify the supported cell geometry and actual minimum-image implementation, and correct the formula accordingly.

W6. Sampling and validity terminology could be more precise. Validity is measured after relaxation, and stratified replay and local refinement modify the training endpoint distribution in the special-case experiments. Please use post-relaxation validity consistently, clarify how those modifications relate to the Boltzmann target, and correct the narrow description of the low-energy coverage evaluation in CatFlow [2].

Questions for authors

Please prioritize the pretrained-only and rigid-body controls (W1–W2). I am particularly interested in whether they preserve the claimed advantage of energy training and flexible generation. Positive results, together with satisfactory stability and basin checks (W3–W4), will strengthen my assessment enough to raise the rating.

  1. Does the periodic-distance implementation handle nonorthogonal cells using a true minimum-image calculation?
  2. How are the 44 ID and 50 OOD systems selected, and with respect to which sampler-training and MLIP-training data is each OOD designation defined?
2 comments

We thank the reviewer for their detailed response. References are to the revised PDF.


W1. Pretrained-only baseline.

We evaluated the pretrained checkpoint under the identical pipeline, including DFT verification (Tables 1–2, "pretraining only" rows; §5.2, L417–426). On its own, it reaches MLIP Pass@1 of 10.6% on ID and 12.2% on OOD, below AdsorbML (16.5% and 17.8%). Energy training raises Pass@1 to 27.2% and 23.8%, the highest of all methods, and more than triples DFT SR@1, from 6.8% and 8.0% to 27.8% and 27.5%. Since the pretrained checkpoint alone falls below AdsorbML, the advantage of AdsorbSample comes from energy training.


W2. Flexible control.

We ran constrained minima hopping (CMH) [1] with UMA and the same GGE restraints under two budgets (§5.3, L496–513; App. E.4, L1472–1496, Table 16). At matched GPU time, CMH returns 18 passing minima in 7 basins and AdsorbSample 5 in 5, with six times fewer UMA evaluations. At matched UMA evaluations (2 GPU-hours of CMH), CMH returns 2 passing minima in 1 basin. AdsorbSample therefore finds more basins per energy evaluation, and CMH finds more per GPU-hour once training cost is counted. Both flexible methods recover the two-ring geometry, which rigid placement misses in 400 trials across 60 RDKit conformers (App. E.4, L1467–1470).

These results support the central claim of §5.3: an energy-trained diffusion sampler scales to a 43-atom adsorbate with seven rotatable bonds and reaches the same low-energy mode as a dedicated global optimizer, with fewer energy evaluations. CMH confirms that the target is hard but reachable by flexible search. The two methods also spend compute differently, as CMH's cost is paid in full for every new adsorbate–surface pair, whereas AdsorbSample's is dominated by training, which a conditional model can in principle amortize across systems. Since this is a single system, we report both budgets in §5.3 (L507–513) and do not claim a general ranking of the two methods.


W3. Stability.

The results hold after releasing the restraint and after DFT relaxation (App. D.1, L1092–1162; Tables 6–9). We released the restraint on all candidates of AdsorbSample, AdsorbML, and AdsorbDiff and relaxed them further with unrestrained UMA, after which about 95% of structures of every method are UMA minima (Table 6). AdsorbSample keeps the highest Pass@k at every budget (Table 7), and its DFT SR@k changes by less than 3 points (Table 8). We then relaxed every SR@1-selected candidate of the three methods with DFT. Every single-point success remains a success, and two single-point failures become successes, so single-point SR@1 is slightly conservative. For the claimed adsorption minima, DFT relaxation confirms 20 of the 22 basins that AdsorbSample keeps at least 0.1 eV below the OC20-Dense reference, against 6 for AdsorbML and none for AdsorbDiff (Table 9; §5.1, L400–405).


W4. Basins.

Added as App. D.3; Tables 11–12. References are now OC20-Dense structures relaxed with unrestrained UMA. At A-RMSD thresholds of 0.05, 0.3, and 1.0 Å, AdsorbSample finds the most low-energy basins on both splits (ID: 497, 260, and 192, against 345, 167, and 125 for AdsorbML), and even at 1.0 Å more than half of its basins match no OC20-Dense reference (53% on ID, 57% on OOD; Table 11). The same ranking holds when the generated structures are relaxed without the restraint, as the references are (Table 12). The 0.02 eV criterion only merges generated configurations into basins and is not used for reference matching, so it cannot create unmatched basins.


W5 and Q1.

Thank you for catching this. Our cells are general triclinic slab cells, and Eq. 41–42 (App. B.3, L927–938) now take the minimum-norm displacement over the neighboring in-plane lattice translations. Recomputing A-RMSD with the exact minimum image changed no value below 2.5 Å, well above the largest threshold we use (1.0 Å), so every reported result is unchanged.


W6.

We use post-relaxation validity throughout (abstract, L025; §4, L324–325). Stratified replay and endpoint refinement appear only in the single-system case studies, and App. E.3 (L1424–1426) now states that they bias these models toward rare low-energy modes, so we use the models as search procedures. The amortized model of §5.1 uses neither and is trained toward the restrained Boltzmann target. The CatFlow description is corrected (§1, L078–082).


Q2.

Added to §4 (L305–313): we use AdsorbDiff's evaluation subsets with no further selection, all held out from our training, with OOD defined as in OC20.


[1] Global Optimization of Adsorbate–Surface Structures While Preserving Molecular Identity

Thank you again for the detailed review. As the discussion period closes today, we wanted to follow up on our response above. For convenience, here is how the revision addresses the points you asked us to prioritize:

  • W1: The pretraining-only checkpoint reaches Pass@1 of 10.6% (ID) and 12.2% (OOD); energy training raises this to 27.2% and 23.8%, respectively.
  • W2: Constrained minima hopping is added as a flexible control. Both flexible methods recover the geometry that rigid placement does not.
  • W3: Every SR@1 success persists under DFT ionic relaxation, and 20 of 22 below-reference basins are confirmed by DFT relaxation.
  • W4: AdsorbSample finds the most basins at every matching threshold.

We would be glad to clarify anything in the remaining time.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to the Area Chair and reviewers

We thank the Area Chair and the reviewers for their careful reading and constructive suggestions. We are encouraged that the reviewers recognized the paper's core contribution: training a sampler from MLIP energies addresses a real supervision bottleneck in adsorbate placement, since energy and force queries are cheap while dense placement labels are expensive. Reviewers also highlighted the well-motivated, problem-specific design of the method, with full-coordinate generation and a surface-aware controller, and the care taken in the evaluation, which includes DFT verification, relaxation controls applied to every method, and ablations that report negative results.

The reviewers' questions focused on how robust these results are and how they compare with further baselines. In response, we added seven new sets of experiments and analyses. The main conclusions are unchanged, and we summarize below how each new result strengthens them.

1. Robustness of the main results

  • Statistical uncertainty (FKiu, hbwv). We now report DFT success over several candidate orderings, with paired bootstrap intervals over systems. The low-budget DFT gain over AdsorbML and the MLIP Pass@k gain at every budget both remain clear, which shows the main result does not depend on one candidate ordering.
  • Restraint release and DFT relaxation (917c, hbwv). After releasing the bond restraint and relaxing further, AdsorbSample remains the best method in MLIP Pass@k, and DFT relaxation confirms the selected candidates and nearly all structures found below the OC20-Dense reference. The reported structures are therefore physically meaningful minima, not products of the restraint or of single-point evaluation.
  • Low-energy basins (917c). With consistently relaxed references and across a range of matching thresholds, AdsorbSample still finds the most low-energy basins, many of them absent from OC20-Dense.

2. Additional baselines and ablations

  • AdsorbFlow and CatFlow (FKiu). We added a second supervised placement model and the only prior full-coordinate generator. AdsorbSample outperforms both in MLIP Pass@k and DFT SR@k at every budget. This extends the central claim to a broader set of methods.
  • Pretraining only (917c, FKiu). The pretrained checkpoint on its own performs below the search-based baseline, and the gains appear only after energy training. This isolates energy training as the source of the improvement and clarifies that structural pretraining serves as an initialization.
  • Compute cost (FKiu, hbwv). Training is a one-time cost that carries over to unseen systems, and inference is 13% cheaper per candidate than AdsorbML, because generation is negligible compared to relaxation and our candidates relax in fewer steps. The gains therefore come at no extra inference cost.

3. Flexible adsorbates

  • Constrained minima hopping control (917c, hbwv). We compared against a non-learned flexible global optimizer under matched budgets. Both recover the low-energy geometry that rigid placement misses. AdsorbSample finds more distinct low-energy structures per energy evaluation, and minima hopping finds more per GPU-hour. This supports the claim that full-coordinate sampling scales to large flexible adsorbates, which we present as a single-system case study.

We also made the corrections and clarifications the reviewers requested. We believe these additions have significantly strengthened the paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.