acceptodds
Withdrawn from ICLR 2027

Adsorption-Site-Guided Message Passing for Catalytic System Energy Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph neural networks (GNNs) have been widely used for energy prediction in catalytic systems. However, existing methods primarily focus on modeling geometric relationships among atoms, while adsorption-site information has not been sufficiently incorporated into message passing. To address this limitation, we propose an adsorption-site-guided gated message-passing framework that encodes the distances between atoms and adsorption-site using radial basis functions (RBFs) and integrates the resulting site features into node or directed-edge representations. A gating mechanism then adaptively modulates the contribution of site information based on the current structural representations, enabling effective integration of adsorption-site information with the backbone representations. The proposed framework is integrated into SchNet, DimeNet++, and PaiNN and evaluated on the Initial Structure to Relaxed Energy (IS2RE) task of the Open Catalyst 2020(OC20) dataset. Experiments across three data regimes—10k, 100k, and the full training set—show that the proposed framework improves upon the corresponding baselines across most models and evaluation metrics, with more pronounced gains in the low-data regime. Ablation studies on SchNet further demonstrate that adsorption-site information and the gating mechanism play complementary roles, and their combination yields consistent performance improvements.These results demonstrate that the proposed framework can effectively exploit structural information associated with adsorption-site and is broadly applicable across different graph neural network architectures.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Adsorption-site-guided message passing, Geometric graph neural networks, Gated message passing, Catalytic energy prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_aAnvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a plug-in module for direct IS2RE prediction on OC20. The module encodes the distance between each atom and the adsorption site with RBFs, adds it to the atom or edge features, and modulates the result with a sigmoid gate at every message-passing layer. The authors apply the module to SchNet, DimeNet++, and PaiNN and report lower MAE on the 10k, 100k, and full training sets, with the largest gains at 10k. An ablation on SchNet suggests that the distance encoding and the gate are complementary.

Critical strengths

S1. The module is simple and applies to backbones with different message-passing schemes. The same distance encoding and gating attach to node features in SchNet, directed-edge features in DimeNet++, and edge RBFs in PaiNN. All nine backbone and data-size combinations show lower aggregated MAE, with the largest gains in the 10k regime.

Critical weaknesses

W1. The evaluation does not establish the module’s value on stronger backbones.
The experiments are restricted to SchNet, DimeNet++, and PaiNN. Published results for direct OC20 IS2RE without auxiliary tasks or Noisy Nodes report val_id MAEs of 0.5088 eV for Equiformer and 0.5161 eV for EquiformerV2, below the proposed method’s best result of 0.5547 eV. These literature results do not constitute a controlled comparison, but they highlight the absence of stronger backbones from the evaluation. Since the contribution is presented as a broadly applicable module, demonstrating that it also improves a stronger backbone is more informative than improvements on older architectures alone. The current experiments leave its practical value beyond the evaluated backbones unresolved.

W2. The methodological advance over existing locality-based modeling is insufficiently established.
The module combines site-distance RBF encodings, feature fusion, and sigmoid gating. These are established architectural components, and the cited local-environment pooling work already exploits the importance of atoms near the adsorbate. The paper’s main distinction is therefore the repeated incorporation of site information into intermediate message-passing representations. However, it does not compare this design with adding the same site-distance features only at the input, injecting them only at readout, or using local-environment pooling under the same backbone and training protocol. The site-only/gate-only ablation tests the combination of components but does not isolate the benefit of their placement across message-passing layers. Consequently, the experiments establish improvements over the original backbones, but do not establish the added value of the central architectural distinction over simpler ways of supplying the same local prior.

Questions for authors

Q1 (experiment request). Could you apply the module to a stronger backbone such as Equiformer or EquiformerV2 under the same direct IS2RE setting? Gains on such a backbone would support the claim that the module is broadly applicable.

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

This paper proposes an adsorption-site-guided message-passing framework for energy prediction in catalytic systems. The framework is a plug-in module that works across architectures. It takes the adsorption-site coordinates provided in the OC20 metadata as an additional input. Each atom’s distance to the adsorption site is encoded with radial basis functions. The resulting site features are fused into the backbone’s intermediate representations through a residual connection followed by layer normalization. A sigmoid gate then modulates the fused features channel-wise. The authors instantiate the framework on three geometric GNNs, each at a different level of representation:

  • SchNet: node features
  • DimeNet++: directed-edge features, weighted by a smooth distance-based relevance score around the site
  • PaiNN: edge RBF features

The method is evaluated on the OC20 IS2RE (direct) task with three training-set sizes (10k, 100k, and All). Results are reported on the in-domain validation split and the three OOD validation splits, using MAE, MSE, and EwT. The site-guided variants generally improve over their baselines. The gains are largest in the low-data regime; for example, DimeNet++ aggregated-OOD MAE drops from 0.893 to 0.744 eV at 10k. With the full training set, the gains shrink to roughly 1–7%.

Critical strengths
  1. Architecture-agnostic plug-in design. The proposed module is lightweight. It requires only one extra input, the adsorption-site coordinate, and the authors integrate it into three backbones with different representation types: node-level (SchNet), directed-edge-level (DimeNet++), and equivariant scalar–vector (PaiNN). This matters for my assessment because a site-aware module that does not depend on a specific architecture could in principle be adopted by many existing GNN pipelines for catalysis, which makes the idea practically relevant beyond a single model.

  2. Consistent gains in the low-data regime, including OOD splits. In the 10k setting, all three site-guided variants improve over their baselines on every metric (MAE, MSE, EwT) and on every validation split, including all three OOD subsets (Table 1 and Table 3). Some of these gains are substantial; for example, DimeNet++ aggregated-OOD MAE drops from 0.893 to 0.744 eV. This is the most interesting finding of the paper. Labeled DFT data is expensive in real catalyst-discovery settings, so a structural prior that improves data efficiency and OOD generalization would be valuable. However, the results come from a single run, and the mechanism behind the gains is not analyzed.

Critical weaknesses
  1. Evaluation limited to IS2RE-direct. All experiments use the direct setting, where the model maps the initial structure to the relaxed energy in one pass. The mainstream pipeline instead relaxes the structure with an S2EF-trained ML potential and then reads off the energy. That pipeline achieves roughly 0.29–0.31 eV MAE on the IS2RE test split (EquiformerV2, with or without DeNS), compared with about 0.55 eV (val-ID) for the best model here. After relaxation, the adsorbate–surface environment is explicit in the geometry. It is therefore unclear whether the site prior still helps in that setting.

  2. Unclear benefit over existing locality and tag information. The backbones already restrict message passing with a distance cutoff, and OC20 tags already mark which atoms belong to the adsorbate. The adsorption-site coordinate corresponds to the binding adsorbate atoms, which are already present in the input. The site feature may therefore act mainly as a soft, site-centred cutoff (Eq. 9–10 for DimeNet++ is literally one). Could authors experiment on tuning the backbone cutoff, and replacing the metadata site with the distance to the nearest adsorbate atom (tag = 2).

  3. More analysis on low-data gains. The 10k improvements are the most interesting result, but they come from a single run with no standard deviations. Table 2 also suggests high variance: at 100k, adding either module alone worsens OOD MAE by about 14%. Please report multiple seeds, a denser learning curve, and a breakdown of the gains by adsorbate class, catalyst type, and distance to the site.

Questions for authors

Please also refer to weaknesses.

  1. Could the authors compare the proposed model with baselines under S2EF settings?

  2. Could the authors provide the base model results under different cutoff settings? Is cut-off sensitive to the base models? Is the effect and mechanism the same as the proposed model?

  3. The 10k improvements are the most interesting result, but the current evidence does not show that they are robust or that they come from the site prior. Could the authors address the following?

  • Could the authors report mean ± std over at least three seeds for both the baseline and the site-guided model at 10k and 100k?
  • At 100k, adding either module alone worsens OOD MAE by about 14% (Table 2). This suggests substantial run-to-run variance, and the 10k gains need to be compared against that variance.
  • Could the authors report a denser learning curve, for example 5k, 10k, 20k, 50k, 100k, 200k, and All.
  • Does the gap shrink smoothly as data increases, or vanish beyond a certain size?
  • How many training samples does the baseline need to match the site-guided model at 10k? An “equivalent-data” number would make the data-efficiency claim concrete.
Reviewer_yb53OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Adsorption-site-guided message passing augments geometric graph neural networks with an explicit site representation for initial-structure-to-relaxed-energy prediction on OC20. Distances to the site are encoded and injected into intermediate representations, with a learned gate controlling their contribution. The framework is instantiated in SchNet, DimeNet++, and PaiNN, and evaluated at three training-set sizes with separate in-domain and out-of-domain results. The reported improvements are most pronounced in the small-data setting, but the architectural symmetry guarantees and the contribution of repeated site conditioning remain insufficiently established.

Critical strengths
  1. The proposed prior is concrete and the intervention is identifiable. The method specifies how site-distance information enters node or edge representations, rather than relying only on a general claim that local chemistry matters. The three backbone adaptations provide useful examples of how to incorporate this information into existing models.
  2. The evaluation covers several architectures and data regimes. SchNet, DimeNet++, and PaiNN are evaluated on 10k, 100k, and full-data training, with the individual OOD subsets reported separately. This allows readers to examine where the method helps and where improvements do not extend to every metric.
  3. The small-data results are promising. For DimeNet++ trained on 10k examples, aggregate OOD MAE decreases from 0.8927 to 0.7438 eV. This is a useful empirical motivation for studying explicit adsorption-site information.
  4. The paper acknowledges meaningful limitations. Dependence on a predefined site is discussed, and the appendix cautions that internal atomic responses should not be interpreted as direct physical contributions to the energy.
Critical weaknesses

W1. The gate's geometric invariance would benefit from clarification. The stated gate definition uses Cartesian site coordinates, so its translation invariance depends on unspecified preprocessing; periodic-image handling is also unspecified. Please clarify the actual gate inputs for each backbone and the site construction, and check existing checkpoints under equivalent translations, rotations, and periodic representations.

W2. The controls could better separate site conditioning from other changes. Only the PaiNN 10k/100k comparisons explicitly match training protocols, while the site module also changes projections and normalization and the gate-only variant may retain site coordinates. A representative comparison should match training and model capacity and list the inputs available to each ablation.

W3. Comparisons with simpler local priors are missing. The experiments omit PhAST-style task-specific controls, input-only site-distance features, and local-environment readout. Please compare these alternatives on the same backbone and budget to assess the benefit of repeated gated injection.

W4. Aggregate and subset OOD results need reconciliation. Reaggregation under the stated pooling rule and standard OC20 subset sizes suggests discrepancies in eight SchNet-Ours entries; for example, the 100k OOD EwT aggregates to approximately 0.02503 rather than the reported 0.0242. Please provide the actual subset sizes and checkpoint identifiers and reconcile the reported values.

W5. Several implementation details would be helpful for reproduction. The PaiNN update and the backbone-specific injection layers, parameter-sharing rules, cutoffs, and residual-coefficient initialization are underspecified, and parameter counts are missing.

W6. The response visualization would benefit from a clearer definition. The visualization uses one example, leaves the scalar response undefined, and omits the corresponding backbone visualization.

Questions for authors

I would first like to understand whether the implemented gate preserves geometric invariance and how the OOD results were aggregated (W1, W4). Resolving these points, with a matched comparison supporting the benefit of site conditioning (W2–W3), will bring my rating to 3/4.

  1. How is the adsorption-site vector constructed from the available metadata or initial structure, and what input is assumed at deployment?
  2. Which layers are modified in PaiNN, and how is parameter sharing configured?

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

Related papers

Discussion (0)

Sign in to comment.