pith. sign in

arxiv: 2605.04102 · v1 · submitted 2026-05-03 · ❄️ cond-mat.mtrl-sci · cs.AI· cs.LG

Meta-LegNet: A Transferable and Interpretable Framework for Surface Adsorption Prediction via Self-Defined Adsorption-Environment Learning

Yifan Li , Arravind Subramanian , Xiaoqing Liu , Qiujie Lyu , Sergey Kozlov , Lei Shen This is my paper

Pith reviewed 2026-05-09 17:00 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci cs.AIcs.LG
keywords adsorption predictiongraph neural networkscatalyst screeningtransfer learninginterpretable machine learningsurface sciencecomputational catalysisequivariant networks
0
0 comments X

The pith

A graph network learns transferable local adsorption environments to propose sites on new catalyst surfaces without full enumeration.

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

The paper develops a graph-based model that captures patterns in atomic arrangements favorable for adsorption by training across multiple catalyst-adsorbate pairs. It encodes these patterns using directional graph updates and multi-scale spatial pooling so the same features apply to materials and molecules not seen during training. If the patterns hold, the model can match new surfaces to a database of learned environments and suggest promising sites directly. This also yields maps that attribute importance to specific local regions, allowing chemists to see why a site is favored. The approach targets the bottleneck in catalyst discovery where checking every possible adsorption configuration with quantum calculations becomes prohibitive for complex systems.

Core claim

Meta-LegNet combines SE(3)-equivariant atom-level message passing with voxel-based multiscale aggregation and cross-domain meta-learning to encode transferable local adsorption environments, then uses the resulting representations and an environment database to identify adsorption sites on unexplored surfaces through template matching.

What carries the argument

The adsorption-environment encoder that extracts invariant radial features and equivariant directional signals from graphs, fuses them with voxel pooling and gated attention to separate local and global context, and produces atom-resolved attribution maps for template-based site proposal.

If this is right

  • Adsorption sites on new surfaces can be proposed by matching to stored environment templates instead of evaluating every candidate location.
  • Atom-level attribution maps directly indicate which neighboring atoms and directions most influence adsorption strength.
  • A single model trained across domains replaces separate models for each catalyst-adsorbate pair.
  • The environment database supports rapid screening of additional adsorbates without retraining from scratch.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same environment representations could serve as input features for predicting full reaction barriers once adsorption is located.
  • Pairing the method with active-learning loops would let the model request DFT data only for the most uncertain new environments.
  • Attribution maps might reveal recurring geometric motifs that guide the design of alloy or defect-engineered surfaces.

Load-bearing premise

Local chemical environments extracted from the training set of surfaces and adsorbates remain sufficiently similar and predictive when applied to entirely new surfaces and adsorbates.

What would settle it

Apply the trained model to a previously unseen catalyst-adsorbate combination, generate proposed sites, and compare their DFT-relaxed energies against those from exhaustive site enumeration; large systematic overestimation of binding strength would falsify transferability.

Figures

Figures reproduced from arXiv: 2605.04102 by Arravind Subramanian, Lei Shen, Qiujie Lyu, Sergey Kozlov, Xiaoqing Liu, Yifan Li.

Figure 1
Figure 1. Figure 1: Comparison between the conventional sequential adsorption workflow and the proposed direct adsorption view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Meta-LegNet. The framework integrates multidimensional adsorption datasets spanning 0D, view at source ↗
Figure 3
Figure 3. Figure 3: Statistical overview of the adsorption benchmark. (a) Elemental diversity and mean number of atoms for each view at source ↗
Figure 4
Figure 4. Figure 4: Architecture of Meta-LegNet. Starting from atomistic graphs, the model combines self-defined adsorption view at source ↗
Figure 5
Figure 5. Figure 5: Multiscale equivariant feature fusion in Meta-LegNet. Coarse-grained graph features provide broader view at source ↗
Figure 6
Figure 6. Figure 6: Cross-domain meta-learning in Meta-LegNet. A reorganized benchmark spanning multidimensional ad view at source ↗
Figure 7
Figure 7. Figure 7: Validation of the learned adsorption environments against geometric reference environments. (a) Mean view at source ↗
Figure 8
Figure 8. Figure 8: Masking-based faithfulness analysis of the learned adsorption environment. Removing the learned environment view at source ↗
Figure 9
Figure 9. Figure 9: SEAM enables direct adsorption-site localization on unseen surfaces. (a) For representative 2D and 3D view at source ↗
Figure 10
Figure 10. Figure 10: Distributional comparison between the learned adsorption environments and geometric reference environ view at source ↗
Figure 11
Figure 11. Figure 11: Dependence of overlap statistics on the size of the learned adsorption environment. (a) IoU with the distance view at source ↗
read the original abstract

A central challenge in computational catalysis is the identification of low-energy and chemically plausible adsorption configurations, as these directly affect adsorption energies, reaction pathways, and catalytic performance. Existing approaches generally rely on enumerating candidate adsorption sites followed by iterative refinement through density functional theory calculations or machine-learning-based relaxations. However, such workflows remain computationally expensive and are difficult to scale to complex surfaces or multi-adsorbate systems. Here, we introduce Meta-LegNet, a graph learning framework that combines SE(3)-equivariant atom-level message passing with voxel-based multiscale aggregation and cross-domain meta-learning to learn transferable representations of local adsorption environments across diverse catalyst--adsorbate systems. Rather than following a conventional regression-only paradigm, Meta-LegNet encodes local chemical environments using invariant radial features and equivariant directional information, and further incorporates broader structural context through coordinate-frame voxel pooling, assignment-based upsampling, and gated feature fusion. The resulting local-global decomposition produces atom-resolved attribution maps, which are processed to identify adsorption-relevant local environments in an interpretable manner. Based on the learned representations, we further construct an adsorption-environment database and develop a template-matching strategy to propose likely adsorption sites on previously unexplored surfaces without exhaustive site enumeration. Overall, our results suggest that learning transferable adsorption environments provides an accurate, interpretable, and practical route for accelerating catalyst screening.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces Meta-LegNet, a graph learning framework that integrates SE(3)-equivariant atom-level message passing, voxel-based multiscale aggregation, coordinate-frame pooling, and cross-domain meta-learning to encode transferable representations of local adsorption environments across catalyst-adsorbate systems. These representations support atom-resolved attribution maps for interpretability and enable a template-matching strategy to propose adsorption sites on previously unseen surfaces without exhaustive enumeration. The central claim is that this approach yields an accurate, interpretable, and practical route to accelerate catalyst screening.

Significance. If the transferability and accuracy claims are quantitatively validated, the work would be significant for computational materials science and catalysis. It addresses the scalability bottleneck of site enumeration and DFT relaxation by learning reusable local environments and providing an interpretable decomposition, which could enable faster screening of complex multi-adsorbate systems. The combination of equivariant networks with meta-learning and voxel pooling is a technically coherent extension of existing GNN approaches in the field.

major comments (2)
  1. [Abstract] Abstract: The statement that 'our results suggest that learning transferable adsorption environments provides an accurate, interpretable, and practical route' supplies no quantitative support (MAE, RMSE, R², out-of-distribution accuracy, ablation on the meta-learning component, or comparison to non-meta baselines). Without these metrics the central claim of transferability to unexplored surfaces cannot be assessed and the practicality argument remains ungrounded.
  2. [Abstract] The template-matching strategy and adsorption-environment database construction are described at a high level but no validation protocol (e.g., held-out catalyst-adsorbate pairs, error on proposed sites versus DFT-relaxed ground truth, or comparison to random or heuristic site selection) is reported. This is load-bearing for the claim that the method avoids exhaustive enumeration while remaining predictive.
minor comments (1)
  1. [Abstract] The title uses 'Self-Defined Adsorption-Environment Learning' but the abstract does not explicitly define what 'self-defined' means in contrast to standard supervised or unsupervised environment learning.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and positive overall assessment of Meta-LegNet. The comments highlight opportunities to strengthen the abstract, and we have revised it accordingly while preserving its concise nature. Below we respond point by point to the major comments.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The statement that 'our results suggest that learning transferable adsorption environments provides an accurate, interpretable, and practical route' supplies no quantitative support (MAE, RMSE, R², out-of-distribution accuracy, ablation on the meta-learning component, or comparison to non-meta baselines). Without these metrics the central claim of transferability to unexplored surfaces cannot be assessed and the practicality argument remains ungrounded.

    Authors: We agree that the original abstract did not include explicit numerical metrics. The main text and supplementary information contain the requested quantitative results, including MAE/RMSE values for adsorption-energy prediction, R² coefficients, out-of-distribution accuracy on held-out catalyst–adsorbate pairs, ablation studies isolating the meta-learning component, and direct comparisons against non-meta baselines. In the revised manuscript we have updated the abstract to incorporate representative quantitative support (e.g., out-of-distribution MAE and ablation outcomes) so that the transferability and practicality claims are grounded within the abstract itself. revision: yes

  2. Referee: [Abstract] The template-matching strategy and adsorption-environment database construction are described at a high level but no validation protocol (e.g., held-out catalyst-adsorbate pairs, error on proposed sites versus DFT-relaxed ground truth, or comparison to random or heuristic site selection) is reported. This is load-bearing for the claim that the method avoids exhaustive enumeration while remaining predictive.

    Authors: The validation protocol for the template-matching strategy—including evaluation on held-out catalyst–adsorbate pairs, site-proposal error relative to DFT-relaxed ground truth, and comparisons against random and heuristic baselines—is presented in the Methods and Results sections of the full manuscript. To address the referee’s concern that this information is not visible in the abstract, we have revised the abstract to concisely state the validation approach and its key outcomes, thereby making the claim that exhaustive enumeration can be avoided while retaining predictive accuracy more explicit and self-contained. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is data-driven ML architecture without self-referential reductions

full rationale

The paper describes Meta-LegNet as a graph neural network framework that learns representations of adsorption environments from diverse training systems using SE(3)-equivariant message passing, voxel pooling, and cross-domain meta-learning. Claims of transferability and practicality for catalyst screening are presented as empirical outcomes from training against external benchmarks rather than any definitional equivalence, fitted parameter renamed as prediction, or load-bearing self-citation. No equations or derivation steps are supplied in the provided text that reduce outputs to inputs by construction; the approach remains a standard supervised learning pipeline with independent content.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract supplies insufficient technical detail to identify or enumerate free parameters, background axioms, or newly postulated entities.

pith-pipeline@v0.9.0 · 5570 in / 1115 out tokens · 38800 ms · 2026-05-09T17:00:43.180664+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

  1. [1]

    Nørskov, Frank Abild-Pedersen, Felix Studt, and Thomas Bligaard

    Jens K. Nørskov, Frank Abild-Pedersen, Felix Studt, and Thomas Bligaard. Density functional theory in surface chemistry and catalysis.Proceedings of the National Academy of Sciences of the United States of America, 108(3):937–943, 2011

    work page 2011

  2. [2]

    Unke, Stefan Chmiela, Michael Gastegger, Kristof T

    Oliver T. Unke, Stefan Chmiela, Michael Gastegger, Kristof T. Schütt, Huziel E. Sauceda, Klaus-Robert Müller, and Alexandre Tkatchenko. Machine learning force fields.Chemical Reviews, 121(16):10142–10186, 2021

    work page 2021

  3. [3]

    Schweitzer, Fanglin Che, and Hongliang Xin

    Tianyou Mou, Hemanth Somarajan Pillai, Siwen Wang, Mingyu Wan, Xue Han, Neil M. Schweitzer, Fanglin Che, and Hongliang Xin. Bridging the complexity gap in computational heterogeneous catalysis with machine learning. Nature Catalysis, 6(2):122–136, 2023

    work page 2023

  4. [4]

    Lars G. M. Pettersson et al. Adsorption energies on transition metal surfaces: Towards an accurate description of surface chemistry.Nature Communications, 13:6853, 2022

    work page 2022

  5. [5]

    Weckhuysen

    Charlotte V ogt and Bert M. Weckhuysen. The concept of active site in heterogeneous catalysis.Nature Reviews Chemistry, 6(2):89–111, 2022

    work page 2022

  6. [6]

    A. F. Usuga, C. S. Praveen, and Aleix Comas-Vives. Local descriptors-based machine learning model refined by cluster analysis for accurately predicting adsorption energies on bimetallic alloys.Journal of Materials Chemistry A, 2024

    work page 2024

  7. [7]

    Local environment interaction-based machine learning framework for molecular adsorption energy prediction.Journal of Materials Informatics, 3:17, 2023

    Yifan Li, Yihan Wu, Yuhang Han, Qujie Lyu, Hao Wu, Xiuying Zhang, and Lei Shen. Local environment interaction-based machine learning framework for molecular adsorption energy prediction.Journal of Materials Informatics, 3:17, 2023

    work page 2023

  8. [8]

    Recent advances in the application of machine-learning algorithms to predict adsorption energies.Trends in Chemistry, 2022

    Aoqi Wang, Jing Li, et al. Recent advances in the application of machine-learning algorithms to predict adsorption energies.Trends in Chemistry, 2022

    work page 2022

  9. [9]

    Adsorbml: a leap in efficiency for adsorption energy calculations using generalizable machine learning potentials.npj Computational Materials, 9(1):172, 2023

    Janice Lan, Aini Palizhati, Muhammed Shuaibi, Brandon M Wood, Brook Wander, Abhishek Das, Matt Uyt- tendaele, C Lawrence Zitnick, and Zachary W Ulissi. Adsorbml: a leap in efficiency for adsorption energy calculations using generalizable machine learning potentials.npj Computational Materials, 9(1):172, 2023. 17 Running Title for Header

    work page 2023

  10. [10]

    Hoffmann, Christoph Engel, Sebastian Matera, and Karsten Reuter

    Max J. Hoffmann, Christoph Engel, Sebastian Matera, and Karsten Reuter. Automated exploitation of the big configuration space of large adsorbates on transition-metal catalysts.Nature Communications, 13:2180, 2022

    work page 2022

  11. [11]

    Open catalyst 2020 (oc20) dataset and community challenges.Acs Catalysis, 11(10):6059–6072, 2021

    Lowik Chanussot, Abhishek Das, Siddharth Goyal, Thibaut Lavril, Muhammed Shuaibi, Morgane Riviere, Kevin Tran, Javier Heras-Domingo, Caleb Ho, Weihua Hu, et al. Open catalyst 2020 (oc20) dataset and community challenges.Acs Catalysis, 11(10):6059–6072, 2021

    work page 2020

  12. [12]

    High-throughput identification of exfoliable two-dimensional materials with active basal planes for hydrogen evolution.ACS Energy Letters, 5(7):2313–2321, 2020

    Tong Yang, Jun Zhou, Ting Ting Song, Lei Shen, Yuan Ping Feng, and Ming Yang. High-throughput identification of exfoliable two-dimensional materials with active basal planes for hydrogen evolution.ACS Energy Letters, 5(7):2313–2321, 2020

    work page 2020

  13. [13]

    2dmatpedia, an open computational database of two-dimensional materials from top-down and bottom-up approaches.Scientific data, 6(1):86, 2019

    Jun Zhou, Lei Shen, Miguel Dias Costa, Kristin A Persson, Shyue Ping Ong, Patrick Huck, Yunhao Lu, Xiaoyang Ma, Yiming Chen, Hanmei Tang, et al. 2dmatpedia, an open computational database of two-dimensional materials from top-down and bottom-up approaches.Scientific data, 6(1):86, 2019

    work page 2019

  14. [14]

    High-throughput calculations of catalytic properties of bimetallic alloy surfaces.Scientific data, 6(1):76, 2019

    Osman Mamun, Kirsten T Winther, Jacob R Boes, and Thomas Bligaard. High-throughput calculations of catalytic properties of bimetallic alloy surfaces.Scientific data, 6(1):76, 2019

    work page 2019

  15. [15]

    The open catalyst 2022 (oc22) dataset and challenges for oxide electrocatalysts.Acs Catalysis, 13(5):3066–3084, 2023

    Richard Tran, Janice Lan, Muhammed Shuaibi, Brandon M Wood, Siddharth Goyal, Abhishek Das, Javier Heras-Domingo, Adeesh Kolluru, Ammar Rizvi, Nima Shoghi, et al. The open catalyst 2022 (oc22) dataset and challenges for oxide electrocatalysts.Acs Catalysis, 13(5):3066–3084, 2023

    work page 2022

  16. [16]

    Model-agnostic meta-learning for fast adaptation of deep networks

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. InInternational conference on machine learning, pages 1126–1135. PMLR, 2017. 5 Supplementary 5.1 S1 Table 3: Ablation study on the LegNet architecture. MAE is reported in eV , and lower values indicate better performance. Dataset / Task Vanil...

    work page arXiv 2017