Research map / Graph learning / Graph self-supervised learning
DR-Label: Label Deconstruction and Reconstruction of GNN Models for Catalysis Systems
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DR-Label, which deconstructs node-level equilibrium displacement labels into unique edge-level projections and reconstructs node predictions via sphere fitting, and builds DRFormer on it for adsorption energy prediction on OC20 and SAA.
Problem
Predicting equilibrium geometry and adsorption energy of catalyst-adsorbate systems with GNNs, where linearly aggregated edge-to-node geometry predictions suffer from edge representation ambiguity and sensitivity to graph structure variations.
Approach
Project node-level displacement labels onto each outward edge direction to get unique edge-level supervision, then reconstruct node displacements from predicted edge projections with a sphere-fitting algorithm (project-and-fit), which is E(3)-equivariant and graph-structure invariant.
Key results
- DR-Label consistently improves 3d-Graphormer, GemNet-OC and SCN; e.g., GemNet-OC AEwT on ID rises from 4.15% to 6.55% on the OC20 test set.
- DRFormer-S-FT reaches 69 meV MAE and 56.0% AEwT on the SAA dataset, a 20.2% relative MAE decrease over MT-MD DimeNet++ (87 meV).
- DRFormer-Average achieves a new state-of-the-art direct-method OC20 IS2RE test average MAE of 0.4340 eV and AEwT of 6.71%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DR-Label to 3d-Graphormer (ablation) | 3d-Graphormer · OC20 IS2RE 10k validation, Average | MAE (meV): 695 -> 689 | Helped |
| add noisy nodes to 3d-Graphormer (ablation) | 3d-Graphormer · OC20 IS2RE 10k validation, Average | MAE (meV): 695 -> 675 | Helped |
| add intermediate positional updates to 3d-Graphormer without DR-Label (ablation) | 3d-Graphormer · OC20 IS2RE 10k validation, Average | MAE (meV): 695 -> 731 | Hurt |
| add DR-Label and noisy nodes (ablation) | 3d-Graphormer · OC20 IS2RE 10k validation, Average | MAE (meV): 695 -> 662 | Helped |
| add noisy nodes and intermediate positional updates without DR-Label (ablation) | 3d-Graphormer · OC20 IS2RE 10k validation, Average | MAE (meV): 695 -> 721 | Hurt |
| add DR-Label, noisy nodes, and intermediate positional updates (ablation) | 3d-Graphormer · OC20 IS2RE 10k validation, Average | MAE (meV): 695 -> 650 | Helped |
| add DR-Label to 3d-Graphormer | 3d-Graphormer · OC20 IS2RE validation, ID | MAE (eV): 0.4329 -> 0.4574 | Hurt |
| add DR-Label to GemNet-OC | GemNet-OC · OC20 IS2RE testing, Average | MAE (eV): 0.630 -> 0.568 | Helped |
| add DR-Label to SCN | SCN · OC20 IS2RE testing, Average | MAE (eV): 0.573 -> 0.564 | Helped |
| add intermediate positional updates to DRFormer (ablation) | DRFormer · OC20 IS2RE validation, Average | MAE (eV): 0.4539 -> 0.4386 | Helped |
| fine-tune an OC20-trained DRFormer-S checkpoint on SAA | DRFormer-S · SAA test | MAE (meV): 83 ± 14 -> 69 ± 15 | Helped |
Surprising findings
- Adding intermediate positional updates alone worsened 3d-Graphormer performance (MAE 731 vs 695 meV in ablation); only combined with DR-Label did they help.
Limitations
- Edge-level projection labels and sphere fitting add model complexity and training cost relative to plain node prediction.
- On some OC20 metrics (e.g., OOD-Cat MAE for GemNet-OC) DR-Label gives little or no gain, and DRFormer does not top every metric.
Experimental setup
Datasets: OC20 SAA
Benchmarks: OC20 IS2RE SAA CO adsorption dataset
Metrics: MAE (eV) AEwT (%) MAE (meV)
Baselines: 3d-Graphormer GNS+Noisy Nodes Equiformer+Noisy Nodes Moleformer GemNet-OC SCN MT-MD DimeNet++ DimeNet++
Builds on: Graphormer Noisy Nodes GemNet-OC SCN Equiformer
Closest papers by meaning
- Learning Condensed Graph via Differentiable Atom Mapping for Reaction Yield Prediction (ICML 2025 · similarity 0.85)
- AdsorbDiff: Adsorbate Placement via Conditional Denoising Diffusion (ICML 2024 · similarity 0.84)
- Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling, and Zero-Shot Transfer (ICLR 2025 · similarity 0.84)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.83)
- On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.83)
- Reaction Graph: Towards Reaction-Level Modeling for Chemical Reactions with 3D Structures (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.