Research map / Graph learning / GNN architectures
PhAST: Physics-Aware, Scalable, and Task-Specific GNNs for Accelerated Catalyst Design
Alexandre Duval, Victor Schmidt, Santiago Miret, Yoshua Bengio, Alex Hernández-García, David Rolnick
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add all PhAST components (graph pruning, rich embeddings, weighted-sum energy head) | SchNet · OC20 IS2RE | average energy MAE: 683 -> 630 | Improved |
| add all PhAST components | SchNet · OC20 IS2RE | inference time (s): 15 -> 5 | Improved |
| add all PhAST components | GemNet-OC · OC20 IS2RE | inference time (s): 868 -> 257 | Improved |
| add tag embeddings to atom embeddings (ablation) | SchNet · OC20 IS2RE | average energy MAE: 683 -> 648 | Improved |
| remove tag-0 sub-surface atoms from the graph (ablation) | SchNet · OC20 IS2RE | inference time (s): 15.40 -> 4.74 | Improved |
| remove tag-0 sub-surface atoms from the graph (ablation) | DimeNet++ · OC20 IS2RE | inference time (s): 110.11 -> 20.15 | Improved |
| remove tag-0 sub-surface atoms from the graph (ablation) | DimeNet++ · OC20 IS2RE | average energy MAE: 628 -> 610 | Improved |
| add one supernode per graph instead of removing tag-0 (ablation) | SchNet · OC20 IS2RE | average energy MAE: 648 -> 654 | No clear effect |
| add hierarchical HoscPool energy head (ablation) | SchNet · OC20 IS2RE | inference time (s): 15.40 -> 53.93 | Hurt |
| add weighted-sum energy head from final embeddings (ablation) | ForceNet · OC20 IS2RE | average energy MAE: 654 -> 660 | Hurt |
| add weighted-sum energy head from initial embeddings (ablation) | SchNet · OC20 IS2RE | average energy MAE: 683 -> 657 | Improved |
| replace gradient-based forces with direct force head (PhAST-Direct) (ablation) | SchNet · OC20 S2EF-2M | energy MAE vs force MAE: E-MAE 890 -> 716; F-MAE 68.4 -> 81.2 | Mixed |
Closest papers by meaning
- The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains (NeurIPS 2024 · similarity 0.80)
- DR-Label: Label Deconstruction and Reconstruction of GNN Models for Catalysis Systems (AAAI 2024 · similarity 0.79)
- Learning Condensed Graph via Differentiable Atom Mapping for Reaction Yield Prediction (ICML 2025 · similarity 0.78)
- Transolver++: An Accurate Neural Solver for PDEs on Million-Scale Geometries (ICML 2025 · similarity 0.78)
- ELoRA: Low-Rank Adaptation for Equivariant GNNs (ICML 2025 · similarity 0.78)
- PAPM: A Physics-aware Proxy Model for Process Systems (ICML 2024 · similarity 0.78)
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.
