Research map / Graph learning / GNN architectures
Learning Physical Dynamics with Subequivariant Graph Neural Networks
Jiaqi Han, Wenbing Huang, Hengbo Ma, Jiachen Li, Joshua B. Tenenbaum, Chuang Gan
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 |
|---|---|---|---|
| Use subequivariance rather than full E(3)-equivariance (ablation) | SGNN · Physion Dominoes, Contain, and Drop | contact prediction accuracy (%): 79.6 -> 89.1; 65.3 -> 78.1; 68.1 -> 74.1 | Improved |
| Add object-aware message passing (ablation) | SGNN · Physion Dominoes, Contain, and Drop | contact prediction accuracy (%): 81.2 -> 89.1; 68.7 -> 78.1; 70.1 -> 74.1 | Improved |
| Add hierarchical modeling (ablation) | SGNN · Physion Dominoes, Contain, and Drop | contact prediction accuracy (%): 83.5 -> 89.1; 72.5 -> 78.1; 71.3 -> 74.1 | Improved |
| Separate inter-object and inner-object edges (ablation) | SGNN · Physion Dominoes, Contain, and Drop | contact prediction accuracy (%): 85.0 -> 89.1; 74.8 -> 78.1; 72.0 -> 74.1 | Improved |
| Train with rotation augmentation instead of relying on equivariance | GNS · Physion Dominoes | contact prediction accuracy (%): 53.3 -> 74.7 | Improved |
| Rotate test data while training on original, directionally biased data | GNS · Physion Dominoes | contact prediction accuracy (%): 78.6 -> 53.3 | Hurt |
| Replace standard EGNN with gravity-aware EGNN-S | EGNN · Physion Dominoes | contact prediction accuracy (%): 61.3 -> 72.0 | Improved |
| Use SGNN instead of the best baseline | Physical dynamics simulators · RigidFall | rollout MSE (× 10⁻²): DPI 1.71 -> SGNN 0.72 (|Train|=200, t=20); DPI 5.37 -> SGNN 2.44 (|Train|=200, t=40) | Improved |
| Use SGNN instead of GNS | Physical dynamics simulators · Physion | contact prediction accuracy (%): 78.6 -> 89.1 (Dominoes, GNS*); 70.7 -> 78.1 (Contain, GNS); 58.0 -> 73.3 (Link, GNS) | Improved |
Closest papers by meaning
- Equivariant Graph Mechanics Networks with Constraints (ICLR 2022 · similarity 0.91)
- Physics-Inspired All-Pair Interaction Learning for 3D Dynamics Modeling (ICLR 2026 · similarity 0.89)
- SEGNO: Generalizing Equivariant Graph Neural Networks with Physical Inductive Biases (ICLR 2024 · similarity 0.88)
- Universally Invariant Learning in Equivariant GNNs (NeurIPS 2025 · similarity 0.88)
- Equivariant Graph Neural Operator for Modeling 3D Dynamics (ICML 2024 · similarity 0.87)
- Large Language-Geometry Model: When LLM meets Equivariance (ICML 2025 · similarity 0.86)
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.
