Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Context-Informed Neural ODEs Unexpectedly Identify Broken Symmetries: Insights from the Poincaré–Hopf Theorem
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.
Shows that context-informed neural ODEs trained only on pre-bifurcation symmetric data can zero-shot recover post-bifurcation broken symmetries (and sometimes hallucinate them), explains this via the Poincaré–Hopf theorem, and proposes a Poincaré–Hopf topological regularizer validated on Landau–Khalatnikov phase transitions.
Problem
Out-of-domain generalization of learned dynamical systems across bifurcations (symmetry-breaking phase transitions), where models trained only on pre-bifurcation data must predict post-bifurcation behavior.
Approach
Context-informed NODEs (CoDA-style hypernetwork with per-environment context vectors) implicitly learn the Poincaré index of the vector field; the Poincaré–Hopf theorem constrains them to create extra fixed points after a center-to-saddle transition, enabling zero-shot recovery of broken symmetry. A global/local Poincaré-index loss is added as a regularizer.
Key results
- On the double-well system, context-informed NODEs adapted with one post-bifurcation trajectory reduced MSE to about 0.15e-3 to 0.25e-3 versus about 1.7e-3 to 3.9e-3 for vanilla NODEs.
- Zero-shot linear extrapolation of the context vector reconstructs the double-well bifurcation diagram despite pre-bifurcation-only training.
- The Poincaré–Hopf regularizer reduces the global index mismatch and improves long-term topological consistency over the vanilla context-informed NODE on the Landau–Khalatnikov system (qualitative heatmaps; no single headline number given).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a context-informed NODE instead of a vanilla NODE | Hamiltonian double-well dynamics · Left well | MSE (×10^-3): 2427 ±808.8 -> 0.1517 ±0.068 | Helped |
| Use a context-informed NODE instead of a vanilla NODE | Hamiltonian double-well dynamics · Right well | MSE (×10^-3): 1724 ±365.1 -> 0.2504 ± 0.033 | Helped |
| Use a context-informed NODE instead of a vanilla NODE | Hamiltonian double-well dynamics · Outer orbit | MSE (×10^-3): 3943 ±904.8 -> 0.1457 ±0.065 | Helped |
| Train on pre-bifurcation data and extrapolate the context without post-bifurcation adaptation | Context-informed NODE · Hamiltonian double-well, post-bifurcation | Phase-topology reconstruction | Helped |
| Train on localized pre-bifurcation trajectories | Context-informed NODE · Hamiltonian double-well, post-bifurcation | Phase-topology reconstruction | Helped |
| Use a context-informed NODE on a linear center-to-saddle system | Linear system (q̇, ṗ) = (p, µq) · Post-bifurcation | Bifurcation identification | Hurt |
| Add Poincaré–Hopf regularization | Context-informed NODE on the Landau–Khalatnikov system · Landau–Khalatnikov parameter plane | Long-term topological consistency | Helped |
| Add only the global Poincaré–Hopf constraint instead of the full global and local regularizer (ablation) | Context-informed NODE on the Landau–Khalatnikov system · Landau–Khalatnikov parameter plane | Long-term topological consistency | Helped |
| Remove the local Poincaré–Hopf constraint while retaining the global constraint (ablation) | Fully regularized context-informed NODE on the Landau–Khalatnikov system · Landau–Khalatnikov parameter plane | Long-term topological consistency | Hurt |
Surprising findings
- Context-informed NODEs trained only on a linear oscillator (no double-well) spuriously produce a double-well structure, i.e., they hallucinate broken symmetry.
- Vanilla NODEs fail to generalize across the bifurcation even with more post-bifurcation samples confined to one well.
Limitations
- Current method is limited to 2-dimensional flows with nearly closed orbits.
- Local regularizer requires system-specific prior knowledge of how the local index changes with parameters; global-only variant performs slightly worse.
- Model can hallucinate spurious bifurcations, and the proposed variance-based diagnostic is only a heuristic.
Experimental setup
Datasets: synthetic trajectories from Hamiltonian double-well system linear oscillator system cusp catastrophe system Landau–Khalatnikov system
Benchmarks: Hamiltonian double-well bifurcation Landau–Khalatnikov long-term consistency
Metrics: MSE topological consistency score MAPE Poincaré index mismatch bootstrapped ensemble variance
Baselines: vanilla NODE Hamiltonian context-informed NODE
Builds on: CoDA Neural ODE Hamiltonian neural network neural adjoint method Poincaré–Hopf theorem
Closest papers by meaning
- Discovering Physics Laws of Dynamical Systems via Invariant Function Learning (ICML 2025 · similarity 0.87)
- Let's do the time-warp-attend: Learning topological invariants of dynamical systems (ICLR 2024 · similarity 0.87)
- Symmetries in Overparametrized Neural Networks: A Mean Field View (NeurIPS 2024 · similarity 0.86)
- Neural Conjugate Flows: A Physics-Informed Architecture with Flow Structure (AAAI 2025 · similarity 0.86)
- Symmetry-Informed Governing Equation Discovery (NeurIPS 2024 · similarity 0.86)
- Out-of-Domain Generalization in Dynamical Systems Reconstruction (ICML 2024 · 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.