Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Neural Conjugate Flows: A Physics-Informed Architecture with Flow Structure
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.
Introduces Neural Conjugate Flows, which wrap an affine flow between invertible coupling-layer networks via topological conjugation, giving exact flow structure and faster training than Neural ODEs on physics-informed ODE tasks.
Problem
PINNs based on generic MLPs do not enforce flow properties (identity, associativity, invertibility), leading to causality and uniqueness failures, while Neural ODEs that do have flow structure are slow due to sequential numerical solvers.
Approach
Model the flow as a topological conjugate H^{-1} ∘ Ψ^t ∘ H, where H is an invertible coupling-layer network (with input duplication) and Ψ is a closed-form affine flow whose matrix can be projected onto a Lie algebra (e.g., skew-symmetric) to enforce topology; this gives exact group structure and parallel-computable flows.
Key results
- On FitzHugh-Nagumo, NCF-T reaches accuracy loss 3.2e-4 and extrapolation loss 1.9e-3 versus MLP's 2.8e+1 extrapolation loss, but trains in 75.8 s versus MLP's 35.7 s.
- On Hodgkin-Huxley, NCF-T achieves extrapolation loss 1.6e-3 vs NODE's 3.8e-2, training in 45.6 s vs NODE's 241.1 s.
- The paper reports NCFs are roughly twice as slow as MLPs and about five times faster than Neural ODEs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| enforce topology in NCF (ablation) | NCF · FitzHugh-Nagumo | extrapolation loss: 2.5\tento{-1} \pm 3.0\tento{-2} -> 1.9\tento{-3} \pm 4.6\tento{-4} | Helped |
| enforce topology in NCF (ablation) | NCF · Hodgkin-Huxley | extrapolation loss: 3.2\tento{-2} \pm 3.2\tento{-1} -> 1.6\tento{-3} \pm 9.8\tento{-4} | Helped |
| enforce topology in NCF (ablation) | NCF · Lotka-Volterra | extrapolation loss: 3.6\tento{-1} \pm 4.8\tento{-1} -> 1.1\tento{-1} \pm 1.8\tento{-1} | Helped |
| enforce topology in NCF (ablation) | NCF · FitzHugh-Nagumo, near-equilibrium initial condition | extrapolation loss: 1.7\tento{0} \pm 1.9\tento{-2} -> 2.3\tento{0} \pm 8.9\tento{-2} | Hurt |
| use NCF instead of MLP | FitzHugh-Nagumo · FitzHugh-Nagumo | extrapolation loss: 2.8\tento{1} \pm 5.4\tento{0} -> 2.5\tento{-1} \pm 3.0\tento{-2} | Helped |
| use NCF instead of NODE | Hodgkin-Huxley · Hodgkin-Huxley | training time (s): 241.1\pm11.2 vs 44.4\pm1.9 | Helped |
| use NCF instead of NODE | Hodgkin-Huxley · Hodgkin-Huxley | extrapolation loss: 3.8\tento{-2} \pm 9.1\tento{-3} -> 3.2\tento{-2} \pm 3.2\tento{-1} | Helped |
| use NCF instead of MLP | Lotka-Volterra · Lotka-Volterra | accuracy loss: 1.1\tento{0} \pm 1.1\tento{-1} -> 9.2\tento{-2} \pm 2.3\tento{-2} | Helped |
| use NCF instead of MLP | FitzHugh-Nagumo, near-equilibrium initial condition · FitzHugh-Nagumo, near-equilibrium initial condition | accuracy loss: 4.4\tento{-5} \pm 2.1\tento{-5} -> 1.4\tento{0} \pm 8.8\tento{-3} | Hurt |
Surprising findings
- On the Lotka-Volterra causality test, vanilla PINNs failed entirely to approximate the solution while NCFs converged to correct solutions.
- Affine NCFs fail to converge on a nonlinear case where the orbit transitions from spiral growth to a limit cycle, a stated limitation of affine conjugates.
Limitations
- Expressiveness limited by affine central flow; poor on strongly nonlinear behavior such as limit cycles and chaos.
- Coupling layers reduce representation power; invertible-network-based conjugation is slower than MLPs.
- Experiments are small-scale ODE tasks on a single machine.
Experimental setup
Datasets: FitzHugh-Nagumo simulation (synthetic) Hodgkin-Huxley simulation (synthetic, noisy) Lotka-Volterra simulation
Benchmarks: FitzHugh-Nagumo Hodgkin-Huxley Lotka-Volterra
Metrics: mean-squared error (accuracy) extrapolation MSE training time (s)
Baselines: MLP-PINN Neural ODE (NODE) NCF (vanilla)
Builds on: Physics-informed neural networks Neural ODEs Coupling layers (RealNVP-style) Gaussian Fourier features ANODEs
Closest papers by meaning
- Identifying Equivalent Training Dynamics (NeurIPS 2024 · similarity 0.87)
- Neural Context Flows for Meta-Learning of Dynamical Systems (ICLR 2025 · similarity 0.87)
- Simulation-Free Training of Neural ODEs on Paired Data (NeurIPS 2024 · similarity 0.87)
- ControlSynth Neural ODEs: Modeling Dynamical Systems with Guaranteed Convergence (NeurIPS 2024 · similarity 0.86)
- How Deep Do We Need: Accelerating Training and Inference of Neural ODEs via Control Perspective (ICML 2024 · similarity 0.86)
- Switched Flow Matching: Eliminating Singularities via Switching ODEs (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.