Research map / Graph learning / GNN architectures
PhyMPGN: Physics-encoded Message Passing Graph Network for spatiotemporal PDE 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 PhyMPGN, a message-passing GNN inside an RK2 numerical integrator with a learnable Laplace-Beltrami block and a boundary-condition padding strategy, to predict spatiotemporal PDE dynamics on coarse irregular meshes from small training sets.
Problem
Predicting spatiotemporal PDE dynamics on irregular, coarse unstructured meshes with complex boundary conditions when only small training datasets are available; purely data-driven neural models need rich data and extrapolate poorly.
Approach
Combine a GNN (encode-process-decode) with a second-order Runge-Kutta time integrator, add a learnable Laplace block that corrects the discrete Laplace-Beltrami operator for the known diffusion term, and enforce Dirichlet/Neumann/Robin/periodic BCs via ghost-node padding in physical and latent space.
Key results
- On Burgers, FN and GS, PhyMPGN reaches MSE 2.99e-4, 2.82e-3 and 4.05e-4, versus best baselines MGN 9.19e-4, MP-PDE 2.88e-2 and MGN 4.44e-3, with leads of 67.5%, 90.2% and 90.9%.
- On cylinder flow at Re 600-1000, PhyMPGN achieves MSE 1.25e-1 versus 5.96e-1 for the best baseline MP-PDE (79.0% lead).
- On the Laplacian approximation task, the Laplace block (5.6k parameters) reaches MSE 151 and RNE 0.033, outperforming SDL-padding (MSE 408) and Mesh Laplace (MSE 2574).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PhyMPGN instead of the strongest listed baseline | PDE prediction · Burgers, FN, and GS | MSE: Burgers: 9.19e-4 -> 2.99e-4; FN: 2.88e-2 -> 2.82e-3; GS: 4.44e-3 -> 4.05e-4 | Helped |
| use PhyMPGN instead of the strongest listed baseline | Cylinder flow prediction · Re [200, 280, 360], [440, 480, 520], and [600, 800, 1000] | MSE: 7.23e-3 -> 2.14e-4; 2.17e-1 -> 3.56e-2; 5.96e-1 -> 1.25e-1 | Helped |
| use PhyMPGN instead of MGN | Burgers with varied boundary conditions · Scenario A, B, C, and D | MSE: Scenario A: 9.19e-4 -> 2.99e-4; Scenario B: 1.68e-3 -> 2.89e-4; Scenario C: 1.89e-3 -> 3.06e-4; Scenario D: 2.70e-3 -> 9 | Helped |
| use the learnable Laplace block instead of Mesh Laplace | coarse-mesh Laplacian approximation · synthetic coarse non-uniform mesh | MSE: 2574 -> 151 | Helped |
| add BC padding | SDL · synthetic coarse non-uniform mesh | MSE: 1210 -> 408 | Helped |
| increase the number of training trajectories | PhyMPGN · Burgers | MSE: 6.35e-4 -> 2.99e-4 -> 1.60e-4 -> 4.43e-5 | Helped |
| remove the learnable Laplace block (ablation) | PhyMPGN · Cylinder flow, Re [440, 480, 520] | MSE | Hurt |
| add the learnable Laplace block (ablation) | PhyMPGN · Cylinder flow, Re [440, 480, 520] | MSE | Helped |
| remove BC padding (ablation) | PhyMPGN · Cylinder flow, Re [440, 480, 520] | MSE | Hurt |
| add BC padding (ablation) | PhyMPGN · Cylinder flow, Re [440, 480, 520] | MSE | Helped |
| remove both the learnable Laplace block and BC padding (ablation) | PhyMPGN · Cylinder flow, Re [440, 480, 520] | MSE | Hurt |
| use a higher-order numerical integrator (ablation) | PhyMPGN · PDE prediction | MSE | Helped |
Surprising findings
- Mesh Laplace (the pure discrete Laplace-Beltrami operator without learned correction) performs worse than the learnable baselines on coarse meshes, motivating the learned correction.
- Even with the Euler forward integrator, PhyMPGN remains well ahead of baselines, and the padding strategy adds only about 0.7 s (10.71 vs 10.02 s) inference time.
Limitations
- Neumann/Robin BCs cannot yet be exactly enforced in latent space.
- Method does not handle nonlinear or high-order diffusion terms (e.g., fourth-order Laplacian) and assumes the diffusion term is known a priori.
- Limited to 2D; extension to 3D is left to future work; correlation decays for Re=1000 and over longer extrapolation horizons.
Experimental setup
Datasets: Burgers' equation (COMSOL-generated) FitzHugh-Nagumo (COMSOL-generated) Gray-Scott (COMSOL-generated) Cylinder flow Navier-Stokes (COMSOL-generated) Synthetic random periodic functions for Laplace block
Benchmarks: Burgers' equation FitzHugh-Nagumo Gray-Scott cylinder flow
Metrics: MSE relative L2 error Pearson correlation
Baselines: DeepONet PA-DGN MGN MP-PDE SDL PENN
Builds on: PeRCNN MeshGraphNets MP-PDE PA-DGN DeepONet
Closest papers by meaning
- P$^2$C$^2$Net: PDE-Preserved Coarse Correction Network for efficient prediction of spatiotemporal dynamics (NeurIPS 2024 · similarity 0.92)
- Graph Neural PDE Solvers with Conservation and Similarity-Equivariance (ICML 2024 · similarity 0.90)
- Mechanistic PDE Networks for Discovery of Governing Equations (ICML 2025 · similarity 0.89)
- Text2PDE: Latent Diffusion Models for Accessible Physics Simulation (ICLR 2025 · similarity 0.89)
- MultiPDENet: PDE-embedded Learning with Multi-time-stepping for Accelerated Flow Simulation (ICML 2025 · similarity 0.89)
- PIORF: Physics-Informed Ollivier-Ricci Flow for Long–Range Interactions in Mesh Graph Neural Networks (ICLR 2025 · similarity 0.88)
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.