Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Towards Universal Mesh Movement Networks
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 UM2N, a graph transformer encoder and GAT decoder trained on a PDE-independent dataset of random Gaussian fields to zero-shot move meshes using a Monge-Ampère-derived target and element volume loss.
Problem
Mesh movement (r-adaptation) for PDE solvers is expensive with Monge-Ampère methods and existing learned mesh movement methods need retraining per PDE or geometry and can produce tangled (inverted) elements.
Approach
A two-stage graph transformer encoder plus GAT decoder that takes only monitor-function values as input, trained on PDE-independent random Gaussian fields with an element volume loss that penalizes inverted elements, enabling zero-shot transfer across PDEs and boundary geometries.
Key results
- On Swirl, UM2N achieves 35.93% error reduction versus 14.15% for M2N and 37.43% for MA, at 17.66 ms versus 5615.78 ms for MA.
- On flow past cylinder, UM2N achieves 61.29% average drag/lift error reduction while MA and M2N fail due to mesh tangling.
- Performance benchmark reports roughly 2347x speedup over MA on RTX 3090 and 3342x on A100 (about 102x on CPU).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UM2N instead of M2N | Mesh movement · Swirl | ER (%): 14.15 -> 35.93 | Helped |
| Use UM2N instead of MA | Mesh movement · Swirl | Time (ms): 5615.78 -> 17.66 | Helped |
| Use UM2N instead of M2N | Mesh movement · Helmholtz | ER (%): 16.32 -> 17.08 | Helped |
| Use UM2N instead of MA | Mesh movement · Helmholtz | Time (ms): 3265.23 -> 4.86 | Helped |
| Use UM2N instead of MA or M2N | Mesh movement · Cylinder | ER (%): Fail -> 61.29 | Helped |
| Use UM2N instead of M2N | Mesh movement · Flow past cylinder (simplified laminar) | Vorticity ER (%): 10.22 -> 27.43 | Helped |
| Add element volume loss instead of coordinate loss (ablation) | UM2N · Swirl | ER (%): 31.21 -> 35.93 | Helped |
| Add element volume loss instead of coordinate loss (ablation) | UM2N · Cylinder | Mesh tangling: Fail -> 61.29 ER (%) | Helped |
| Use monitor values instead of PDE solution as input (ablation) | UM2N · Swirl | ER (%): -0.35 -> 35.93 | Helped |
| Use monitor values instead of PDE solution as input (ablation) | UM2N · Helmholtz | ER (%): 13.52 -> 17.08 | Helped |
| Use monitor values instead of PDE solution as input (ablation) | UM2N · Cylinder | ER (%): 39.94 -> 61.29 | Helped |
| Use UM2N instead of MA on CPU | Mesh movement inference · Swirl benchmark (11,833 vertices) | Speedup: ~102× | Helped |
Surprising findings
- Using PDE solution as input instead of monitor values yields -0.35% error reduction on Swirl, far below the monitor-based model.
- The full Monge-Ampère method fails on the Tohoku tsunami complex coastline boundary, while UM2N avoids tangling.
Limitations
- No theoretical guarantee that mesh tangling is fully prevented.
- GAT constrains movement to within one hop, limiting large deformations in a single step.
- Best performance depends on a manually chosen monitor function; cannot increase vertex count for coarse meshes.
Experimental setup
Datasets: randomly generated Gaussian-sum generic fields (PDE-independent)
Benchmarks: Swirl advection Cylinder flow (Navier-Stokes) Helmholtz Tohoku tsunami
Metrics: PDE error reduction ratio (%) mesh tangling (fail) inference time (ms)
Baselines: MA (Movement relaxation) M2N UM2N-coord UM2N-sol
Builds on: Monge-Ampère mesh movement graph attention network M2N
Closest papers by meaning
- UGM2N: An Unsupervised and Generalizable Mesh Movement Network via M-Uniform Loss (NeurIPS 2025 · similarity 0.92)
- Better Neural PDE Solvers Through Data-Free Mesh Movers (ICLR 2024 · similarity 0.89)
- PhyMPGN: Physics-encoded Message Passing Graph Network for spatiotemporal PDE systems (ICLR 2025 · similarity 0.88)
- MultiPDENet: PDE-embedded Learning with Multi-time-stepping for Accelerated Flow Simulation (ICML 2025 · similarity 0.86)
- PIORF: Physics-Informed Ollivier-Ricci Flow for Long–Range Interactions in Mesh Graph Neural Networks (ICLR 2025 · similarity 0.86)
- MC^2: Monte Carlo Correction for Fast Elliptic PDE Solving (NeurIPS 2026 · 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.