Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Geometric and Physical Constraints Synergistically Enhance Neural PDE Surrogates
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 staggered-grid-compatible equivariant input/output layers and mass/momentum conservation hard constraints for neural PDE surrogates, showing symmetry and physics constraints combine synergistically on shallow water and decaying turbulence tasks.
Problem
Neural PDE surrogates generalize poorly to new initial conditions and accumulate autoregressive errors, and existing symmetry-equivariant layers cannot be applied to staggered CFD grids, so the combined effect of symmetry and physical constraints has not been systematically studied.
Approach
Novel equivariant input and output layers that respect the transformation of staggered (Arakawa C-grid) velocity fields under rotations and reflections, combined with hard mass and momentum conservation constraints (vector-potential curl and mean subtraction) in a modern U-net, evaluated systematically with and without each constraint.
Key results
- On shallow water equations, p4m/M (symmetry + mass constraint) reached NRMSE-zeta 0.032 at 1h and 0.14 at 25h, versus NaN divergence for unconstrained p1/none at 25h.
- On decaying turbulence, p4m/M+rho*u double-constrained U-net achieved NRMSE-u 1.20 at 60s versus 3.32 for unconstrained p1 U-net; with pushforward training it reached 1.06 at 60s.
- On real ocean currents at 120h, p4m/M+rho*u achieved nRMSE 1.10 versus 1.14 for the Equ_rot Unet and 1.66 for p1/none.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add rotation-reflection equivariance and mass conservation | Modern U-Net on closed shallow water · Closed shallow water, 1h | NRMSE-ζ: 0.14 ±0.02 -> 0.032±0.004 | Helped |
| Add mass conservation to p4m equivariance (ablation) | Modern U-Net on closed shallow water · Closed shallow water, 1h | NRMSE-ζ: 0.045±0.005 -> 0.032±0.004 | Helped |
| Add rotation-reflection equivariance to the unconstrained network (ablation) | Modern U-Net on closed shallow water · Closed shallow water, 1h | NRMSE-ζ: 0.14 ±0.02 -> 0.045±0.005 | Helped |
| Add rotation-reflection equivariance and mass conservation | Modern U-Net on decaying turbulence · Decaying turbulence, 60s | NRMSE-u: 3.32 ±0.12 -> 1.20±0.05 | Helped |
| Add mass conservation to p4m equivariance and momentum conservation (ablation) | Modern U-Net on decaying turbulence · Decaying turbulence, 60s | NRMSE-u: 1.40±0.06 -> 1.20±0.05 | Helped |
| Add rotation-reflection equivariance to the unconstrained network (ablation) | Modern U-Net on decaying turbulence · Decaying turbulence, 60s | NRMSE-u: 3.32 ±0.12 -> 1.42±0.05 | Helped |
| Add pushforward training | p4m/M+ρu Modern U-Net · Decaying turbulence, 36s | NRMSE-u: 1.29±0.04 -> 1.07 ±0.01 | Helped |
| Add data augmentation (ablation) | p4m equivariant surrogates · Decaying turbulence | NRMSE-u | No clear effect |
| Add momentum conservation (ablation) | INS surrogates · Decaying turbulence | momentum conservation | Helped |
| Use rotation-reflection equivariance, mass conservation, and momentum conservation | Modern U-Net on real ocean currents · Real ocean currents, 12h | nRMSE: 1.52±0.04 -> 0.88±0.02 | Helped |
| Train p1/∅ with data augmentation | Modern U-Net on decaying turbulence · Decaying turbulence | NRMSE-u | Helped |
Surprising findings
- Pushforward training of unconstrained p4m/none did not help, while pushforward on doubly-constrained surrogates gave the most accurate result overall.
- Data augmentation gave no effect on p4 and p4m surrogates, as expected since they are already equivariant.
Limitations
- Only mass and momentum conservation and square 2D grid symmetries considered
- Constraints benefit mainly for longer rollouts; unclear how constraints limit error accumulation
- Evaluated on two 2D fluid PDEs and one ocean dataset
Experimental setup
Datasets: closed shallow water simulations (semi-implicit Fortran solver) decaying turbulence simulations (jax-cfd) real-world ocean current observations
Benchmarks: closed shallow water system decaying turbulence (INS) real ocean currents
Metrics: NRMSE correlation energy spectrum error mass conservation error total energy error
Baselines: FNO Dilated ResNet (drnet) p1 unconstrained U-net Equ_rot Unet data augmentation pushforward training input noise
Builds on: escnn modern U-net Dilated ResNet pushforward trick Arakawa C-grid
Closest papers by meaning
- Compute-optimal data scaling for neural surrogates via multi-fidelity training (NeurIPS 2026 · similarity 0.90)
- Scaling physics-informed hard constraints with mixture-of-experts (ICLR 2024 · similarity 0.88)
- Active Learning with Selective Time-Step Acquisition for PDEs (ICML 2025 · similarity 0.87)
- Breakeven complexity: A new perspective on neural partial differential equation solvers (NeurIPS 2026 · similarity 0.87)
- CALM-PDE: Continuous and Adaptive Convolutions for Latent Space Modeling of Time-dependent PDEs (NeurIPS 2025 · similarity 0.87)
- P$^2$C$^2$Net: PDE-Preserved Coarse Correction Network for efficient prediction of spatiotemporal dynamics (NeurIPS 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.