Research map / Probabilistic and scientific modeling / Differential equations and PDEs
PhysicsCorrect: A Training-Free Approach for Stable Neural PDE Simulations
Xinquan Huang, Paris Perdikaris
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add PhysicsCorrect residual correction | FNO · 2D Navier-Stokes one-step | relative L2 error: 3.3e-5 -> 5.5e-6 | Improved |
| add PhysicsCorrect residual correction | FNO, UNet, ViT · 2D Navier-Stokes 1000-step rollout | rollout stability | Improved |
| add PhysicsCorrect residual correction | FNO, UNet, ViT · 2D wave equation 100-step rollout | relative L2 error | Improved |
| add PhysicsCorrect with cached Jacobian | FNO · Kuramoto-Sivashinsky 1000-step rollout | relative L2 error | Improved |
| cache Jacobian pseudoinverse instead of recomputing (ablation) | PhysicsCorrect · 2D Navier-Stokes | computational cost: 163x cost reduction, accuracy preserved | Improved |
| recompute Jacobian every 3-10 steps instead of caching once (ablation) | PhysicsCorrect · Kuramoto-Sivashinsky | relative L2 error | No clear effect |
| use semi-implicit discretization for both Jacobian and residual (ablation) | PhysicsCorrect · Kuramoto-Sivashinsky | relative L2 error | Hurt |
| subtract reference residual before correction (idealized) (ablation) | PhysicsCorrect · 2D Navier-Stokes one-step | relative L2 error: 3.3e-5 -> 6.1e-7 (text); corrected 5.5e-6 without it | Improved |
| predict second-order increment u(t+1)+u(t-1)-2u(t) instead of u(t+1)-u(t) (ablation) | neural operators · 2D wave equation | stability | Improved |
| add correction step to inference | FNO · 2D Navier-Stokes 200 steps | inference time (s): 0.69 -> 0.90 | Hurt |
Closest papers by meaning
- P$^2$C$^2$Net: PDE-Preserved Coarse Correction Network for efficient prediction of spatiotemporal dynamics (NeurIPS 2024 · similarity 0.90)
- INC: An Indirect Neural Corrector for Auto-Regressive Hybrid PDE Solvers (NeurIPS 2025 · similarity 0.90)
- Learning a Neural Solver for Parametric PDEs to Enhance Physics-Informed Methods (ICLR 2025 · similarity 0.90)
- Calibrated Physics-Informed Uncertainty Quantification (ICML 2025 · similarity 0.89)
- Compute-optimal data scaling for neural surrogates via multi-fidelity training (NeurIPS 2026 · similarity 0.89)
- Text2PDE: Latent Diffusion Models for Accessible Physics Simulation (ICLR 2025 · similarity 0.89)
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.
