Research map / Probabilistic and scientific modeling / Differential equations and PDEs
INC: An Indirect Neural Corrector for Auto-Regressive Hybrid PDE Solvers
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 Indirect Neural Corrector (INC), which injects learned corrections as a right-hand-side term inside the PDE time integration instead of as direct state updates, supported by an error-amplification analysis for autoregressive hybrid solvers.
Problem
Hybrid solvers that combine coarse numerical solvers with learned direct state corrections accumulate amplified perturbations during long autoregressive rollouts, especially in chaotic regimes, causing instability and error growth.
Approach
Move the learned correction G_theta into the governing equation's right-hand side so it is integrated by the solver's time stepping (indirect correction), rather than added to the output state; theory claims the error-dominance ratio between direct and indirect perturbations scales as 1/dt + L.
Key results
- INC improves long-term trajectory R^2 by up to 158.7% over SITL for a UNet on chaotic KS.
- On KS2 with first-order RK at dt=0.5, INC reaches R^2 0.78 versus 0.65 for SITL, a 20.0% improvement.
- For 3D turbulent channel flow, INC reaches a 330x speed-up over the No Model 256 reference at equivalent statistical accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use indirect neural correction instead of SITL direct correction | Hybrid solver on KS2 · KS2 | R²: 0.65 -> 0.78; 20.0% improvement | Helped |
| Use indirect neural correction instead of SITL* direct pre-correction | Hybrid solver on KS2 · KS2 | MSE: 36.9% decrease | Helped |
| Use indirect neural correction instead of SITL direct correction | Hybrid solver with oversized time steps · CFL=2.0 | R²: 0.74 -> 0.80 | Helped |
| Use indirect neural correction instead of direct correction | Hybrid solver with oversized time steps · CFL=2.0 | MSE: 26.9% | Helped |
| Use indirect neural correction instead of SITL* direct pre-correction | Hybrid solver with oversized time steps · CFL=2.0 | MSE: 36.5% | Helped |
| Use indirect neural correction instead of SITL | Hybrid solver on BFS · Backward facing step (BFS) | MSE: 97.1% | Helped |
| Use indirect neural correction instead of SITL | Hybrid solver on 3D TCF · 3D turbulent channel flow (TCF) | MSE on statistics: 80.7% | Helped |
| Use INC rather than the fine-resolution No Model baseline | 3D turbulent channel flow at equivalent statistical accuracy · 3D turbulent channel flow (TCF) | speedup: 330× | Helped |
| Use INC rather than the high-resolution No Model simulation | Hybrid solver on BFS · Backward facing step (BFS) | speedup: 7× | Helped |
| Inject perturbations indirectly rather than directly into the solution | Hybrid solver on Burgers · 1D Burgers | stability | Helped |
| Inject perturbations indirectly rather than directly into the solution | Hybrid solver on KS · 1D KS | error growth | Helped |
Surprising findings
- For FNO, INC improves only about 3% over direct corrections, which the authors attribute to FNO's similarity to the pseudo-spectral solver.
- The RNN baseline without solver consistently underperforms hybrid solvers on KS and Burgers.
Limitations
- Advantage depends on dynamical stiffness and the time integrator's stability region; benefits may diminish for well-damped problems or unconditionally stable implicit schemes.
- Theoretical analysis relies on linearized error dynamics and small-perturbation assumptions.
- Requires differentiable solver and unrolled training, increasing memory cost.
Experimental setup
Datasets: high-resolution simulation data for Kuramoto-Sivashinsky Burgers equation simulations Karman vortex shedding backward facing step turbulent channel flow (3D)
Benchmarks: KS equation Burgers equation Karman vortex shedding backward facing step turbulent channel flow
Metrics: R^2 MSE speed-up CFL stability
Baselines: SITL SITL* CSM RNN No Model
Builds on: solver-in-the-loop CoSTA dresdner2022learning Fourier neural operator ETDRK
Closest papers by meaning
- Breakeven complexity: A new perspective on neural partial differential equation solvers (NeurIPS 2026 · similarity 0.90)
- MC^2: Monte Carlo Correction for Fast Elliptic PDE Solving (NeurIPS 2026 · similarity 0.89)
- Hierarchical Implicit Neural Emulators (NeurIPS 2025 · similarity 0.89)
- APEBench: A Benchmark for Autoregressive Neural Emulators of PDEs (NeurIPS 2024 · similarity 0.88)
- P$^2$C$^2$Net: PDE-Preserved Coarse Correction Network for efficient prediction of spatiotemporal dynamics (NeurIPS 2024 · similarity 0.88)
- UGrid: An Efficient-And-Rigorous Neural Multigrid Solver for Linear PDEs (ICML 2024 · 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.