Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Solver-in-the-Loop: Learning from Differentiable Physics to Interact with Iterative PDE-Solvers
Kiwon Um, Robert Brand, Yun (Raymond) Fei, Philipp Holl, Nils Thuerey
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 |
|---|---|---|---|
| train corrector with differentiable solver in the loop (SOL32) instead of precomputed corrections (PRE) | CNN corrector · Wake Flow | MAE velocity: 0.031 -> 0.013 | Improved |
| train with solver in the loop (SOL128) instead of non-interacting supervised (NON) | CNN corrector · Buoyancy | MAE velocity: 1.080 -> 0.620 | Improved |
| solver in the loop vs NON | CNN corrector · Forced advection-diffusion | MAE velocity: 0.159 -> 0.158 | No clear effect |
| train CG initial-guess net through differentiable CG solver instead of physics-based residual loss | CNN initial guess · CG Solver Poisson | iterations to 0.001: 79.03 -> 29.59 | Improved |
| solver in the loop vs NON | CNN corrector · 3D Wake | MAE velocity: 0.144 -> 0.130 | Improved |
| increase look-ahead horizon (SOLs -> SOL) (ablation) | SOL · Wake Flow | MAE velocity: 0.041 -> 0.013 | Improved |
| increase look-ahead horizon to 128 steps (ablation) | SOL · Buoyancy | relative improvement: up to 40% (1-4 steps) -> 60% (128) | Improved |
| increase look-ahead horizon (ablation) | SOL · Adv.-diff. | MAE velocity: 0.152 -> 0.158 | No clear effect |
| perturb network inputs with Gaussian noise during training (ablation) | corrector · Buoyancy | relative improvement: 30.6% -> 34.5% | No clear effect |
| add learned correction to coarse solver | coarse PDE solver · 3D Wake | runtime vs reference: 913.2s -> 13.3s | Improved |
Closest papers by meaning
- Learning a Neural Solver for Parametric PDEs to Enhance Physics-Informed Methods (ICLR 2025 · similarity 0.92)
- PRDP: Progressively Refined Differentiable Physics (ICLR 2025 · similarity 0.90)
- Mosaic: A Benchmark Suite for Differentiable Physics Solvers (NeurIPS 2026 · similarity 0.89)
- Physics-informed Reduced Order Modeling of Time-dependent PDEs via Differentiable Solvers (NeurIPS 2025 · similarity 0.89)
- Active Learning for Neural PDE Solvers (ICLR 2025 · similarity 0.89)
- INC: An Indirect Neural Corrector for Auto-Regressive Hybrid PDE Solvers (NeurIPS 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.
