Research map / Architectures and efficiency / RNNs and normalization
Test-Time Accuracy-Cost Control in Neural Simulators via Recurrent-Depth
Harris Abdul Majid, Pietro Sittoni, Francesco Tudisco
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 recurrent-depth (RecurrFNO 0.8B, 100 epochs) in place of a 1.6B FNO | FNO · 3D Compressible Navier-Stokes (Density) | MSE x10^-2: 7.61 -> 7.37 | Improved |
| use RecurrFNO 0.8B trained for 82 epochs against 1.6B FNO | FNO · 3D Compressible Navier-Stokes (Pressure) | MSE x10^-2: 7.59 -> 7.51 | Improved |
| replace ViT with RecurrViT using 58% of the parameters | ViT · Active Matter (Steps 0:3) | MSE x10^-2: 2.91 -> 0.54 | Improved |
| replace ViT with RecurrViT | ViT · Active Matter (Steps 0:12) | MSE x10^-2: 43.16 -> 5.68 | Improved |
| replace UPT with RecurrUPT (92M vs 164M parameters) | UPT · ShapeNet-Car (32^3) | MSE x10^-2: 2.31 -> 2.19 | Improved |
| increase backpropagation window from B=1 to B=2 (ablation) | RecurrSim · Burgers (long-horizon, ~1M params) | trajectory error (L2): 0.0849 -> 0.0315 | Improved |
| increase backpropagation window from B=2 to B=4 (ablation) | RecurrSim · Korteweg-De Vries (long-horizon, ~1M params) | trajectory error (L2): 0.0522 -> 0.0317 | Improved |
| increase backpropagation window from B=4 to B=32 (ablation) | RecurrSim · Burgers (long-horizon, ~1M params) | trajectory error (L2): 0.0199 -> 0.0178 | No clear effect |
| increase backpropagation window from B=4 to B=32 (ablation) | RecurrSim · Kuramoto-Sivashinsky (long-horizon, ~1M params) | trajectory error (L2): 1.3972 -> 1.3910 | No clear effect |
| increase mean training recurrent iterations K-bar beyond 32 (ablation) | RecurrSim · Burgers, KdV, KS | best trajectory error | No clear effect |
| train with larger K-bar (ablation) | RecurrSim · Burgers, KdV, KS | trajectory error at small K | Hurt |
| increase recurrent iterations K at test time up to 16 (Burgers) or 8 (KdV) | RecurrFNO · Burgers and KdV | trajectory error (L2) | Improved |
Closest papers by meaning
- Derived Fields Preserve Fine-Scale Detail in Budgeted Neural Simulators (NeurIPS 2026 · similarity 0.88)
- Compute-optimal data scaling for neural surrogates via multi-fidelity training (NeurIPS 2026 · similarity 0.86)
- Geometric and Physical Constraints Synergistically Enhance Neural PDE Surrogates (ICML 2025 · similarity 0.84)
- Breakeven complexity: A new perspective on neural partial differential equation solvers (NeurIPS 2026 · similarity 0.84)
- Hierarchical Implicit Neural Emulators (NeurIPS 2025 · similarity 0.84)
- Simple ReFlow: Improved Techniques for Fast Flow Models (ICLR 2025 · similarity 0.84)
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.
