Research map / Architectures and efficiency / RNNs and normalization
Overtone: Cyclic Patch Modulation for Cleaner, Faster Physics Emulators
Payel Mukhopadhyay, Michael McCabe, Ruben Ohana, Miles Cranmer
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 |
|---|---|---|---|
| cycle stride/patch size at inference (CSM) instead of fixed patch | Axial ViT (50M) · 10-step rollout, Shear Flow | VRMSE: 0.0793 -> 0.0565 | Improved |
| cycle stride/patch size at inference (CSM) instead of fixed patch | Vanilla ViT (100M) · 10-step rollout, Supernova | VRMSE: 1.75 -> 1.20 | Improved |
| cycle stride/patch size at inference (CSM) instead of fixed patch | Vanilla ViT (100M) · 10-step rollout, Active Matter | VRMSE: 0.370 -> 0.359 | Improved |
| use 4->8->16 stride order instead of random order (ablation) | CSM Vanilla ViT · 10-step rollout, Shear Flow | VRMSE: 0.0433 -> 0.0375 | Improved |
| use 4->8->16 stride order instead of 8->4->16 (ablation) | CSM Vanilla ViT · 10-step rollout, Shear Flow | VRMSE: 0.0442 -> 0.0375 | Improved |
| use 4->8->16 stride order instead of 16->4->8 (ablation) | CSM Vanilla ViT · 10-step rollout, Supernova | VRMSE: 1.325 -> 1.204 | Improved |
| reduce patch size from 16 to 4 at inference | CSM/CKM · Active Matter 256^2 | VRMSE | Improved |
| train one flexible CSM model instead of separate fixed patch models | Vanilla ViT (100M) · Next-step Shear Flow, 2048 tokens | VRMSE: 0.00677 -> 0.00546 | Improved |
| use flexible CKM CViT encoder instead of fixed patch | CViT · Shear Flow 128/16 | VRMSE: 0.319 -> 0.127 | Improved |
| use flexible CKM CViT encoder instead of fixed patch | CViT · Turbulent Radiative Layer 2D 3072/4 | VRMSE: 0.299 -> 0.302 | No clear effect |
| omit patch size 8 from training (ablation) | CSM/CKM · excluded patch size | VRMSE | Hurt |
| use warm-up then coarse schedule or two-step dwell instead of 4->8->16 | CSM Vanilla ViT · rollout schedules | VRMSE | No clear effect |
Closest papers by meaning
- Lost in Latent Space: An Empirical Study of Latent Diffusion Models for Physics Emulation (NeurIPS 2025 · similarity 0.84)
- APEBench: A Benchmark for Autoregressive Neural Emulators of PDEs (NeurIPS 2024 · similarity 0.83)
- PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations (ICML 2025 · similarity 0.83)
- Compute-optimal data scaling for neural surrogates via multi-fidelity training (NeurIPS 2026 · similarity 0.83)
- Neural Emulator Superiority: When Machine Learning for PDEs Surpasses its Training Data (NeurIPS 2025 · similarity 0.83)
- Mosaic: A Benchmark Suite for Differentiable Physics Solvers (NeurIPS 2026 · similarity 0.83)
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.
