Research map / Architectures and efficiency / RNNs and normalization
Multi-Scale Wavelet Transformers for Operator Learning of Dynamical Systems
Xuesong Wang, Michael Groom, Rafael Oliveira, He Zhao, Terence O'kane, Edwin V. Bonilla
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 wavelet attention block (ablation) | MSWT · CKF step 64 | Rel L2: 1.09 -> 0.35 | Improved |
| use wavelet-preserving down/up-sampling instead of strided convolutions (ablation) | MSWT · CKF step 64 | Rel L2: 0.71 -> 0.35 | Improved |
| add patch tokenizer (vs patch size 1) (ablation) | MSWT · CKF step 64 | Rel L2: 0.18 -> 0.35 | Hurt |
| add patch tokenizer (vs patch size 1), training time (ablation) | MSWT · CKF | train time: 3h4' -> 1h17' | Improved |
| increase patch size from 4 to 8 (ablation) | MSWT · CKF step 64 | Rel L2: 0.5535 -> 0.9466 | Hurt |
| increase patch size from 2 to 4 (ablation) | MSWT · CKF step 64 | Rel L2: 0.5776 -> 0.5535 | No clear effect |
| scale up MSWT parameters (ablation) | MSWT · CKF final rollout step | Rollout rel. L2: 1.237877 -> 0.289241 | Improved |
| scale up FNO parameters (ablation) | FNO · CKF final rollout step | Rollout rel. L2: 0.806708 -> 0.911812 | No clear effect |
| replace FNO with MSWT | neural operator · CKF | Rel L2 step 64: 0.91 -> 0.33 | Improved |
| replace FNO with MSWT | neural operator · SWE | Rel L2 step 81: 0.58 -> 0.42 | Improved |
| use MSWT instead of HFS at one step | neural operator · SWE step 1 | EMLR: 0.01 -> 0.02 | Hurt |
| replace LUCIE with MSWT | climate emulator · ERA5 10-year rollout | climatology RMSE: 2.97 -> 1.31 (surface pressure) | Improved |
Closest papers by meaning
- Wavelet Flow Matching for Multi-Scale Physics Emulation (NeurIPS 2026 · similarity 0.85)
- Chaos Meets Attention: Transformers for Large-Scale Dynamical Prediction (ICML 2025 · similarity 0.85)
- Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization (ICML 2025 · similarity 0.84)
- Learning Interpretable Hierarchical Dynamical Systems Models from Time Series Data (ICLR 2025 · similarity 0.84)
- Wavelet Diffusion Neural Operator (ICLR 2025 · similarity 0.83)
- WaveletMixer: A Multi-Resolution Wavelets Based MLP-Mixer for Multivariate Long-Term Time Series Forecasting (AAAI 2025 · 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.
