Research map / Image generation / Fast diffusion sampling
TCP-Diffusion: A Multi-modal Diffusion Model for Global Tropical Cyclone Precipitation Forecasting with Change Awareness
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 TCP-Diffusion, a multi-modal conditional diffusion model that forecasts 12-hour global tropical cyclone rainfall fields by predicting adjacent rainfall residuals (ARP) and conditioning on environmental variables and ERA5-IFS forecasts.
Problem
Forecasting tropical cyclone (TC) precipitation around a moving TC center globally is under-studied, and existing DL rainfall nowcasting methods accumulate errors, lack physical consistency, and ignore TC-related meteorological factors and NWP forecasts.
Approach
Replace absolute rainfall prediction with adjacent residual prediction (predicting 3-hourly rainfall changes that are accumulated onto the latest observation), inside a diffusion model whose 3D U-Net is conditioned on multi-modal historical encoders and an NWP future-forecast encoder.
Key results
- TCP-Diffusion achieves ETS-24 of 0.14703 and ETS-60 of 0.00644 with TP_MAE 0.42344, outperforming U-Net, PreDiff, NowcastNet; U-Net is best on ETS-6 (0.44169 vs 0.43788).
- Against ECMWF-IFS on the TIGGE comparison, TCP-Diffusion reaches ETS-6 0.41238 vs 0.30193 and TP_MAE 0.47446 vs 0.50661.
- Ablation: adding ARP, multi-modal encoders, and NWP future input improves ETS-6 from 0.39109 to 0.43788.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add adjacent residual prediction (ARP) (ablation) | Spatio-temporal diffusion baseline · TC precipitation test set, 3–12 h | ETS-24: 0.12474 -> 0.13014 | Helped |
| Add multi-modal TC environment information (ablation) | ARP model · TC precipitation test set, 3–12 h | ETS-24: 0.13014 -> 0.14253 | Helped |
| Add NWP forecast inputs (ablation) | ARP + multi-modal environment model · TC precipitation test set, 3–12 h | ETS-24: 0.14253 -> 0.14703 | Helped |
| Use TCP-Diffusion instead of U-Net | TC precipitation forecasting · TC precipitation test set, 3–12 h | ETS-6: 0.44169 -> 0.43788 | Hurt |
| Use TCP-Diffusion instead of Persistence | TC precipitation forecasting · TC precipitation test set, 3–12 h | ETS-24: 0.14530 -> 0.14703 | Helped |
| Use TCP-Diffusion instead of ECMWF-IFS | NWP comparison · NWP comparison test samples | ETS-24: 0.08348 -> 0.12790 | Helped |
| Use TCP-Diffusion instead of FuXi | TC precipitation forecasting · TC precipitation, 2020 samples, 6–12 h | ETS-24: 0.03113 -> 0.10348 | Helped |
Surprising findings
- Persistence forecasting outperforms the other deep learning baselines (U-Net, PreDiff, NowcastNet) on most metrics, except TCP-Diffusion.
- FuXi, a large global weather model, performs very poorly on TC rainfall (ETS-6 0.02770) and strongly overestimates intensity.
Limitations
- Performance degrades with lead time, and the model is limited to 12-hour forecasts.
- Underperforms U-Net on light rainfall (ETS-6) and performs worse during TC formation and dissipation phases.
- Uses ERA5-IFS ensemble-mean forecasts for training rather than the higher-quality ECMWF-IFS, and the ECMWF comparison uses different test samples.
Experimental setup
Datasets: IBTrACS MSWEP ERA5 ERA5-IFS TIGGE ECMWF-IFS
Benchmarks: MSWEP-based TC rainfall test set (126 TCs, 2018-2020) TIGGE ECMWF-IFS comparison
Metrics: ETS-6 ETS-24 ETS-60 TP_MAE RAPSD rainfall frequency distribution
Baselines: U-Net PreDiff NowcastNet Persistence ECMWF-IFS ERA5-IFS FuXi
Builds on: denoising diffusion probabilistic models 3D U-Net PreDiff NowcastNet
Closest papers by meaning
- TC-Diffuser: Bi-Condition Multi-Modal Diffusion for Tropical Cyclone Forecasting (AAAI 2025 · similarity 0.92)
- VQLTI: Long-Term Tropical Cyclone Intensity Forecasting with Physical Constraints (AAAI 2025 · similarity 0.85)
- OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales (NeurIPS 2025 · similarity 0.84)
- Integrating Multi-Source Data for Long Sequence Precipitation Forecasting (AAAI 2025 · similarity 0.83)
- CasCast: Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling (ICML 2024 · similarity 0.83)
- Continuous Ensemble Weather Forecasting with Diffusion models (ICLR 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.