Research map / Probabilistic and scientific modeling / Molecules and proteins
Long-Term Typhoon Trajectory Prediction: A Physics-Conditioned Approach Without Reanalysis Data
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 LT3P, a real-time typhoon trajectory predictor that uses Unified Model (UM) forecast fields with a bias corrector toward ERA5 and a diffusion-based trajectory head, forecasting 72 hours ahead at 6-hour steps, and releases the PHYSICS TRACK dataset.
Problem
Typhoon trajectory forecasts beyond 24 hours are needed for damage control, but data-driven methods rely on ERA5 reanalysis, which is delayed by days and so unusable in real time, and NWP outputs require expert interpretation.
Approach
Pretrain a 3D spatio-temporal physics encoder on ERA5 fields, freeze it, then train a bias corrector mapping real-time UM fields to ERA5-like representations, and fuse these physics features into a trajectory predictor via cross-attention so it can run on UM data without reanalysis.
Key results
- LT3P (Bias-corrected UM)-Ens reaches 143.03 km FDE at 72 h, versus 225.59 km for NCEP-GEFS and 210.71 km for ECMWF-EPS in the ensemble comparison.
- LT3P (ERA5 Only)-Ens reaches 70.94 km FDE at 72 h, the best among compared methods, but it relies on reanalysis data unavailable in real time.
- In stochastic (20-sample) prediction, LT3P (Bias-corrected UM) achieves 65.24 km FDE at 72 h versus 164.99 km for SocialGAN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use bias-corrected UM instead of UM-only inputs | LT3P · Western North Pacific, 72 h ensemble | FDE (km): 390.92 -> 143.03 | Helped |
| Add joint training to UM-only training (ablation) | LT3P · Western North Pacific | FDE (km): 390.92 -> 190.75 | Helped |
| Add phase 1 pre-training to joint training (ablation) | LT3P · Western North Pacific | FDE (km): 190.75 -> 198.11 | Hurt |
| Add bias correction to joint training with pre-training (ablation) | LT3P · Western North Pacific | FDE (km): 198.11 -> 143.03 | Helped |
| Use SocialGAN instead of MID as the LT3P backbone (ablation) | LT3P · Western North Pacific | FDE (km): 143.03 -> 155.23 | Hurt |
| Use PECNet instead of MID as the LT3P backbone (ablation) | LT3P · Western North Pacific | FDE (km): 143.03 -> 152.07 | Hurt |
| Use real-time bias-corrected UM rather than ERA5-only inputs | LT3P · Western North Pacific, 72 h ensemble | FDE (km): 70.94 -> 143.03 | Hurt |
Surprising findings
- Human trajectory predictor MID performs poorly in ensemble forecasting but gains substantially when physics features from UM are added.
- Bias-corrected UM-based model (143.03 km at 72 h) lags the ERA5-based version substantially despite the correction, showing a persistent domain gap.
Limitations
- Only applied to typhoons in the Western North Pacific; not evaluated on other tropical cyclones.
- Bias-corrected UM model trails the ERA5-based model considerably, and the UM record only begins in 2010.
- Baseline comparisons with MMSTN and MGTCF were evaluated only on 2019 because best track data were unavailable.
Experimental setup
Datasets: ERA5 IBTrACS Best Track Unified Model forecast data PHYSICS TRACK
Benchmarks: PHYSICS TRACK
Metrics: ADE FDE
Baselines: SocialGAN STGAT PECNet MID MMSTN MGTCF JTWC JMA-GEPS ECMWF-EPS NCEP-GEFS UKMO-EPS
Builds on: MID SocialGAN PECNet MGTCF
Closest papers by meaning
- ChaosBench: A Multi-Channel, Physics-Based Benchmark for Subseasonal-to-Seasonal Climate Prediction (NeurIPS 2024 · similarity 0.84)
- ClimODE: Climate and Weather Forecasting with Physics-informed Neural ODEs (ICLR 2024 · similarity 0.83)
- VQLTI: Long-Term Tropical Cyclone Intensity Forecasting with Physical Constraints (AAAI 2025 · similarity 0.82)
- TC-Diffuser: Bi-Condition Multi-Modal Diffusion for Tropical Cyclone Forecasting (AAAI 2025 · similarity 0.82)
- TCP-Diffusion: A Multi-modal Diffusion Model for Global Tropical Cyclone Precipitation Forecasting with Change Awareness (ICML 2025 · similarity 0.81)
- Generalizing Weather Forecast to Fine-grained Temporal Scales via Physics-AI Hybrid Modeling (NeurIPS 2024 · similarity 0.81)
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.