Research map / Vision-language and multimodal / Cross-modal learning
TC-Diffuser: Bi-Condition Multi-Modal Diffusion for Tropical Cyclone Forecasting
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 TC-Diffuser, a bi-condition multi-modal diffusion model with a PGSA-LSTM module that forecasts tropical cyclone trajectory, pressure and wind speed from 1D track data, 2D geopotential maps and environmental data.
Problem
Multi-attribute tropical cyclone forecasting from multi-modal data (1D TC attributes, 2D geopotential height maps, environmental data) is hampered by noise, interdependence among attributes, and a semantic gap between 1D and 2D modalities.
Approach
A diffusion model guided sequentially by two conditions: a specific condition (environmental data and TC age, for attribute-specific traits) and a shared condition (spatio-temporal 1D/2D features from the PGSA-LSTM module, where 1D data prompts 2D temporal fusion), applied in the order specific then shared.
Key results
- TC-Diffuser achieves 20.50 km trajectory error at 6h and 75.15 km at 24h, versus 23.14 and 93.08 km for MGTCF.
- Reports 11%-48% trajectory, 18%-71% pressure, and 6%-71% wind speed improvements over MGTCF, and beats the CMO NWP model on all metrics using a single A6000 GPU.
- Specific-then-shared condition ordering performs best on long-term metrics (e.g., 24h distance 75.15 km vs 84.56 km for combined conditions).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TC-Diffuser instead of MGTCF | TC forecasting on CMA-BST WP · CMA-BST WP, 6h–24h | distance (km): 23.14 -> 20.50; 43.37 -> 22.63; 67.09 -> 40.85; 93.08 -> 75.15 | Helped |
| Use TC-Diffuser instead of MGTCF | TC forecasting on CMA-BST WP · CMA-BST WP, 6h–24h | pressure error (hPa): 1.37 -> 1.12; 2.04 -> 0.60; 2.66 -> 1.59; 3.29 -> 2.67 | Helped |
| Use TC-Diffuser instead of MGTCF | TC forecasting on CMA-BST WP · CMA-BST WP, 6h–24h | wind speed error (m/s): 0.73 -> 0.69; 1.17 -> 0.34; 1.55 -> 0.88; 1.86 -> 1.50 | Helped |
| Use TC-Diffuser instead of CMO | TC forecasting on CMA-BST WP · CMA-BST WP, 6h–24h | distance (km): 37.08 -> 20.50; 52.93 -> 22.63; 60.69 -> 40.85; 75.49 -> 75.15 | Mixed |
| Use TC-Diffuser instead of CMO | TC forecasting on CMA-BST WP · CMA-BST WP, 6h–24h | pressure error (hPa): 2.67 -> 1.12; 4.30 -> 0.60; 5.04 -> 1.59; 6.31 -> 2.67 | Helped |
| Use TC-Diffuser instead of CMO | TC forecasting on CMA-BST WP · CMA-BST WP, 6h–24h | wind speed error (m/s): 2.29 -> 0.69; 3.45 -> 0.34; 2.75 -> 0.88; 5.00 -> 1.50 | Helped |
| Guide 2D data with 1D data using PGSA-LSTM (ablation) | TC-Diffuser · CMA-BST WP | prediction error: 13.34%, 16.87%, 12.62% | Helped |
| Remove the Bi-condition (ablation) | TC-Diffuser · CMA-BST WP | distance (km): -; -; 81.28; 163.12 | Hurt |
| Remove the shared condition (ablation) | TC-Diffuser · CMA-BST WP, 24h | distance (km): 306.49 | Hurt |
| Remove the specific condition (ablation) | TC-Diffuser · CMA-BST WP, 24h | distance (km): 81.18 | Hurt |
| Move 1D data from shared condition to specific condition (ablation) | TC-Diffuser · CMA-BST WP, 24h | distance (km): 117.96 | Hurt |
| Guide with shared condition before specific condition rather than specific before shared (ablation) | TC-Diffuser · CMA-BST WP, 24h | distance (km): 75.15 -> 81.13 | Hurt |
Surprising findings
- Testing with randomly wrong expert knowledge (future move direction and intensity change direction) has negligible impact on predictions.
- The 6-hour predictions are suboptimal relative to longer horizons, attributed to the shared bi-condition overemphasizing global historical information.
Limitations
- Inference is slower than other deep learning baselines due to multi-step denoising.
- Short-term focus; 6-hour predictions are not improved as much as longer horizons.
- Evaluated only on the Western North Pacific dataset in main results; other oceans appear only in supplementary material.
Experimental setup
Datasets: CMA-BST (MGTCF dataset) TC best track geopotential height maps environmental data
Benchmarks: CMA-BST / MGTCF western North Pacific test set (2017-2021)
Metrics: trajectory distance error (km) pressure error (hPa) wind speed error (m/s)
Baselines: GRU NMPT DLM TCIF-fusion SGAN GBRNN MMSTN MGTCF TAM-CL CMO Pangu
Builds on: MGTCF Trajectron++ Self-Attention ConvLSTM denoising diffusion probabilistic models
Closest papers by meaning
- TCP-Diffusion: A Multi-modal Diffusion Model for Global Tropical Cyclone Precipitation Forecasting with Change Awareness (ICML 2025 · similarity 0.92)
- VQLTI: Long-Term Tropical Cyclone Intensity Forecasting with Physical Constraints (AAAI 2025 · similarity 0.86)
- IDOL: Meeting Diverse Distribution Shifts with Prior Physics for Tropical Cyclone Multi-Task Estimation (NeurIPS 2025 · similarity 0.84)
- DiffDA: a Diffusion model for weather-scale Data Assimilation (ICML 2024 · similarity 0.84)
- OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales (NeurIPS 2025 · similarity 0.84)
- Elucidated Rolling Diffusion Models for Probabilistic Forecasting of Complex Dynamics (NeurIPS 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.