Organizations: School of Computer Science, School of Software, Nanjing University of Information Science and Technology, Nanjing, 210044, China. · School of Atmospheric Sciences, Nanjing University of Information Science and Technology, Nanjing, 210044, China. · Shanghai Typhoon Institute, China Meteorological Administration, Shanghai, 200030, China. · School of Computer Science, Nanjing University of Posts and Telecommunications, Nanjing, 210023, China.
Tropical cyclones (TCs) rank among the most destructive natural hazards, yet their forecasting faces fundamental trade-offs: numerical weather prediction (NWP) models are computationally prohibitive and struggle to leverage historical data, while existing deep learning (DL)-based intelligent models are variable-specific and deterministic, which fail to generalize across different forecasting variables. Here we present CycloneMAE, a scalable multi-task forecasting model that learns transferable TC representations from multi-modal data using a TC structure-aware masked autoencoder. By coupling a discrete probabilistic gridding mechanism with a pre-train/fine-tune paradigm, CycloneMAE simultaneously delivers deterministic forecasts and probability distributions. Evaluated across five global ocean basins, CycloneMAE outperforms leading NWP systems in pressure and wind forecasting up to 120 hours and in track forecasting up to 24 hours. Attribution analysis via integrated gradients reveals physically interpretable learning dynamics: short-term forecasts rely predominantly on the internal core convective structure from satellite imagery, whereas longer-term forecasts progressively shift attention to external environmental factors. Our framework establishes a scalable, probabilistic, and interpretable pathway for operational TC forecasting.
Figures & tables
Figure 1: Schematic of the CycloneMAE model. a , Pre-training architecture. A TC structure-aware MAE model is designed to learn generalizable representations from multi-modal input data. b , Fine-tuning architecture. The pre-trained encoders are transferred to downstream forecasting tasks and only the forecasting head is needed to train. For each forecasting head, a discrete probabilistic gridding scheme is designed to generate probabilistic forecasts.
Figure 2: Performance comparison between CycloneMAE and global operational numerical models in three different basins. a, Forecast results achieved by CycloneMAE, ECMWF-IFS, NCEP-GFS and CMA-GFS in the WP basin. For each model, three forecasting variables (i.e., MSLP, MSW and Track) are evaluated using mean absolute error. b and c are the same as a but in the EP basin and the NA basin, respectively.
Figure 3: Year-by-year performance comparisons between CycloneMAE and operational numerical models in the WP basin. a-e , Forecast errors of MSLP achieved by different models in the year 2020 to 2024 sequentially. Errors are evaluated by using mean absolute error. f-j and k-o are the same as a-e but for the MSW and the Track errors, respectively.
Figure 4: Case studies of full lifecycle evolution and uncertainty estimation for representative TCs. Analysis of In-fa (WP), Doksuri (WP), Earl (NA), and Teddy (NA), showcasing trajectory tracking (Rows 1 and 3) and probabilistic boxplots at specific target times (Rows 2 and 4).
Figure 5: IG attribution analysis for CycloneMAE. The relative contribution weights of 16 predictors are evaluated across different lead times for MSW, MSLP, and Track forecasts. Predictors are color-coded into four groups: satellite imagery (purple), surface fields (blue), 850 hPa lower troposphere (green), and 200 hPa upper troposphere (red).
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Data
Source
Spatial Resolution
Time Resolution
Selected Variables
GridSat-B1 Knapp et al. (2011)
NOAA (NCDC)
0.07∘
3 h
2 Channels: Infrared Window Brightness Temperature (IR, ∼11μm ); Water Vapor Brightness Temperature (WVP, ∼6.7μm ).
ERA5 Hersbach et al. (2020)
ECMWF
0.25∘
1 h
14 Channels: Geopotential Height ( z850,z200 ), Temperature ( t850,t200 ), Specific Humidity ( q850,q200 ), U-wind ( u850,u200 ), V-wind ( v850,v200 ), 10m wind ( u10,v10 ), 2m Temperature ( t2m ), Mean Sea Level Pressure ( msl ).
IBTrACS Knapp et al. (2010) / CMA Lu et al. (2021)
NOAA/CMA
N/A
6 h
3 Variables: MSW, MSLP, Track (Longitude and Latitude)
TIGGE Bougeault et al. (2010)
ECMWF (C)
0.5∘
12 h
14 Channels: Variables are strictly consistent with ERA5, which are used as substitute input for environmental fields during the testing phase.
Note: NOAA stands for National Oceanic and Atmospheric Administration. NCDC stands for National Climatic Data Center. In the TIGGE dataset source, (C) specifies the unperturbed Control forecast run, which is selected as the operational substitute for ERA5 during testing.
Appendix
Table 1: Detailed description about each data used in this study.
Figure 6: Visualization of spatial cropping for multi-modal inputs.
Figure 7: Global distribution of observed versus forecasted TCs in 2020 to 2024. The visualization compares the best track data with CycloneMAE’s forecast results with a lead time of 6 hours across five major ocean basins. The color bar represent MSW with green-to-dark-green indicating observed MSW and purple-to-dark-purple indicating forecasted MSW.
Basin
Model
MSW (m/s)
MSLP (hPa)
Track (km)
6h
24h
48h
72h
96h
120h
6h
24h
48h
72h
96h
120h
6h
24h
48h
72h
SI
CycloneMAE
3.4
4.7
4.8
5.2
5.5
6.1
5.5
6.2
6.7
7.7
8.0
8.6
55.1
90.6
239.8
322.9
HWRF
4.4
5.7
7.4
8.9
10.8
12.3
5.0
6.7
9.6
10.9
13.1
15.4
60.7
111.0
183.1
240.0
COAMPS-TC
4.5
5.7
7.1
8.3
8.8
10.4
4.8
6.3
7.9
9.5
10.5
12.3
58.8
94.9
157.9
220.6
UKM
5.2
6.9
8.6
9.8
10.6
11.2
5.8
7.6
10.0
12.0
13.2
14.6
60.8
92.6
140.5
191.3
SP
CycloneMAE
3.4
4.6
5.1
5.6
5.9
6.1
5.2
5.4
6.7
6.6
8.0
8.2
62.9
103.1
182.3
260.8
Appendix
Table 2: Forecast performance of different models in terms of mean absolute error in the SI and SP basins. The forecast lead times are up to 120 hours for MSW/MSLP and 72 hours for track. The comparison operational NWP models include HWRF, COAMPS-TC and UKM.
Tropical cyclone (TC) forecasting is critical for disaster warning and emergency response. Deep learning methods address computational challenges but often neglect physical relationships between TC attributes, resulting in predictions lacking physical consistency. To address this, we propose Phys-Diff, a physics-inspired latent diffusion model that disentangles latent features into task-specific components (trajectory, pressure, wind speed) and employs cross-task attention to introduce prior physics-inspired inductive biases, thereby embedding physically consistent dependencies among TC attributes. Phys-Diff integrates multimodal data including historical cyclone attributes, ERA5 reanalysis data, and FengWu forecast fields via a Transformer encoder-decoder architecture, further enhancing forecasting performance. Experiments demonstrate state-of-the-art performance on global and regional datasets.
Lei Liu, Xiaoning Yu, Kang Chen +4
University of Science and Technology of China, Hefei, China · Hefei University of Technology, Hefei, China
Tropical cyclones (TCs) pose severe risks from strong winds and heavy rainfall. However, forecasting their track and intensity remains challenging due to chaotic atmosphere and the rapid amplification of initial condition errors, leading to growing forecast uncertainty. While numerical weather prediction (NWP) and deep learning models have made progress, they remain computationally demanding and often fail under complex meteorological scenarios. Here, we present Tianmu-TC, a physics-constraints generative framework for global TC forecasting. Trained on Western North Pacific data, Tianmu-TC leverages physics-constraints to generate controllable outputs with reduced uncertainty thus improving forecast reliability. Experiments show Tianmu-TC outperforms deterministic and ensemble meteorological artificial intelligence models and authoritative NWP systems such as ECMWF in global ocean basins, with significantly lower computational cost. We further show Tianmu-TC performs well in challenging scenarios such as data sparsity, anomaly tracks, rapid intensification and weakening. These findings suggest physics-constraints generative AI offers a promising approach for reliable, efficient global TC forecasting.
Shiqi Zhang, Pan Mu, Cheng Huang +6
College of Computer Science and Technology, Zhejiang University of Technology, Hangzhou & 310023, China. · School of Control Science and Engineering, Shandong University, Jinan & 250061, China. · School of Computer Science and Engineering, Tianjin University of Technology, Tianjin & 300384, China. +1
Tropical cyclones are growing more destructive in a changing climate, and efficient forecasting of their structure and track has become a necessity. Deep generative models promise an alternative to computationally expensive numerical weather prediction (NWP), yet current systems produce either satellite imagery or atmospheric fields, never both; they need many sampling steps, putting them out of reach of modest hardware; and their storm tracks come from regression heads with no physical link to the generated atmosphere. This work presents a single-pass model that jointly forecasts GRIDSAT-B1 infrared imagery and four ERA5 atmospheric fields (U-wind, V-wind, air temperature, and surface pressure) out to nine hours. A five-channel variational autoencoder compresses each 5 x 256 x 256 frame to a 4 x 64 x 64 latent, and a conditional rectified-flow UNet with a factorized temporal-attention module predicts the next three frames from three past frames, their best-track coordinates, and timestamps. The model is then reward-fine-tuned (DRaFT) against a differentiable track error derived from the predicted winds through a steering-flow calculation. On held-out 2022 storms the model reaches 16.35 dB PSNR and 0.759 SSIM, ahead of a reproduced cascaded-diffusion baseline at every lead time (+0.84 dB at +9 h) while sampling ~30x faster (56 ms vs. 1673 ms). Track error at +9 h is 62.4 km, 15% below the baseline, and a reward fine-tuning study demonstrates a further 8-11% track-error reduction across sampler budgets.
Meheru Zannat, Sk. Md. Masudul Ahsan
Department of Computer Science and Engineering Khulna University of Engineering & Technology Khulna 9203, Bangladesh