3DTCR: Vortex-following generative reconstruction improves tropical cyclone intensity prediction from global AI weather forecasts
Authors: Jun Liu, Xiaohui Zhong, Kai Zheng, Jiarui Li, Yifei Li, Tao Zhou, Wenwen Zhang, Hao Li
Organizations: Artificial Intelligence Innovation and Incubation Institute, Fudan University, Shanghai, 200433, China. · Department of Atmospheric and Oceanic Sciences, Institute of Atmospheric Sciences, Fudan University, Shanghai, 200438, China. · Shanghai Academy of Artificial Intelligence for Science, Shanghai, 200232, China.
Global numerical weather prediction (NWP) and artificial intelligence (AI) models often underestimate tropical cyclone (TC) intensity, partly because coarse grids underresolve fine-scale three-dimensional inner-core structures that influence TC intensity evolution. Kilometer-scale NWP simulations can better resolve these structures but are too costly for operational use, which also makes high-resolution TC data limited. Here we construct a 3-km, TC-centered dataset for 2018--2024 using the Weather Research and Forecasting (WRF) model and develop 3DTCR, a vortex-following generative framework that reconstructs three-dimensional TC structure from FuXi global forecasts. 3DTCR combines conditional flow matching (CFM) with two-stage training to learn a coarse-to-fine mapping from center-aligned reanalysis simulation pairs and then adapt it to forecast inputs through latent space alignment. Evaluated on 2024 western North Pacific (WNP) forecasts, 3DTCR reduces the root mean square error (RMSE) and the magnitude of the mean bias error (MBE) in maximum 10-m wind speed by 36.5% and 60.7% relative to FuXi, and by 8.9% and 28.2% relative to the European Centre for Medium-Range Weather Forecasts high-resolution system (ECMWF-HRES), respectively. Case studies suggest that reconstructions capture wind gradients between the eye and eyewall, vertical structure of horizontal winds, and vortex asymmetries, while inheriting the FuXi-predicted TC tracks. 3DTCR generates 5-day TC-centered atmospheric fields in approximately 3 minutes on one GPU, whereas a full-basin 3-km WRF simulation takes about 400 hours on 32 CPU cores. These results demonstrate that such a vortex-following generative reconstruction can add localized three-dimensional detail to global forecasts and improve TC intensity prediction without basin-wide kilometer-scale numerical integration.
Figures & tables
Method
Strategy
Loss function
Description
Baselines
FuXi
-
-
Global AI weather forecast
ERA5
-
-
Reanalysis data
ECMWF-HRES
-
-
Operational NWP forecast
Ours
3DTCR-Pre (ERA5)
Pretraining
LRF
Pretrained on ERA5, evaluated with ERA5 input
Table 1: Configurations and experimental roles of the 3DTCR variants. All variants use the same reconstruction backbone and WRF 3-km target. The data source in parentheses denotes the input used at inference. SFT denotes supervised fine-tuning, and MMD denotes maximum mean discrepancy.
Figure 1: Comparison of maximum 10-m wind speed (WS10M) among FuXi, ERA5, and WRF 3-km. Left panels show the spatial distribution of WS10M for Typhoon DOKSURI at 12 UTC 24 July 2023. Middle panels compare maximum WS10M with IBTrACS observations for 2242 forecasts covering forecast lead times of 6–120 hours in the 2024 WNP testing set. Right panels show kernel density estimates of the intensity distributions. Solid curves denote IBTrACS, and dashed curves denote intensity from ERA5, FuXi, and WRF. Repeated observations arise from 5-kt best-track increments and from forecasts with different initialization times that verify against the same observation. A random jitter of ±2 m/s is applied to the observations for visualization only. All metrics are computed using the original values.
Figure 2: Forecast performance for maximum tropical cyclone (TC) 10-m wind speed (WS10M). (a) TC maximum WS10M RMSE relative to IBTrACS best-track data. Shading denotes 95% confidence intervals obtained from 1000 bootstrap resamples. (b) Distributions of forecast bias, defined as the predicted minus the best-track maximum WS10M. Boxes indicate the interquartile range, center lines denote the median, plus signs denote the mean, and whiskers extend from the 5th to the 95th percentiles. Points show a random subset of forecast samples. Numbers in parentheses indicate the sample size at each lead time.
Figure 3: Near-surface structure of Typhoon KONG-REY at a forecast lead time of 18 h. The forecast was initialized at 12 UTC on 29 October 2024. Columns show 2-m temperature (T2M), 10-m u and v wind components (U10M and V10M), mean sea level pressure (MSL), and 10-m wind speed (WS10M). Rows show bilinearly interpolated FuXi fields, the 3DTCR reconstruction, and the WRF 3-km reference.
Figure 4: Multilevel wind-speed distributions and spectra for Typhoon KONG-REY. From top to bottom, rows show winds at 200 hPa, 500 hPa, 700 hPa, 850 hPa, and at 10 m. From left to right, columns show radial wind speed profiles, probability density functions, wind speed power spectra, and kinetic energy spectra. Results are shown for ERA5 (red), FuXi (blue), 3DTCR (purple), and WRF 3-km (green). Both spectra are plotted on logarithmic axes.
Figure 5: Ablation analysis of 3DTCR training strategies. Critical success index (CSI) for 10-m wind speed (WS10M) exceeding six thresholds from 13.9 to 32.7 m/s at lead times of 1–5 days (panels a – f ). Bar groups compare FuXi with 3DTCR variants trained with pretraining (3DTCR-Pre), supervised fine-tuning (3DTCR-SFT), and end-to-end training (3DTCR-E2E). Numbers in parentheses denote the sample size at each lead day. Parenthetical labels denote the inference input source. For example, 3DTCR-Pre (FuXi) denotes the pretrained model applied to FuXi inputs at inference. The labels “w/ MMD” and “w/o MMD” indicate whether MMD-based latent space alignment is applied.
Figure 6: Overview of the 3DTCR framework. (a) During pretraining, the model learns a coarse-to-fine mapping from spatially aligned ERA5 inputs to WRF 3-km targets using Rectified Flow. During supervised fine-tuning (blue arrows), ERA5 and FuXi forecast inputs are introduced, and a maximum mean discrepancy (MMD) loss aligns their latent space feature distributions at the U-Net bottleneck. (b) At inference, the TC tracker locates the TC in each FuXi forecast, a TC-centered window is extracted, and 3DTCR reconstructs the high-resolution vortex at each lead time.
Figure S1 : Atmospheric frontal systems and typical tropical cyclone structure. ( a ) Cross section through a gradually ascending warm front generating precipitation through rising warm air masses. ( b ) Cross section through a steep cold front generating precipitation through rising warm air masses. ( c ) Three-dimensional cutaway diagram of a tropical cyclone revealing its internal components, structural organization, and wind circulation patterns.
Figure S2 : WRF simulation framework and dataset construction. ( a ) WRF simulation domain (pink shaded area) and tropical cyclone tracks used for dataset construction. Historical TC trajectories are overlaid to show the spatial coverage of extreme weather events within the modeling domain. ( b ) WRF numerical simulation workflow. The processing chain includes data ingestion from terrestrial and gridded meteorological sources, preprocessing via WPS, model initialization through REAL, WRF simulation, and post-processing analysis. ( c ) Multilevel vertical structure of selected meteorological variables in the WRF dataset. Stacked horizontal slices show how the spatial distributions of these variables vary with height.
Variable Category
Variable Name
Abbreviation
Upper-air Variables
Geopotential
Z
Temperature
T
U component of wind
U
V component of wind
V
Relative humidity
R
Specific humidity
Q
Table S2: Complete variable information for the WRF post-processing dataset.
Figure S3 : Comparison of tropical cyclone fields across datasets and resolutions. Spatial distribution of key meteorological variables for Typhoon DOKSURI at 1200 UTC on 24 July 2023. The rows (from top to bottom) compare fields from FuXi, ERA5, the WRF 25-km simulation, and the WRF 3-km simulation. The columns show: (a, e, i, m) 10-m U wind component (U10M, m/s), (b, f, j, n) 10-m V wind component (V10M, m/s), (c, g, k, o) mean sea level pressure (MSL, hPa), and (d, h, l, p) 10-m wind speed (WS10M, m/s). The black contour lines represent coastlines. The WRF 3-km simulation displays finer inner-core wind features and sharper pressure gradients than FuXi, ERA5, and the WRF 25-km configuration. Color scales are indicated at the bottom of each column.
Figure S4 : Comparison of 10-m wind speed (WS10M) relative frequencies across wind speed bins at different spatial scales. Panels ( a – c ) show the field at the global scale, the field over the western North Pacific basin, and the 10 ∘ × 10 ∘ TC vortex sample at 3 km resolution, respectively. Panels ( d – f ) show the corresponding WS10M relative frequencies across different wind speed bins. For each scale, relative frequency is the fraction of grid points within the corresponding sampling domain that fall in each WS10M bin. Vertical dashed lines denote different TC intensity categories.
Parameter
Level
Radius
Threshold
Comments
WS10M
Surface
278 km
>8 m/s
Only required over land.
Vorticity
850 hPa
278 km
absolute value ≥5×10−5s−1
Geopotential thickness
850 hPa and 200 hPa
278 km
A maximum in thickness is necessary only after the TC has transitioned into an extratropical cyclone.
Table S3 : Parameters and thresholds related to the TC tracking.
Figure S5 : Schematic of the TC tracking algorithm. The algorithm first uses TC observations to determine an initial estimate of the TC position. It then coarsens the forecast data and searches for candidate TC centers. If a candidate satisfies the predefined TC criteria, the TC center is updated; otherwise, tracking stops. For subsequent tracking steps, an advection vector derived from the steering flow is used, either alone or in combination with linear extrapolation, depending on whether the current step is the first forecast step.
Figure S6 : Spatial displacement of 10-m wind speed (WS10M) for Typhoon MAWAR. FuXi, ERA5, ECMWF-HRES, and WRF 3-km fields are shown for Days 1–3 and 5 from the forecast initialized at 0000 UTC 24 May 2023. All fields are displayed in the common WRF 3-km domain. Shading denotes WS10M, and black crosses indicate the TC center in each field, defined by the minimum mean sea level pressure.
Figure S7 : Spatial displacement of 10-m wind speed (WS10M) for Typhoon DOKSURI. FuXi, ERA5, ECMWF-HRES, and WRF 3-km fields are shown for Days 1–3 and 5 from the forecast initialized at 0000 UTC 22 July 2023. All fields are displayed in the common WRF 3-km domain. Shading denotes WS10M, and black crosses indicate the TC center in each field, defined by the minimum mean sea level pressure.
Method
24 h
48 h
72 h
96 h
120 h
Baselines
FuXi
20.19
21.64
22.80
22.48
23.91
ERA5
18.92
19.77
20.04
19.73
21.20
ECMWF-HRES
13.96
14.90
15.63
16.83
19.12
Ours
3DTCR-E2E (FuXi, w/ MMD)
13.55
15.06
16.67
16.56
17.49
Table S4: Root mean square error (RMSE) of maximum 10-m wind speed (WS10M) predictions (m/s; lower is better ↓ ).
Figure S8 : Comparison of predicted and observed maximum 10-m wind speed (WS10M). Panels ( a – f ) compare predictions with IBTrACS observations for 2242 samples pooled across all forecast lead times in the 2024 WNP test dataset: ( a ) ECMWF-HRES, ( b ) 3DTCR-Pre (ERA5), ( c ) 3DTCR-Pre (FuXi), ( d ) 3DTCR-E2E (FuXi, w/ MMD), ( e ) 3DTCR-SFT (FuXi, w/o MMD), and ( f ) 3DTCR-SFT (FuXi, w/ MMD; full model). Solid lines show linear fits with 95% confidence intervals, and dashed lines indicate the 1:1 reference. The rightmost column shows marginal distributions using box plots and kernel density estimates (KDEs); inset values report the overall mean bias and RMSE. Because best-track intensities are typically reported in 5 kt increments, and forecasts from different initialization times can correspond to the same verifying observation, repeated observed values may appear in the scatter plots. To reduce overplotting, a random jitter of ±2 m/s was applied only along the horizontal axis to the plotted observational values for visualization purposes. All verification metrics were computed using the original, unjittered observational values.
Figure S9 : TC track errors across forecast lead times. Track error is calculated as the great-circle distance between the predicted and IBTrACS TC center coordinates. (a) Track MAE for 1–5-day forecasts, with shading denoting the 95% bootstrap confidence intervals estimated from 1000 resamples. (b) Distributions of TC track errors for 1–5-day forecasts. Boxes indicate the interquartile range, center lines the median, plus signs the mean, and whiskers the 5th–95th percentiles. Points represent randomly selected samples, and numbers in parentheses indicate the sample sizes. Because 3DTCR uses the TC centers provided by FuXi or ERA5, its track error curves and distributions overlap with those of the corresponding inputs.
Figure S10 : Typhoon KONG-REY multivariable field reconstruction at 96 h forecast lead time. The forecast was initialized at 1200 UTC 26 October 2024. Columns show 2-m temperature (T2M), 10-m u and v wind components (U10M and V10M), mean sea level pressure (MSL), and 10-m wind speed (WS10M). Rows show bilinearly interpolated FuXi fields, the 3DTCR reconstruction, and the WRF 3-km reference.
Figure S11 : Typhoon YAGI multivariable field reconstruction at 18 h forecast lead time. The forecast was initialized at 1200 UTC 3 September 2024. Columns show 2-m temperature (T2M), 10-m u and v wind components (U10M and V10M), mean sea level pressure (MSL), and 10-m wind speed (WS10M). Rows show bilinearly interpolated FuXi fields, the 3DTCR reconstruction, and the WRF 3-km reference.
Figure S12 : Typhoon YAGI multivariable field reconstruction at 108 h forecast lead time. The forecast was initialized at 0000 UTC 2 September 2024. Columns show 2-m temperature (T2M), 10-m u and v wind components (U10M and V10M), mean sea level pressure (MSL), and 10-m wind speed (WS10M). Rows show bilinearly interpolated FuXi fields, the 3DTCR reconstruction, and the WRF 3-km reference.
Figure S13 : Typhoon YINXING multivariable field reconstruction at 72 h forecast lead time. The forecast was initialized at 1200 UTC 4 November 2024. Columns show 2-m temperature (T2M), 10-m u and v wind components (U10M and V10M), mean sea level pressure (MSL), and 10-m wind speed (WS10M). Rows show bilinearly interpolated FuXi fields, the 3DTCR reconstruction, and the WRF 3-km reference.
Figure S14 : Vertical wind structure of Typhoon KONG-REY at 0600 UTC on 30 October 2024. (a, b) Interpolation, (c, d) 3DTCR, and (e, f) WRF 3-km results. The left and right columns show latitudinal and longitudinal cross sections, respectively. Shading and contours indicate wind speed (m/s), with pressure levels (hPa) on the vertical axis. The central white region and adjacent regions of high wind speed correspond to the eye and eyewall, respectively.
Figure S15 : Vertical cross sections of tangential and radial wind for Typhoon KONG-REY (0600 UTC on 30 October 2024). (a) 3DTCR tangential wind; (b) WRF 3-km tangential wind; (c) Tangential wind difference (3DTCR minus WRF 3-km); (d) 3DTCR radial wind; (e) WRF 3-km radial wind; (f) Radial wind difference (3DTCR minus WRF 3-km). The horizontal axis represents the radius from the TC center (km), and the vertical axis represents pressure levels (hPa).
Figure S16 : Typhoon KONG-REY three-dimensional multivariable field reconstruction at 18 h lead time. Comparison across vertical pressure levels ( 200 , 500 , 700 , and 850 hPa ) for geopotential, temperature, u/v wind components, and 10-m surface wind speed, initialized at 1200 UTC 29 October 2024. Columns from left to right: Interpolation (FuXi), 3DTCR reconstruction, and WRF 3-km simulation (reference target).
Figure S17 : Comparison of 500 hPa wind and geopotential height fields for Typhoon KONG-REY (0600 UTC on 30 October 2024). (a) ERA5 interpolation data, (b) FuXi forecast interpolation data, (c) 3DTCR, (d) WRF 3-km. Color shading represents wind speed (m/s), black contours show geopotential height (gpm), and wind barbs indicate the 500 hPa wind field.
Figure S18 : Comparison of 850 hPa wind and geopotential height fields for Typhoon KONG-REY (0600 UTC on 30 October 2024). (a) ERA5 interpolation data, (b) FuXi forecast interpolation data, (c) 3DTCR, (d) WRF 3-km. Color shading represents wind speed (m/s), black contours show geopotential height (gpm), and wind barbs indicate the 850 hPa wind field.
Figure S19 : Comparison of near-surface wind and mean sea level pressure fields for Typhoon KONG-REY (0600 UTC on 30 October 2024). (a) ERA5 interpolation data, (b) FuXi forecast interpolation data, (c) 3DTCR, (d) WRF 3-km. Color shading represents wind speed (m/s), black contours show mean sea level pressure (hPa), and wind barbs indicate the 10-m wind field.
Figure S20 : Comparison of cloud-top brightness temperature and 10-m wind fields for Typhoon KONG-REY (0600 UTC 30 October 2024). (a) ERA5; (b) FuXi; (c) 3DTCR; (d) WRF 3-km simulation. Grayscale shading shows cloud-top brightness temperature (°C), colored wind barbs indicate 10-m wind speed (m/s) and direction, yellow lines denote coastlines, and red plus signs mark the IBTrACS best-track position.
Figure S21 : Temporal evolution of mean sea level pressure and 10-m wind field for Typhoon KONG-REY (0600 UTC on 29 October to 0000 UTC on 31 October 2024). Rows 1 and 3 show the 3DTCR results, while rows 2 and 4 show the corresponding WRF 3-km simulation results. Each group of four columns represents consecutive valid times. Color shading represents mean sea level pressure, wind barbs indicate 10-m wind speed and direction, the yellow star marks the predicted TC center, and the black solid line shows the IBTrACS best track. The time interval is 6 hours.
Figure S22 : Vertical wind structure of Typhoon YAGI at 0600 UTC on 4 September 2024. (a, b) Interpolation, (c, d) 3DTCR, and (e, f) WRF 3-km results. The left and right columns show latitudinal and longitudinal cross sections, respectively. Shading and contours indicate wind speed (m/s), with pressure levels (hPa) on the vertical axis. The central white region and adjacent regions of high wind speed correspond to the eye and eyewall, respectively.
Figure S23 : Vertical cross sections of tangential and radial wind for Typhoon YAGI (0600 UTC 4 September 2024). (a) 3DTCR tangential wind; (b) WRF 3-km tangential wind; (c) Tangential wind difference (3DTCR minus WRF 3-km); (d) 3DTCR radial wind; (e) WRF 3-km radial wind; (f) Radial wind difference (3DTCR minus WRF 3-km). The horizontal axis represents the radius from the TC center (km), and the vertical axis represents pressure levels (hPa).
Figure S24 : Typhoon YAGI three-dimensional multivariable field reconstruction at 18 h lead time. Comparison across vertical pressure levels ( 200 , 500 , 700 , and 850 hPa ) for geopotential, temperature, u/v wind components, and 10-m surface wind speed, initialized at 1200 UTC 3 September 2024. Columns from left to right: Interpolation (FuXi), 3DTCR reconstruction, and WRF 3-km simulation (reference target).
Figure S25 : Comparison of 500 hPa wind and geopotential height fields for Typhoon YAGI (0600 UTC 4 September 2024). (a) ERA5 interpolation data, (b) FuXi forecast interpolation data, (c) 3DTCR, (d) WRF 3-km. Color shading represents wind speed (m/s), black contours show geopotential height (gpm), and wind barbs indicate the 500 hPa wind field.
Figure S26 : Comparison of 850 hPa wind and geopotential height fields for Typhoon YAGI (0600 UTC 4 September 2024). (a) ERA5 interpolation data, (b) FuXi forecast interpolation data, (c) 3DTCR, (d) WRF 3-km. Color shading represents wind speed (m/s), black contours show geopotential height (gpm), and wind barbs indicate the 850 hPa wind field.
Figure S27 : Comparison of near-surface wind and mean sea level pressure fields for Typhoon YAGI (0600 UTC 4 September 2024). (a) ERA5 interpolation data, (b) FuXi forecast interpolation data, (c) 3DTCR, (d) WRF 3-km. Color shading represents wind speed (m/s), black contours show mean sea level pressure (hPa), and wind barbs indicate the 10-m wind field.
Figure S28 : Comparison of cloud-top brightness temperature and 10-m wind fields for Typhoon YAGI (0600 UTC 4 September 2024). (a) ERA5; (b) FuXi; (c) 3DTCR; (d) WRF 3-km simulation. Grayscale shading shows cloud-top brightness temperature (°C), colored wind barbs indicate 10-m wind speed (m/s) and direction, yellow lines denote coastlines, and red plus signs mark the IBTrACS best-track position.
Figure S29 : Temporal evolution of mean sea level pressure and 10-m wind field for Typhoon YAGI (from 0000 UTC 4 September to 1800 UTC 5 September 2024). Rows 1 and 3 show the 3DTCR results, while rows 2 and 4 show the corresponding WRF 3-km simulation results. Panels (a–p) show the temporal evolution at consecutive valid times. Color shading represents mean sea level pressure, wind barbs indicate 10-m wind speed and direction, the yellow star marks the predicted TC center, and the black solid line shows the IBTrACS best track. The time interval is 6 hours.
Figure S30 : Multivariable TC vortex reconstruction for Tropical Storm JOHN over the eastern Pacific at a 12 h forecast lead time. The forecast was initialized at 0000 UTC on 26 September 2024. Columns show 2-m temperature (T2M), 10-m u and v wind components (U10M and V10M), mean sea level pressure (MSL), and 10-m wind speed (WS10M). Rows show bilinearly interpolated FuXi fields and the 3DTCR reconstruction.
Figure S31 : Multivariable TC vortex reconstruction for Hurricane MILTON over the North Atlantic at a 114 h forecast lead time. The forecast was initialized at 1200 UTC 5 October 2024. Columns show 2-m temperature (T2M), 10-m u and v wind components (U10M and V10M), mean sea level pressure (MSL), and 10-m wind speed (WS10M). Rows show bilinearly interpolated FuXi fields and the 3DTCR reconstruction.
Figure S32 : Multivariable TC vortex reconstruction for Hurricane MILTON over the North Atlantic at a 120 h forecast lead time. The forecast was initialized at 1200 UTC 5 October 2024. Columns show 2-m temperature (T2M), 10-m u and v wind components (U10M and V10M), mean sea level pressure (MSL), and 10-m wind speed (WS10M). Rows show bilinearly interpolated FuXi fields and the 3DTCR reconstruction.
Figure S33 : Multilevel wind speed distributions and spectra for Typhoon YAGI. From top to bottom, rows show winds at 200, 500, 700 and 850 hPa and at 10 m. From left to right, columns show radial wind speed profiles, probability density functions, wind speed power spectra and kinetic energy spectra. Results are shown for ERA5 (red), FuXi (blue), 3DTCR (purple) and WRF 3-km (green). The wind speed and kinetic energy spectra are plotted on log–log axes.
Figure S34 : Wavenumber decomposition of wind speed for Typhoon KONG-REY from WRF 3-km (0600 UTC on 30 October 2024). (a) Original wind speed; (b)–(s) wavenumbers 0–17; (t) wavenumber 20. Color shading shows wind speed (m/s). Axes show distance from the center (km).
Figure S35 : Wavenumber decomposition of wind speed for Typhoon KONG-REY (0600 UTC 30 October 2024). The first through fifth columns show the original wind speed, wavenumber 0 (axisymmetric component), wavenumber 1 (dipole), wavenumber 2 (quadrupole), and the wavenumber 0–20 synthesis. The first row shows interpolation, the second row shows 3DTCR, and the third row shows WRF 3-km. Color shading represents wind speed (m/s), and axes indicate distance from the TC center (km). All data are interpolated to WRF 3-km resolution and transformed to polar coordinates centered on the TC center with a radius of 300 km; a Fourier transform is applied to the azimuthal data at each radial distance to extract the wavenumber components.
Figure S36 : Wavenumber decomposition of wind speed for Typhoon YAGI from WRF 3-km (0600 UTC on 4 September 2024). (a) Original wind speed; (b)–(s) wavenumbers 0–17; (t) wavenumber 20. Color shading shows wind speed (m/s). Axes show distance from center (km).
Figure S37 : Wavenumber decomposition of wind speed for Typhoon YAGI (0600 UTC 4 September 2024). The first through fifth columns show the original wind speed, wavenumber 0 (axisymmetric component), wavenumber 1 (dipole), wavenumber 2 (quadrupole), and the wavenumber 0–20 synthesis. The first row shows interpolation, the second row shows 3DTCR, and the third row shows WRF 3-km. Color shading represents wind speed (m/s), and axes indicate distance from the TC center (km). All data are interpolated to WRF 3-km resolution and transformed to polar coordinates centered on the TC center with a radius of 300 km; a Fourier transform is applied to the azimuthal data at each radial distance to extract the wavenumber components.
Figure S38 : Comparison of multivariable TC vortex fields for Typhoon AMPIL across different ablation variants. The forecasts are initialized at 1200 UTC 14 August 2024 with a lead time of 72 h. Columns show T2M, U10M, V10M, MSL, and WS10M. Rows show 3DTCR-E2E (FuXi, w/ MMD), 3DTCR-SFT (FuXi, w/o MMD), 3DTCR-SFT (FuXi, w/ MMD; full model), and the WRF 3-km reference, from top to bottom.
Method
Grid spacing
Grid size (Nlat×Nlon)
Computing resource
Runtime
WRF 25-km (WNP)
0.250∘
201×280
32 CPU cores
∼ 1 h
WRF 12-km (WNP)
0.125∘
401×560
32 CPU cores
∼ 6 h
WRF 3-km (WNP)
0.027∘
1852×2592
32 CPU cores
∼ 400 h
WRF 3-km (TC vortex)
0.027∘
370×370
32 CPU cores
∼ 12 h
3DTCR (TC vortex)
0.027∘
370×370
1 NVIDIA A100 GPU
∼ 3 min
Table S5: Computational resource requirements and runtime comparison for 5-day WNP and TC-centered tasks. All WRF simulations used the same Intel Xeon Platinum 8369B CPU platform at 2.90 GHz.
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 (TCs) pose severe threats to life, infrastructure, and economies in tropical and subtropical regions, underscoring the critical need for accurate and timely forecasts of both track and intensity. Recent advances in AI-based weather forecasting have shown promise in improving TC track forecasts. However, these systems are typically trained on coarse-resolution reanalysis data (e.g., ERA5 at 0.25 degree), which constrains predicted TC positions to a fixed grid and introduces significant discretization errors. Moreover, intensity forecasting remains limited especially for strong TCs by the smoothing effect of coarse meteorological fields and the use of regression losses that bias predictions toward conditional means. To address these limitations, we propose BaguanCyclone, a novel, unified framework that integrates two key innovations: (1) a probabilistic center refinement module that models the continuous spatial distribution of TC centers, enabling finer track precision; and (2) a region-aware intensity forecasting module that leverages high-resolution internal representations within dynamically defined sub-grid zones around the TC core to better capture localized extremes. Evaluated on the global IBTrACS dataset across six major TC basins, our system consistently outperforms both operational numerical weather prediction (NWP) models and most AI-based baselines, delivering a substantial enhancement in forecast accuracy. Remarkably, BaguanCyclone excels in navigating meteorological complexities, consistently delivering accurate forecasts for re-intensification, sweeping arcs, twin cyclones, and meandering events. Our code is available at https://github.com/DAMO-DI-ML/Baguan-cyclone.
AI weather models are in the process of revolutionising weather forecasting. While these models have been shown to achieve superior performance to physics-based NWP in forecasting tropical cyclone (TC) tracks, they tend to dramatically underestimate intensity. Here we present AIFS-TC, a simple correction to the AIFS-Single model that is competitive with the operational state-of-the-art for forecasting maximum wind speed and minimum central pressure at lead times of 12 h to seven days. This performance also holds for rapid intensification events. Notably, the entire system was autonomously designed and built by a large language model (Claude Fable 5) in a few hours, directed through a small number of natural-language prompts by a single domain scientist. That the operational frontier can be reached with an open-source AI forecast model (AIFS-Single) and relatively simple, cheap post-processing is significant for TC science, and points to agentic coding as a route to rapid exploration and progress in life-saving early-warning systems in other domains.
Anna Allen, Wessel P. Bruinsma, Michael Maier-Gerber +3
University of Cambridge · 2European Centre for Medium-Range Weather Forecasts