FuXi-TC: Diffusion-Based Tropical Cyclone Forecast

• FuXi-TC is a diffusion-based generative forecasting framework that improves tropical cyclone intensity predictions using conditional denoising diffusion.
• It integrates FuXi’s large-scale track information with high-resolution WRF intensity fields to correct systematic biases in traditional forecasts.
• The system reduces 5-day max wind speed RMSE by up to 53% and delivers forecasts in seconds, offering significant operational efficiency gains.

Searching arXiv for FuXi-TC and closely related FuXi tropical-cyclone papers. FuXi-TC is a diffusion-based generative forecasting framework for tropical cyclone prediction that combines the track prediction skill of the FuXi model with the intensity representation of numerical weather prediction simulations. It is trained on outputs from the Weather Research and Forecasting model driven by FuXi background fields, with the stated aim of correcting systematic biases and reconstructing realistic tropical-cyclone structures without retraining global models. In evaluations across 21 tropical cyclones in 2024, it substantially reduces root mean square error in 5-day intensity forecasts relative to both FuXi and ERA5, achieves skill comparable to the high-resolution deterministic forecasts of the European Centre for Medium-Range Weather Forecasts, and produces forecasts within seconds rather than the tens of minutes required by regional dynamical downscaling (Guo et al., 22 Aug 2025).

1. Forecasting problem and conceptual role

Tropical cyclones are described as among the most devastating natural hazards, while their intensity remains notoriously difficult to predict. Within the framing of FuXi-TC, two limitations motivate the method. Numerical weather prediction models are constrained by both computational demands and intrinsic predictability, whereas state-of-the-art deep learning-based weather forecasting models tend to underestimate tropical-cyclone intensity because of biases in reanalysis-based training data. FuXi-TC is positioned specifically against that intensity-underestimation problem rather than as a replacement for the global FuXi forecast model (Guo et al., 22 Aug 2025).

The framework is explicitly formulated as a bias-correcting and structure-reconstructing system. Its inputs preserve FuXi’s large-scale background and track-relevant environment, while its targets inherit the high-resolution intensity structure of Weather Research and Forecasting simulations. The paired dataset $\{(c,x_0)\}$ is therefore described as fusing FuXi track skill with WRF’s intensity realism. This suggests that FuXi-TC is designed as a learned regional refinement stage: it leaves the global model intact, but replaces the weakly expressed storm-core intensity signal with a generative reconstruction conditioned on FuXi’s synoptic context (Guo et al., 22 Aug 2025).

A recurrent misunderstanding in the FuXi literature is to equate all post-processing with simple sharpening. FuXi-TC is not presented as a generic image-enhancement layer. It is trained on physics-based regional simulations, uses explicit meteorological conditioning, and targets realistic eye-wall wind speeds and pressure gradients rather than only visual detail (Guo et al., 22 Aug 2025).

2. Diffusion formulation

FuXi-TC treats the problem of “correcting” FuXi intensity forecasts as a conditional denoising diffusion process. In its notation, $x_0$ is the high-resolution WRF field, described as the “ground truth” intensity structure, and $c$ denotes the FuXi forecast fields that provide the large-scale background. A standard DDPM is run on $x_0$ while conditioning on $c$ (Guo et al., 22 Aug 2025).

The forward process adds Gaussian noise in $T$ discrete steps according to a predefined noise schedule $\{\beta_t\}$, while the reverse process is parameterized as

$$p_{\theta}(x_{t-1}\mid x_t,c) = \mathcal{N}\bigl(x_{t-1};\mu_{\theta}(x_t,t,c),\,\Sigma_{\theta}(t)\bigr).$$

In practice, $\Sigma_{\theta}(t)$ is fixed or tied to $\beta_t$, and the learning objective focuses on $x_0$0. Training adopts the “$x_0$1-prediction” re-parameterization and minimizes the mean-squared error between true noise and network-predicted noise. The diffusion configuration uses $x_0$2 timesteps with a linear $x_0$3-schedule from $x_0$4 to $x_0$5 (Guo et al., 22 Aug 2025).

The significance of this formulation lies in what is being modeled conditionally. The conditioning variables are not abstract embeddings but meteorological fields that encode steering flows, environmental shear, surface pressure, near-surface winds, and precipitation. The diffusion model is therefore tasked with generating physically plausible mesoscale intensity structure under a given large-scale environment rather than forecasting the entire atmosphere from scratch. This suggests a division of labor in which FuXi supplies the global trajectory and synoptic setting, while the generative model restores the storm-core intensity patterns that are otherwise smoothed or biased low (Guo et al., 22 Aug 2025).

3. Architecture and conditioning strategy

FuXi-TC is implemented as a lightweight U-Net from the HuggingFace Diffusers library, modified to ingest multiple meteorological variables together with noise level and timestep embeddings. The architecture contains 3 down-blocks and 3 up-blocks with channel widths $x_0$6. Each block uses two $x_0$7 convolution layers, SiLU activations, and GroupNorm. The bottleneck contains 12 global self-attention layers, included to capture long-range dependencies such as spiral rainbands and convective asymmetries. Timestep embeddings are added via FiLM to all convolution blocks, and the background fields $x_0$8 are concatenated to the noisy input $x_0$9 at each stage (Guo et al., 22 Aug 2025).

The conditioning fields are FuXi-2.0 forecasts used as static background inputs. At each forecast lead time, the model stacks 2D maps of Z500, MSLP, T2M, U10M, V10M, WS10M, and total precipitation. These fields are described as supplying the synoptic steering flows and environmental shear that govern tropical-cyclone track and large-scale structure. The target variables extracted from WRF use the same channels, thereby defining a direct field-to-field mapping between FuXi background state and WRF-resolved storm intensity structure (Guo et al., 22 Aug 2025).

WRF is run at $c$0 resolution over $c$1–$c$2 E and $c$3–$c$4 N with spectral nudging to FuXi large scales. In this setup, WRF is said to produce realistic eye-wall wind speeds and pressure gradients. The model architecture is therefore not only a generic image denoiser; it is a conditional meteorological generator whose inputs and outputs are aligned channelwise across dynamically related forecast fields (Guo et al., 22 Aug 2025).

4. Training data pipeline and optimization

The training pipeline uses FuXi-2.0 global forecasts from 2019–2023 both as background conditioning and to initialize or ring WRF boundary conditions. Regional WRF v4.3 runs from 2019–2023 are driven by FuXi initial and boundary fields, with spectral nudging above 850 hPa. Outputs at each 6 h lead time form $c$5, and testing is conducted on withheld 2024 typhoons comprising 21 storms for final evaluation (Guo et al., 22 Aug 2025).

This construction is described as a form of bias-correction and physics-based synthesis. Running WRF on FuXi fields is said to correct the underestimation of peak winds present in both ERA5 and FuXi. The resulting paired dataset therefore transfers information from a physics-based regional model into a trainable generative corrector, while preserving consistency with the upstream FuXi forecast. A plausible implication is that the framework uses WRF less as an operational inference engine than as a mechanism for manufacturing a higher-fidelity supervision signal tailored to FuXi’s own forecast distribution (Guo et al., 22 Aug 2025).

The optimization setup is specified in detail. Training uses 60,000 iterations on two NVIDIA A100 GPUs with batch size 4, split as 2 per GPU. The optimizer is AdamW with $c$6, $c$7, learning rate $c$8, and weight decay $c$9. Gradient norm clipping is set at $x_0$0, and an exponential moving average of weights is used for sampling stability (Guo et al., 22 Aug 2025).

5. Quantitative skill and runtime

The central reported result is the reduction of 5-day intensity error across 21 Western North Pacific typhoons in 2024, spanning April to October. At 120 h lead time, the root mean square error of max-WS10M forecasts is approximately $x_0$1 m/s for ERA5, $x_0$2 m/s for FuXi, $x_0$3 m/s for FuXi-TC, $x_0$4 m/s for ECMWF-HRES at $x_0$5 deterministic resolution, and $x_0$6 m/s for WRF at $x_0$7 downscaled resolution (Guo et al., 22 Aug 2025).

System	120 h max-WS10M RMSE
ERA5	$x_0$8 m/s
FuXi	$x_0$9 m/s
FuXi-TC	$c$0 m/s
ECMWF-HRES	$c$1 m/s
WRF	$c$2 m/s

Relative to FuXi and ERA5 at 120 h, FuXi-TC reduces RMSE by about 46% and about 53%, respectively. At 72 h and 96 h leads, similar reductions of roughly 40–50% are reported. The 5-day intensity skill is described as statistically indistinguishable from the 9 km ECMWF deterministic run with $c$3, even though FuXi-TC is trained only on FuXi plus WRF fields. The reduction in RMSE versus FuXi is reported as robust across all 21 storms under a paired $c$4-test with $c$5 (Guo et al., 22 Aug 2025).

The computational contrast is similarly explicit. Runtime per 120 h forecast is approximately 83 min for WRF on 32 Intel Xeon CPUs and approximately 2 s for FuXi-TC on 1 NVIDIA A100 GPU, corresponding to a speed-up greater than 2,400 times over WRF and greater than 100 times over a typical NWP regional nest (Guo et al., 22 Aug 2025).

System	Runtime per 120 h forecast
WRF (32 × Intel Xeon CPUs)	$c$6 min
FuXi-TC (1 × NVIDIA A100 GPU)	$c$7 s

These numbers define the framework’s operational claim. Because inference is 2 s per forecast, the paper states that sub-minute generation of 1000-member intensity ensembles on a single GPU node becomes possible, along with hourly or sub-hourly update cycles. By combining FuXi’s track forecasts with FuXi-TC’s rapid, NWP-quality intensity fields, meteorological services are described as being able to deliver high-resolution wind and rain hazard guidance in seconds, including probabilistic forecasts of wind gusts, surge potential, and rainfall extremes (Guo et al., 22 Aug 2025).

6. Relationship to adjacent FuXi systems and nomenclature

FuXi-TC emerges within a broader sequence of FuXi-based efforts addressing tropical cyclones and extremes, but it occupies a distinct methodological niche. “FuXi-Extreme” uses a DDPM to restore finer-scale detail in FuXi surface forecasts and improves extreme total precipitation, 10-meter wind speed, and 2-meter temperature; however, in tropical-cyclone evaluation it still lags HRES in intensity because it only enhances surface fields and does not recover the low-pressure core or eyewall winds completely (Zhong et al., 2023). “FuXi-ENS” introduces a learnable perturbation generator and shows lower track errors than ECMWF-ENS, but the paper states that intensity forecasts remain biased low in both systems (Liu et al., 27 Oct 2025). “FuXi_CNOP” adds Orthogonal Conditional Nonlinear Optimal Perturbations to FuXi and improves deterministic and probabilistic track forecasting, with CRPS improvements of up to 29.2% at 120 h, but that work is centered on ensemble track uncertainty rather than the diffusion-based reconstruction of storm intensity fields (Li et al., 26 Feb 2026).

Within the multimodal “FuXi-Uni” framework, the name FuXi-TC is also used for a tropical-cyclone forecasting and editing instantiation driven by language prompts. In that setting, for 20 Western North Pacific storms from May to October 2024, track MAE at 120 h is reported as approximately 123 km for FuXi-Uni original forecasts, 138 km for HRES, and 106 km after editing; the corresponding WS10M RMSE values at 120 h are approximately 14.50 m/s for FuXi-Uni original forecasts, 13.93 m/s for HRES, and 13.84 m/s for the edited FuXi-TC output (Yang et al., 4 Jan 2026). This suggests that the label “FuXi-TC” can refer either to the dedicated diffusion-based intensity-correction framework of (Guo et al., 22 Aug 2025) or to the tropical-cyclone editing component inside FuXi-Uni (Yang et al., 4 Jan 2026).

There is also an unrelated use of the same acronym in superconducting-circuit research, where “FuXi-TC” denotes a “fluxonium–transmon–fluxonium” coupling architecture for two-qubit gates rather than any weather-forecasting system (Ding et al., 2023). In arXiv practice, therefore, the term is not globally unique. In atmospheric-science usage, however, it refers primarily to the tropical-cyclone framework that integrates FuXi forecasts, WRF-derived supervision, and conditional diffusion to improve intensity prediction (Guo et al., 22 Aug 2025).