FuXi-Atmosphere: ML Weather Forecasting

• FuXi-Atmosphere is a family of ML systems for global weather prediction using cascaded transformer architectures tailored for different forecast ranges.
• It integrates advanced data assimilation and ensemble methods to enhance forecast accuracy, probabilistic outcomes, and subseasonal predictions.
• Specialized modules such as FuXi-Extreme and FuXi-TC, along with hybrid ML–physics systems, address extremes and mesoscale hazards to overcome traditional model limitations.

Searching arXiv for FuXi-family atmospheric papers to ground the article. FuXi-Atmosphere is not explicitly defined as a formal product name in the cited FuXi literature. The most evidence-grounded interpretation is that it denotes the atmospheric and weather-focused branch of the broader FuXi model family: a set of machine-learning systems for global weather forecasting, data assimilation, subseasonal prediction, ensemble forecasting, and phenomenon-specific refinement. In this reading, the core atmospheric line begins with FuXi as a cascaded 15-day global weather forecasting system, extends through FuXi-2.0 and FuXi Weather to higher-frequency and end-to-end analysis-to-forecast operation, and branches into probabilistic, subseasonal, hazard-oriented, and hybrid ML–physics systems (Chen et al., 2023).

1. Terminological scope and family structure

The FuXi literature consistently names concrete systems such as FuXi, FuXi Weather, FuXi-DA, FuXi-2.0, FuXi-ENS, FuXi-S2S, FuXi-Extreme, FuXi-TC, and FuXi-Nowcast, rather than a formally defined umbrella called “FuXi-Atmosphere.” The closest explicit framing appears in descriptions that place these models within a broader FuXi family of atmospheric prediction models and identify FuXi Weather as the closest documented counterpart to an atmospheric FuXi system. This suggests that “FuXi-Atmosphere” is best treated as a convenient umbrella label for the atmospheric/weather forecasting line, not as an officially introduced model name (Sun et al., 2024).

Within that atmospheric line, the systems organize around forecast range, operational role, and target phenomenon. The original FuXi addresses global deterministic medium-range weather forecasting to 15 days. FuXi-2.0 extends that line to 1-hourly global weather forecasts and a broader variable inventory, while also adding ocean-surface outputs in a joint atmosphere–ocean-surface configuration. FuXi Weather and FuXiWeather2 move further toward a full operational loop by combining learned data assimilation with learned forecasting. FuXi-ENS provides medium-range ensemble prediction, and FuXi-S2S extends the family into subseasonal-to-seasonal forecasting. Separate modules then target specific deficiencies of base global models, such as extreme-event smoothing in FuXi-Extreme, tropical-cyclone intensity in FuXi-TC, and convective nowcasting in FuXi-Nowcast. Hybrid systems such as FuXi-SHTM integrate FuXi’s large-scale atmospheric forecasts with regional physics-based typhoon models (Zhong et al., 2024).

A useful way to understand the family is as a layered atmospheric stack. The base layer is a multivariable global atmospheric forecast model. Around that base sit an initialization layer (FuXi-DA, FuXi Weather, FuXiWeather2), a probabilistic layer (FuXi-ENS, FuXi-S2S), specialized hazard refiners (FuXi-Extreme, FuXi-TC, FuXi-Nowcast), and hybrid regional downscaling or ML–physics fusion systems (FuXi-SHTM, FuXi-CFD). This layered interpretation is an inference from the naming and functional division reported across the cited works.

2. Core atmospheric forecasting architecture and state representation

The foundational atmospheric model is FuXi, described as “a cascade machine learning forecasting system for 15-day global weather forecast” operating at $0.25^\circ$ spatial resolution and 6-hourly temporal resolution. It predicts 70 variables: 5 upper-air variables across 13 pressure levels—geopotential $(Z)$, temperature $(T)$, u wind $(U)$, v wind $(V)$, and relative humidity $(R)$—plus 5 surface variables: $T2M$, $U10$, $V10$, $MSL$, and $(Z)$0. Its input tensor is $(Z)$1, corresponding to two preceding time steps, all forecast variables, and the global $(Z)$2 grid (Chen et al., 2023).

The central design principle in the original FuXi is the cascade. Rather than using one autoregressive model over the entire 15-day horizon, the system uses FuXi-Short for 0–5 days, FuXi-Medium for 5–10 days, and FuXi-Long for 10–15 days. The paper explicitly argues that using a single model is insufficient for achieving the best performance for both short and long lead times, because of accumulation error during long autoregressive rollouts. This cascade therefore specializes models to distinct forecast windows and trains later-stage models on forecast-generated intermediate states rather than clean reanalysis states (Chen et al., 2023).

Architecturally, the original FuXi combines cube embedding, a U-Transformer, and a fully connected layer. The input is compressed by a 3D convolution with kernel and stride $(Z)$3 into a latent field with channel dimension 1536, and the backbone uses 48 repeated Swin Transformer V2 blocks inside a U-shaped structure. The paper gives the attention form as

$(Z)$4

with $(Z)$5 as relative position bias and $(Z)$6 as a learnable scalar (Chen et al., 2023).

Later atmospheric models preserve the same broad state representation while changing cadence, variable inventory, or training regime. FuXi-2.0 uses a two-stage forecasting system: a primary 6-hourly model and a secondary 1-hourly interpolation model. Its tensor shape is $(Z)$7, reflecting a larger variable set that includes the atmospheric core plus sea-surface temperature and ocean-wave outputs. The atmospheric subset remains structured around 5 variables at 13 pressure levels plus a larger set of practical surface fields such as cloud cover and radiation products. The backbone is patch-based and transformer-based, with 30 consecutive Swin Transformer blocks and window size $(Z)$8, and the hourly model is organized into 5-block subsets, each followed by a transposed convolution layer, to emit 6 consecutive hourly forecasts within a 6-hour window (Zhong et al., 2024).

The family therefore exhibits a stable atmospheric design philosophy: global multivariable state prediction on regular latitude–longitude grids, pressure-level channels rather than explicit volumetric operators, transformer-heavy spatiotemporal encoding, and autoregressive rollout. A plausible implication is that “FuXi-Atmosphere” refers less to one immutable architecture than to a design program centered on learned evolution of a unified atmospheric state.

3. Data assimilation and the move to end-to-end atmospheric systems

A major transition in the FuXi atmospheric line is the shift from forecast-only systems initialized by conventional analyses to learned assimilation-plus-forecast pipelines. FuXi-DA is explicitly presented as the data-assimilation counterpart to the FuXi forecast model. It takes a 6-hour FuXi forecast as the background $(Z)$9, combines it with real satellite observations, and produces an analysis $(T)$0 intended both to reduce current-state error and to improve subsequent FuXi forecasts. The paper starts from the standard 3D-Var equation

$(T)$1

but replaces explicit observation operators and covariance matrices with learned encoders, multimodal fusion, and an analysis-increment formulation (Xu et al., 2024).

The demonstrated observation source in FuXi-DA is AGRI satellite data from Fengyun-4B. The atmospheric state contains 65 upper-air channels and 5 surface variables, for 70 variables total, on a $(T)$2 global grid. The architecture uses three U-Net branches: a background information flow, an observation information flow, and a mixed information flow, with learned fusion modules mediating interaction between them. Its primary supervised objective is a latitude-weighted $(T)$3 loss, and its defining system-level feature is a multi-time-step loss in which the FuXi-DA analysis is fed through FuXi and optimized jointly over future forecast steps: $(T)$4 with $(T)$5 in the study. This directly ties data assimilation to downstream atmospheric forecast skill (Xu et al., 2024).

FuXi Weather goes further by integrating satellite data preprocessing, FuXi-DA, and the FuXi forecast model into a 6-hourly DA and forecasting cycle. It uses a deliberately limited set of observing platforms—microwave sounders on FY-3E, Metop-C, and NOAA-20, plus GNSS radio occultation—and claims to achieve all-grid, all-surface, all-channel, and all-sky DA and forecasting. The underlying forecast state remains the 70-variable atmospheric cube used in FuXi, but the system closes the loop from raw observations to analyses to 10-day forecasts. Its headline benchmark is that the skillful lead time of $(T)$6, defined by $(T)$7, increases from 9.25 days for ECMWF HRES to 9.5 days for FuXi Weather when background forecasts are included in assimilation (Sun et al., 2024).

FuXiWeather2 generalizes this into a unified end-to-end atmospheric state estimation and forecasting system. It defines an assimilation module

$(T)$8

and a forecast module

$(T)$9

with $(U)$0 hours, and trains them in a recursively unrolled closed loop. Its stated motivation is to avoid prior ML systems acting merely as “reanalysis emulators.” The training objective combines a state-space loss against ERA5 and an observation-space loss against real in-situ observations: $(U)$1 The paper reports that its analysis fields surpass the NCEP-GFS across most variables and show superior accuracy over both ERA5 and the ECMWF-HRES system in lower-tropospheric and surface variables, while deterministic forecasts exceed HRES in 91\% of evaluated metrics (Xu et al., 16 Mar 2026).

Taken together, these works indicate that the atmospheric FuXi program is no longer limited to learned forecast dynamics. It now includes a learned analysis operator, end-to-end cycling, and explicit handling of train–deployment mismatch in self-generated backgrounds. This suggests a transition from “ML weather model” to “ML atmospheric system.”

4. Temporal extension, uncertainty, and coupled forecasting

The atmospheric FuXi line extends along two major axes beyond deterministic medium-range prediction: uncertainty quantification and longer-range forecasting.

For medium range, FuXi-ENS provides 6-hourly global ensemble weather forecasts up to 15 days at $(U)$2 resolution. It predicts 78 variables total, using 5 upper-air variables at 13 pressure levels plus 13 surface variables, and it frames its main innovation as a VAE-inspired perturbation model coupled to a transformer-based forecast model. The training loss replaces the standard $(U)$3 VAE objective with a CRPS + KL objective,

$(U)$4

with $(U)$5, in order to directly optimize probabilistic forecast quality. The paper reports that FuXi-ENS outperforms ensemble forecasts from ECMWF on 98.1\% of 360 variable and forecast lead time combinations on CRPS, but also notes a clear limitation: the ensemble is underdispersive, with SSR < 1 across the 15-day period (Zhong et al., 2024).

For subseasonal forecasting, FuXi-S2S extends the family to global daily mean forecasts up to 42 days on a $(U)$6 grid. It predicts 76 variables, comprising 5 upper-air variables at 13 pressure levels plus 11 surface variables, and uses a probabilistic latent perturbation architecture built on an enhanced FuXi backbone. Its input tensor shape is $(U)$7, and it uses autoregressive training, multi-step loss, curriculum schedule, and explicit lead-time conditioning. The paper reports that FuXi-S2S outperforms the ECMWF S2S system in deterministic and probabilistic skill for precipitation and OLR, and extends skillful MJO prediction from 30 days to 36 days in one evaluation setup (Chen et al., 2023).

A later diagnostic paper focuses on the Madden–Julian Oscillation and attributes part of FuXi-S2S’s subseasonal improvement to better representation of low-frequency background moisture gradients over the tropical western Pacific. In that study, FuXi-S2S extends skillful MJO prediction from 28 days to 35 days relative to ECMWF S2S under the ACC = 0.5 criterion. The proposed mechanism centers on more accurate prediction of the area-averaged meridional gradient of low-frequency background moisture, which then improves meridional moisture transport and downstream convection forecasts. This suggests that the subseasonal branch is not only statistically skillful but can also be interrogated in physically interpretable terms (Cao et al., 22 Aug 2025).

Coupling is introduced explicitly in FuXi-2.0, which the paper describes as an “atmosphere-ocean coupled model” in the sense of joint prediction of atmospheric variables plus SST and six wave variables. The paper is careful to note that the coupling is implemented through a joint multivariate state vector and shared model, not through explicit exchange equations or a traditional modular coupler. Its practical evidence for the value of this coupled configuration is that tropical-cyclone intensity forecasts improve over FuXi-1.0, while track skill remains similar (Zhong et al., 2024).

These developments show that the atmospheric FuXi line is not monolithic. It spans deterministic and ensemble forecasting, medium-range and subseasonal regimes, and atmosphere-only and atmosphere–ocean-surface state vectors. A plausible implication is that “FuXi-Atmosphere” refers to the atmospheric forecasting core across these variants, even when adjacent components such as ocean-surface outputs are added.

5. Specialized hazard and application modules

A conspicuous feature of the FuXi ecosystem is the use of specialized downstream modules to compensate for known failure modes of global deterministic atmospheric forecasts.

FuXi-Extreme addresses one such failure mode directly: the tendency of ML weather forecasts to become increasingly smooth with lead time and therefore underestimate extremes. It is described as a frozen FuXi-Short base model plus a denoising diffusion probabilistic model (DDPM) that refines only surface variables—$(U)$8, $(U)$9, $(V)$0, $(V)$1, and $(V)$2—within the 0–5 day range. Instead of standard $(V)$3-prediction, it predicts the clean target directly with

$(V)$4

The paper reports that FuXi-Extreme has the highest SEDI for $(V)$5, $(V)$6, and $(V)$7 for all lead times, even though base FuXi remains slightly better on global bulk RMSE and ACC (Zhong et al., 2023).

FuXi-TC addresses a narrower but operationally important weakness: tropical-cyclone intensity underestimation stemming from ERA5-based training targets. It is a diffusion-based generative forecasting framework trained not on ERA5 but on WRF simulations driven by FuXi background fields. The framework is explicitly regional, covering the western North Pacific over a $(V)$8 grid at $(V)$9 resolution, with a 120 h forecast horizon. The paper states that across 21 TCs in 2024, FuXi-TC substantially reduces RMSE relative to FuXi and ERA5, achieving 5-day intensity forecast skill comparable to HRES, while running in 2.0 seconds on a single NVIDIA A100 GPU compared with 83.0 minutes for WRF on 32 Intel Xeon Platinum 8369B CPUs (Guo et al., 22 Aug 2025).

FuXi-Nowcast appears to be a nowcasting-oriented member of the broader FuXi family, but the supplied supplementary-information stub contains almost no technical content. The only evidence-grounded conclusion is that a work named FuXi-Nowcasting exists and is presented as supplementary information to a paper of that name. The naming suggests a nowcasting branch within the FuXi atmospheric ecosystem, but the provided material does not substantiate architectural, variable, or benchmarking details. Any stronger claim about its relationship to FuXi-2.0 or to a larger “FuXi-Atmosphere” framework would exceed the supplied evidence (Chen et al., 3 Dec 2025).

At the application layer, FuXi-CFD is positioned as a terrain-aware downscaling system that complements kilometer-scale AI atmospheric forecasts. It ingests coarse 10 m wind components $(R)$0 on a $(R)$1 grid at 1 km spacing together with 30 m elevation and roughness, and predicts full 3D wind structure—$(R)$2, $(R)$3, $(R)$4, and $(R)$5—at 30 m horizontal resolution and 27 vertical levels up to approximately 300 m AGL. The paper does not describe it as a native component of the core FuXi atmospheric model, but rather as a downstream extension that can consume the kind of near-surface wind outputs that FuXi-like weather systems produce (Lin et al., 19 May 2025).

These modules collectively indicate a modular strategy: use a global atmospheric model for synoptic evolution, then attach task-specific refiners where deterministic reanalysis-trained forecasts are weakest, especially for extremes, convection, terrain-mediated wind fields, and TC inner-core structure.

6. Hybrid ML–physics systems, performance profile, and open issues

A recurrent conclusion in the FuXi literature is that the atmospheric core is especially strong in large-scale flow prediction, while weaknesses persist for intensity, inner-core structure, cloud microphysics, and some extremes. This assessment is made most explicitly in the tropical-cyclone hybrid papers.

FuXi-SHTM couples FuXi’s global forecast with the Shanghai Typhoon Model (SHTM), a WRF-based regional model. The fusion mechanism is spectral nudging, described conceptually as

$(R)$6

where only the large-scale component of the FuXi–SHTM difference is used. In the operational configuration, only u, v, and virtual temperature are nudged; nudging is not applied below 850 hPa; and the cutoff wavelength is around 1000 km. The paper’s practical conclusion is that FuXi supplies the evolving synoptic-scale steering and thermal environment, while SHTM provides mesoscale and storm-core physics (Niu, 1 Mar 2025).

A broader 2024 evaluation reports that, relative to FuXi alone, FuXi-SHTM reduces typhoon track forecast errors by 16.5\% and 5.2\% at lead times of 72 h and 120 h, respectively, and reduces intensity forecast errors by 59.7\% and 47.6\%. The same study reports that FuXi-SHTM simulates cloud structures more realistically than SHTM and achieves superior 10-m wind fields compared with FuXi and SHTM. Increasing the physical-model resolution from 9 km to 3 km improves intensity forecasts further, highlighting that the hybrid system’s remaining bottlenecks are strongly tied to high-resolution physics rather than synoptic steering (Niu et al., 29 Apr 2025).

This hybrid evidence clarifies the present status of the atmospheric FuXi program. The global atmospheric model is competitive or superior for many standard variables and lead times, but it is not yet a universal replacement for physically explicit regional hazard models. The literature repeatedly identifies several persistent issues: dependence on ERA5 or reanalysis-like targets, underrepresentation of extremes, limited physical interpretability in some settings, incomplete uncertainty calibration, and incomplete coverage of operational variables such as visibility. Some papers also note evaluation caveats, such as unequal ensemble size in FuXi-S2S versus ECMWF S2S comparisons, or the absence of direct comparisons against mature operational DA systems in the FuXi-DA work (Chen et al., 2023).

At the same time, the family’s performance claims are substantial. The original FuXi is described as the first ML-based weather forecasting system to achieve comparable forecast performance to ECMWF ensemble mean in 15-day forecasts, with skillful lead times of 10.5 days for $(R)$7 and 14.5 days for $(R)$8 under ACC > 0.6 (Chen et al., 2023). FuXi-2.0 reports more accurate 1-hourly forecasts than ECMWF HRES for key variables used in wind energy, solar energy, and aviation, and it documents better tropical-cyclone intensity than the atmosphere-only FuXi-1.0 (Zhong et al., 2024). FuXi Weather and FuXiWeather2 further suggest that end-to-end AI atmospheric systems can approach or exceed operational NWP skill while drastically simplifying or bypassing conventional variational DA infrastructure (Sun et al., 2024).

In encyclopedic terms, the most defensible synthesis is that FuXi-Atmosphere denotes an inferred atmospheric model family rather than a single formally named system. Its core characteristics are global multivariable atmospheric state prediction, transformer-based autoregressive rollout, progressive integration of learned assimilation, expansion toward ensemble and subseasonal prediction, and a modular architecture in which specialized refiners or regional hybrids correct weaknesses in extremes and mesoscale hazards. The current literature suggests a trajectory from deterministic global atmospheric forecasting toward a fuller operational AI atmosphere system, while also indicating that hybrid ML–physics coupling remains important for the last-mile prediction of high-impact weather.