Pangu-Weather: Data-Driven Global Forecasting

• Pangu-Weather is a data-driven global weather forecasting system that employs a 3D transformer architecture with hierarchical temporal aggregation.
• It uses 43 years of ERA5 reanalysis data to train dedicated models for various lead times, achieving higher efficiency and competitive accuracy compared to ECMWF IFS.
• While excelling in medium-range deterministic forecasts and dynamical structure representation, it faces challenges like bias drift, forecast smoothness, and calibration limitations.

Pangu-Weather is a global, purely data-driven weather forecasting system trained on ERA5 reanalysis and formulated around a three-dimensional transformer architecture with hierarchical temporal aggregation. It operates on a $0.25^\circ\times0.25^\circ$ latitude–longitude grid, predicts upper-air and surface meteorological fields on 13 pressure levels, and composes dedicated $1\,\mathrm{h}$, $3\,\mathrm{h}$, $6\,\mathrm{h}$, and $24\,\mathrm{h}$ submodels to reach medium-range lead times. In its original presentation, it was reported to outperform ECMWF’s operational Integrated Forecasting System in latitude-weighted RMSE and ACC for all variables and all lead times from one hour to one week; later studies have situated it within a broader transition toward operationally relevant data-driven weather prediction, while also documenting bias drift, forecast smoothness, and variable-specific limitations (Bi et al., 2022, Ben-Bouallegue et al., 2023).

1. Position within data-driven weather forecasting

Pangu-Weather emerged in a period of rapid progress in global data-driven weather models trained end-to-end on reanalysis data. It is described as an “all-machine” forecasting system trained purely on ERA5, without explicit physics in the forecast model itself, and as part of an emerging ML-based forecasting paradigm in which state-of-the-art analyses or reanalyses provide the initial conditions and training targets (Bremnes et al., 2023, Ben-Bouallegue et al., 2023).

The system’s historical significance in the literature is tied to two related claims. First, the original paper presents Pangu-Weather as a “deep learning based system for fast and accurate global weather forecast,” emphasizing both accuracy and computational speed. Second, comparative studies in operational-like settings reported that deterministic forecasts from Pangu-Weather could match the quality of one of the leading global NWP systems, ECMWF IFS, for both global metrics and selected extremes, while requiring substantially lower computational cost (Bi et al., 2022, Ben-Bouallegue et al., 2023).

A central feature of this research trajectory is that Pangu-Weather is not treated merely as a pattern-recognition surrogate for conventional NWP. Subsequent work subjected it to operational station verification, probabilistic calibration, uncertainty quantification, compatibility tests with multiple operational analyses, and classical dynamical experiments. Taken together, these studies position Pangu-Weather as both a competitive forecasting model and an object of scientific scrutiny regarding what atmospheric structure and dynamics a purely data-driven system can represent (Bremnes et al., 2023, Bülte et al., 2024).

2. State representation and neural architecture

In the original formulation, Pangu-Weather represents the atmosphere and surface as a three-dimensional cubic tensor whose axes correspond to pressure level, latitude, and longitude. The upper-air input comprises 13 pressure levels, 721 latitudes, 1440 longitudes, and 5 variables: geopotential $Z$, specific humidity $Q$, temperature $T$, and zonal and meridional wind components. The surface input comprises one surface “level” with 4 variables—2 m temperature, 10 m $u$-wind, 10 m $v$-wind, and mean sea-level pressure—together with three constant land–sea/topography masks (Bi et al., 2022).

The architectural core is the “3D Earth-Specific Transformer” (3DEST), which is also summarized in later work as a 3D vision-transformer in a U-Net–style encoder–decoder arrangement. Patch embedding converts the upper-air and surface streams into latent volumes that are concatenated and processed by an encoder–decoder stack. In the original description, the encoder has 8 layers and the decoder has 8 layers; patch-merging and patch-expanding follow the Swin Transformer design. The full-resolution embedding dimension is $1\,\mathrm{h}$0, with 6 attention heads in full-resolution layers and 12 in downsampled layers, and the total number of trainable parameters is approximately 256 million (Bi et al., 2022).

A distinctive component is the Earth-Specific Positional Bias, $1\,\mathrm{h}$1, introduced into windowed multi-head self-attention:

$1\,\mathrm{h}$2

Unlike a standard relative positional bias, $1\,\mathrm{h}$3 depends on vertical and latitudinal window indices as well as intra-window offsets. This lets the model encode latitude- and level-dependent behavior such as differences between equatorial and polar circulation regimes or between upper- and lower-tropospheric structure (Bi et al., 2022).

Temporal rollout is not produced by a single recurrent step. Instead, Pangu-Weather trains four separate lead-specific models for $1\,\mathrm{h}$4 and composes them greedily at inference time. The original paper frames this as “hierarchical temporal aggregation,” introduced to reduce cumulative rollout error. A 71 h forecast, for example, is constructed by successive calls to the 24 h, 6 h, 3 h, and 1 h models rather than by repeated application of a single short-step model (Bi et al., 2022).

3. Training data, objective functions, and computational profile

The training corpus is ERA5 reanalysis on the native $1\,\mathrm{h}$5 grid with hourly temporal resolution. The original paper states that 43 years of hourly global weather data were downloaded from ERA5, with training on 1979–2017, validation on 2019, and testing on 2018, 2020, and 2021; secondary summaries also describe 39 years of ERA5 hourly fields used for parameter estimation in related evaluations (Bi et al., 2022, Bremnes et al., 2023).

The principal optimization target is mean squared error on forecast fields or next-step increments. One reported form is

$1\,\mathrm{h}$6

where $1\,\mathrm{h}$7 predicts the increment from analysis state $1\,\mathrm{h}$8 to verifying truth $1\,\mathrm{h}$9. Another equivalent training description emphasizes a cosine-latitude–weighted RMSE over grid points, variables, and leads, while noting that in practice the square loss is minimized during training (Bremnes et al., 2023, Ben-Bouallegue et al., 2023).

Optimization is reported with Adam in the original paper and with AdamW in later technical summaries. The original training schedule is 100 epochs, approximately 15 days on 192 NVIDIA V100 GPUs, with batch size 192, weight decay $3\,\mathrm{h}$0, and DropPath 0.2. A closely related later summary reports training over 39 years of data on 192 V100 GPUs for 16 days, again emphasizing the large offline training investment followed by very low inference cost (Bi et al., 2022, Bremnes et al., 2023).

The inference profile is central to Pangu-Weather’s scientific and operational interest. Once trained, a full 10 day forecast can be generated in seconds on a single GPU. One later comparison gives $3\,\mathrm{h}$1 for a full 10-day forecast at 28 km resolution on an NVIDIA A100, versus approximately 1 h wall-clock time for ECMWF IFS high resolution on a supercomputer CPU cluster. The same source reports approximately $3\,\mathrm{h}$2 FLOPs for Pangu-Weather versus approximately $3\,\mathrm{h}$3 FLOPs for IFS per 10-day forecast, i.e. a two-order-of-magnitude gap (Ben-Bouallegue et al., 2023).

4. Verification framework and deterministic forecast skill

The core deterministic verification metrics used across studies are latitude-weighted RMSE and ACC, with additional use of bias, mean absolute error, and CRPS in station-based and probabilistic evaluations. Representative definitions include

$3\,\mathrm{h}$4

and

$3\,\mathrm{h}$5

where $3\,\mathrm{h}$6, $3\,\mathrm{h}$7, and $3\,\mathrm{h}$8 denote forecast, verifying analysis or observation, and climatology, respectively (Ben-Bouallegue et al., 2023).

On the 2018 test set reported in the original paper, Pangu-Weather was stated to beat both FourCastNet and operational ECMWF IFS across every variable and every lead time from 1 h to 168 h. The headline numbers at 5 days include: Z500 RMSE 296.7 for Pangu-Weather versus 333.7 for IFS; T850 RMSE 1.79 K versus 2.06 K; 2 m temperature RMSE 1.53 K versus 1.75 K; and 10 m wind RMSE approximately 2.53 m/s versus approximately 2.90 m/s. The same paper describes a “forecast-time gain” of 10–24 h relative to competitors at equivalent RMSE (Bi et al., 2022).

Later operational-like verification qualifies this picture. In a Northern Hemisphere evaluation against IFS analysis and SYNOP observations, Pangu-Weather was found to be very close to IFS for global deterministic skill. For Z500, it matches IFS by day 3 at approximately 25 m RMSE and achieves ACC $3\,\mathrm{h}$9 at day 2 and 0.85 at day 5; for T850, RMSE is approximately 1.5 K at day 3 and within 0.02 ACC units of IFS out to day 7. For European 2 m temperature, summer RMSE is approximately 1.8 K at day 5 versus approximately 1.9 K for IFS, while winter RMSE is approximately 2.2 K at day 6 versus approximately 2.0 K for IFS. The same study identifies increasing forecast smoothness and bias drift with lead time as current drawbacks of ML-based forecasts (Ben-Bouallegue et al., 2023).

Station-focused verification over Norway provides a more localized view. Using 183 SYNOP stations with observations every 6 h and daily forecasts issued at 00 UTC for lead times +6 to +60 h, Pangu-Weather, ECMWF HRES, ECMWF ENS, and MEPS were compared for 2 m temperature and 10 m wind speed. The results show that the global models are on the same level, with Pangu-Weather being slightly better than the ECMWF models for temperature and slightly worse for wind speed, while the high-resolution regional MEPS model clearly provided the best forecasts for both parameters (Bremnes et al., 2023).

At a representative +24 h lead time in that Norwegian study, Pangu-Weather’s raw 2 m temperature MAE and CRPS were 1.35 °C and 1.40 °C, versus 1.45 °C and 1.50 °C for HRES. For 10 m wind speed, Pangu-Weather’s raw MAE and CRPS were 1.75 m/s and 1.80 m/s, versus 1.70 m/s and 1.75 m/s for HRES. MEPS remained best, with raw MAE 1.10 °C for temperature and 1.40 m/s for wind speed at the same lead time (Bremnes et al., 2023).

5. Probabilistic calibration, uncertainty quantification, and operational compatibility

Although Pangu-Weather is fundamentally deterministic in its original form, several studies turn it into a probabilistic forecasting system through statistical post-processing, machine-learning calibration, or perturbed initial conditions. In the Norwegian station study, probabilistic calibration was implemented with Bernstein Quantile Networks (BQN), trained on 2021 data with inputs including raw forecasts, lead time, day-of-year sine and cosine terms, station embeddings, and, for wind, the zonal and meridional wind components plus magnitude. The model used Bernstein polynomials of degree $6\,\mathrm{h}$0, two dense layers, ELU activations, softplus constraints for monotonic quantiles, Adam optimization, and early stopping at 200 epochs (Bremnes et al., 2023).

That post-processing considerably improved forecast quality for all models. For Pangu-Weather at +24 h, calibrated 2 m temperature MAE and CRPS dropped from 1.35 °C and 1.40 °C to 0.75 °C and 0.80 °C; calibrated 10 m wind speed MAE and CRPS dropped from 1.75 m/s and 1.80 m/s to 1.05 m/s and 1.10 m/s. The study reports approximately 45% improvement for Pangu-Weather, HRES, and ENS in temperature-related metrics, compared with approximately 35% for MEPS, and emphasizes that the overall model ranking remained essentially unchanged after calibration (Bremnes et al., 2023).

A broader European uncertainty-quantification study compared three initial-condition ensemble strategies—Gaussian-noise perturbations, random-field perturbations, and IFS-based perturbations—with two post-hoc methods: EasyUQ based on isotonic distributional regression, and a distributional regression network (DRN) assuming a Gaussian predictive distribution. For T2m over Europe, mean CRPS at lead times 6–48 h was 0.57 for the ECMWF ensemble, 0.50 for random-field perturbations, 0.43 for EasyUQ, and 0.41 for DRN; at 48–120 h it was 0.75, 0.74, 0.69, and 0.67, respectively. Beyond 120 h, ECMWF regained a small edge for T2m, with CRPS 1.05 versus 1.10 for DRN and 1.13 for EasyUQ. PIT histograms and spread–skill plots showed GNP and IFSP to be strongly underdispersive, RFP underdispersive but less so, and DRN/EasyUQ nearly calibrated (Bülte et al., 2024).

Operational integration also depends on initialization compatibility. A case-study paper tested Pangu-Weather with ERA5, ECMWF-IFS, NOAA-GFS, CMA-GRAPES, YHGSM, and a hybrid “ecmfpadGFS” initial condition. It reported that Pangu-Weather initialized from GFS systematically outperformed native GFS forecasts by approximately 5–10% reduction in RMSE and approximately 0.02–0.05 increase in ACC across most fields in a representative 72 h case, and that local improvement in initial-condition quality yielded the largest regional skill gains. The same study details the preprocessing steps required for such use: bilinear interpolation to the $6\,\mathrm{h}$1 target grid, log-pressure interpolation to the 13 target levels, and harmonization of variable units and definitions (Cheng et al., 2023).

6. Dynamical fidelity, extreme events, limitations, and derivative lines of work

A major question about Pangu-Weather is whether its performance reflects learned atmospheric dynamics or merely statistical pattern matching. Classical dynamical experiments suggest that the model encodes a substantial amount of physically meaningful structure. When a steady tropical heat source was added to the model physics, it generated a Matsuno–Gill-type response near the forcing and a stationary extratropical wave train. A localized disturbance to the winter-mean North Pacific jet produced realistic extratropical cyclones and fronts, including spontaneous emergence of polar lows. Geostrophic adjustment experiments showed relaxation from a 500 hPa height perturbation to a wind–pressure-balanced state over approximately 6 h. In tropical-cyclone experiments, low-pressure perturbations intensified into Atlantic hurricanes when the initial amplitude exceeded about 5 hPa, while setting humidity to zero eliminated hurricane development (Hakim et al., 2023).

Even so, event-based validation reveals variable- and impact-specific shortcomings. In a comparison of ML models and ECMWF HRES on the 2021 Pacific Northwest heatwave, the 2023 South Asian humid heatwave, and the 2021 North American winter storm, Pangu-Weather locally achieved similar accuracy to HRES on the Pacific Northwest event but underperformed when errors were aggregated over space and time. For the humid heatwave, because Pangu-Weather does not predict surface relative humidity, the study used RH at 1000 hPa as a substitute and found a large negative heat-index bias, with danger levels over Bangladesh underestimated. By contrast, on the North American winter storm, Pangu-Weather forecast the compound wind-chill event substantially better than HRES, with RMSE approximately 5–8 K versus approximately 10–12 K at 2–3 day lead times at College Station (Pasche et al., 2024).

The documented limitations are consistent across several strands of the literature. Oversmoothing of mesoscale features, systematic bias drift with lead time, poor representation of tropical-cyclone intensity and structure at 28 km resolution, and the absence of a precipitation prognostic are explicitly identified in one operational comparison (Ben-Bouallegue et al., 2023). The Norwegian station study associates some of Pangu-Weather’s raw biases with its coarse $6\,\mathrm{h}$2 resolution, which cannot resolve local orographic effects or sharp wind-speed gradients (Bremnes et al., 2023).

Later tropical-wave diagnostics refine that picture further. In free-run output, Pangu-Weather shows clear intraseasonal Kelvin-wave and Rossby-wave spectral bands, with a Kelvin-wave phase speed estimated at approximately 11.9 m/s and composite structures that reproduce lower/upper-level convergence–divergence dipoles and vertical tilts. The same analysis, however, finds that equatorial Rossby-wave temperature anomalies have the wrong sign alignment relative to vertical motion and that Pangu-Weather is among the weakest of the compared AI models at reproducing the Rossby wave’s full baroclinic structure. This suggests that physically important large-scale variability can be partly captured while remaining inconsistent across dynamical and thermodynamical fields (Jalan et al., 10 Jul 2025).

The architecture has also seeded derivative developments. A regional weather forecasting system over India modified the Pangu-Weather architecture for CPU-only training on IMDAA reanalysis data and found that a hierarchical forecasting approach again performed best among static, autoregressive, and hierarchical rollouts. Separately, PW-FouCast used Pangu-Weather forecasts as spectral priors for precipitation nowcasting, treating them as a source of large-scale meteorological context in a Fourier-domain fusion framework (Choudhury et al., 17 Mar 2025, Qin et al., 23 Mar 2026).

Across these studies, Pangu-Weather appears neither as a direct replacement for all operational NWP components nor as a narrow benchmark artifact. The published record instead supports a more specific characterization: a fast global deterministic forecaster with strong medium-range skill, broad compatibility with modern verification frameworks, substantial learned dynamical structure, and clear deficiencies in calibration, small-scale detail, and selected variables that motivate post-processing, hybridization, and architecture-specific extensions (Ben-Bouallegue et al., 2023)