Pangu-Weather: AI Global Forecasting

• Pangu-Weather is a deep-learning global forecasting system that uses ERA5 reanalysis and a 3D Earth-Specific Transformer for high-resolution medium-range predictions.
• It employs lead-time-specific networks with hierarchical temporal aggregation to reduce cumulative error across 1 to 24-hour forecast intervals.
• The model outperforms traditional NWP systems on key deterministic metrics and serves as a benchmark for probabilistic forecasting, severe-weather guidance, and control studies.

Pangu-Weather is a deep-learning global weather forecasting system introduced as a $0.25^\circ\times0.25^\circ$ high-resolution model for deterministic medium-range prediction, using upper-air fields on 13 pressure levels together with surface variables and multiple lead-time-specific networks. Its core design combines a 3D Earth-Specific Transformer with hierarchical temporal aggregation, and it has since become both a benchmark for AI weather prediction and a substrate for work on probabilistic forecasting, severe-weather guidance, dynamical testing, and control-oriented sensitivity analysis (Bi et al., 2022).

1. Origins and defining characteristics

Pangu-Weather was introduced by Bi et al. as a fast, data-driven global weather forecasting system trained on ERA5 and evaluated against ECMWF’s Integrated Forecast System. In the original benchmark setting, it was reported to outperform state-of-the-art NWP in latitude-weighted RMSE and ACC for all evaluated variables and all lead times from one hour to one week, while operating at $0.25^\circ\times0.25^\circ$ resolution and supporting downstream scenarios such as tropical cyclone tracking and large-member ensembles (Bi et al., 2022). Later summaries characterize it as a fully data-driven global forecast model trained on ERA5 reanalysis and producing forecasts on 13 pressure levels plus surface variables, with four separate lead-time models at 1, 3, 6, and 24 hours (Bonavita, 2023).

The system is usually described as a global deterministic forecaster rather than a coupled Earth-system simulator. In the original formulation, the upper-air state comprises geopotential, temperature, specific humidity, and horizontal wind components on 13 pressure levels from 1000 to 50 hPa, while the surface state comprises 2 m temperature, 10 m wind components, and mean sea-level pressure (Bi et al., 2022). A later operational-control study describes the same model family as using 13 atmospheric layers plus surface variables, with approximately 64 million parameters per 3D network; this is consistent with the original description of a few networks with about 256 million parameters in total across lead times (Peng et al., 21 May 2026).

Pangu-Weather occupies a distinctive position in the AI weather literature because it is both a forecasting system and, increasingly, a reference architecture. Operational-compatibility studies emphasize that it is open source for non-commercial use at inference time and can be driven by multiple analysis systems once those fields are mapped into the required variable set, levels, and $0.25^\circ$ grid (Cheng et al., 2023).

2. Architecture and data representation

The model’s central architectural idea is to represent the atmosphere as a 3D volume in latitude, longitude, and pressure level, rather than as a stack of 2D maps. The original paper formulates forecasting as learning a mapping

$$f(\cdot;\boldsymbol{\theta}): \mathbf{A}_{t}^{\ast} \mapsto \hat{\mathbf{A}}_{t+\Delta t}^{\ast},$$

where $\mathbf{A}_{t}^{\ast}$ denotes the full multivariate atmospheric state at time $t$ (Bi et al., 2022). Upper-air fields are patch-embedded into a cubic tensor, surface fields are embedded separately and concatenated as an additional height slice, and the resulting volume is processed by an encoder–decoder transformer with skip connections (Bi et al., 2022).

The defining block is the 3D Earth-Specific Transformer, which adapts windowed self-attention to the geometry of the Earth and to the vertical structure of the atmosphere. Instead of treating the global grid as an ordinary image, Pangu-Weather uses Earth-Specific Positional Bias, so that attention weights depend on vertical position, latitude band, and relative longitude offset. The attention rule is written as

$$\mathrm{Attention}(\mathbf{Q}, \mathbf{K}, \mathbf{V}) = \mathrm{SoftMax}\left( \frac{\mathbf{Q}\mathbf{K}^\top}{\sqrt{D}} + \mathbf{B} \right)\mathbf{V},$$

with $\mathbf{B}$ carrying the Earth-specific positional bias (Bi et al., 2022). This design explicitly addresses the regular lat–lon grid’s nonuniform physical spacing with latitude and the strong vertical anisotropy of atmospheric dynamics.

Subsequent analyses describe Pangu-Weather as a vision-style or Swin-like transformer operating on 3D atmospheric cubes, with local 3D processing as an inductive prior. In that framing, nearby pressure levels interact locally within attention windows, while shifted windows propagate information across neighborhoods (Couairon et al., 2024). This architectural choice later became the target of direct critique and modification in follow-on models such as ArchesWeather, which argued that Pangu’s 3D local processing is computationally sub-optimal and replaced local cross-level interactions with explicit column-wise attention (Couairon et al., 2024).

The model’s data basis is ERA5 reanalysis, interpolated globally to $0.25^\circ$ resolution for input and output fields. Independent studies that reuse the published implementation characterize the core input/output set as five 3D atmospheric fields on 13 pressure levels and four surface fields, exactly matching the original system definition (Bülte et al., 2024).

3. Forecast generation and hierarchical temporal aggregation

Pangu-Weather does not rely on a single autoregressive time step. Instead, it trains four separate models for direct forecasts at 1, 3, 6, and 24 hours, and longer forecasts are constructed by combining these lead-time-specific networks through hierarchical temporal aggregation (Bi et al., 2022). The purpose is to reduce cumulative forecast error relative to repeatedly applying a short-step model. A 7-day forecast, for example, can be assembled from seven 24-hour steps rather than 168 one-hour steps (Bi et al., 2022).

This design is important because repeated application of an imperfect model leads to rapid error growth. The hierarchical scheme was introduced precisely to alleviate that cumulative-error problem, and later assessments identify it as one of the two central innovations of the original system, alongside the 3D Earth-Specific Transformer (Bi et al., 2022). Another analysis notes that Pangu thereby avoids strictly fixed-step autoregression and provides 1-hour temporal granularity through combinations of the 1, 3, 6, and 24 hour models, though at the cost of possible temporal discontinuities when switching among models (Bonavita, 2023).

In practical downstream usage, however, the 6-hour model often becomes the operational step. A typhoon-sensitivity testbed built around Pangu-Weather uses a repeated 6-hour control cycle: perturb an ECMWF-style initial field, run Pangu forward 6 hours, diagnose storm-center metrics, update a PID controller, and repeat (Peng et al., 21 May 2026). The same study treats Pangu as a black-box nonlinear plant in control-theoretic terms and integrates it through an ONNX-based MATLAB–Python bridge because MATLAB’s ONNX importer could not directly load the supplied model versions (Peng et al., 21 May 2026). This suggests that, beyond its original forecast role, Pangu-Weather is increasingly used as a fast surrogate dynamics operator.

4. Forecast skill, computational profile, and operational evaluation

In the original ERA5-based benchmark, Pangu-Weather reported better latitude-weighted RMSE and ACC than ECMWF IFS for all evaluated variables and all lead times from 1 hour to 7 days (Bi et al., 2022). Representative numbers from that study include 500 hPa geopotential RMSE of 134.5 at 3 days and 296.7 at 5 days, versus 152.8 and 333.7 for IFS, and 850 hPa temperature RMSE of 1.14 at 3 days and 1.79 at 5 days, versus 1.37 and 2.06 for IFS (Bi et al., 2022). Surface verification showed the same pattern: for 2 m temperature and 10 m winds, Pangu outperformed both ECMWF IFS and FourCastNet in the reported benchmark (Bi et al., 2022).

Its computational profile is one of the model’s most consequential properties. The original paper states that Pangu achieves more than $10{,}000\times$ speedup relative to operational IFS (Bi et al., 2022), and a later control-oriented implementation reports about 1.4 seconds per global forecast on local hardware (Peng et al., 21 May 2026). This speed is what enables many of the later ensemble, control, and post-processing applications.

Independent operational-style evaluations are more nuanced than the original global benchmark. Over 183 Norwegian SYNOP stations, Pangu-Weather was found to be slightly better than ECMWF global models for 2 m temperature and slightly worse for 10 m wind speed, while the 2.5 km MEPS regional model clearly provided the best forecasts for both parameters (Bremnes et al., 2023). After Bernstein Quantile Network post-processing, forecast quality improved considerably for all models, with larger relative gains for the coarse global models because of stronger systematic deficiencies (Bremnes et al., 2023).

Pangu-Weather’s compatibility with non-ERA5 initial conditions has also been examined directly. Case studies using GFS, CMA-GRAPES, and YHGSM analyses showed that Pangu could ingest different operational analyses once interpolated to $0.25^\circ\times0.25^\circ$0 and mapped to the required variables and levels, and that improving local initial-condition quality—for example by replacing East Asia pressure-level fields in GFS with ECMWF fields—substantially improved regional forecast skill over East Asia (Cheng et al., 2023). This suggests that the model is not restricted to training-consistent ERA5 initialization, even if initial-condition quality remains a dominant control on forecast quality.

5. Physical behavior, realism, and scientific assessment

Pangu-Weather has been subjected to two sharply contrasting lines of scientific assessment. One line argues that it encodes substantial atmospheric dynamics. In a set of canonical dynamical experiments, the model produced a Matsuno–Gill-type response to steady tropical heating, realistic baroclinic wave packets and extratropical cyclones from a localized jet perturbation, geostrophic adjustment from a height-only perturbation over roughly 6 hours, and hurricane genesis from localized subtropical low-pressure seeds, with development suppressed when initial humidity was set to zero (Hakim et al., 2023). These tests do not resemble standard training samples and were interpreted as evidence that the model has internalized physically meaningful wave propagation, balance adjustment, moist sensitivity, and storm development mechanisms (Hakim et al., 2023).

A second line of assessment argues that forecast skill and physical fidelity are not equivalent. A detailed comparison with ERA5 hindcasts and ECMWF IFS found that Pangu’s spectra diverge from ERA5 already at total wavenumbers 60–80, corresponding to an effective horizontal resolution nearer 500–700 km, with increasing spectral damping as lead time grows (Bonavita, 2023). The same study reported systematically weakened ageostrophic and divergent flow, diagnosed vertical velocities roughly 40–50% smaller than in ERA5 hindcasts and IFS, and markedly shallower tropical cyclones, exemplified by Typhoon Doksuri forecasts that were much weaker than both IFS and observations (Bonavita, 2023). Its central conclusion is that Pangu’s gains on deterministic metrics are at least partly entangled with smoothing and reduced physical consistency (Bonavita, 2023).

Extreme-event evaluations reinforce those concerns. For record-breaking heat, cold, and wind events, ECMWF HRES was found to outperform both research and operational versions of Pangu-Weather across nearly all lead times, with Pangu systematically underestimating the frequency and intensity of record-breaking events (Zhang et al., 21 Aug 2025). A separate study of 2 m temperature during 60 U.S. heat waves found persistent cold biases before and during heat waves in most seasons, with the exception of winter, when Pangu exhibited a mean warm bias before onset (Ennis et al., 29 Apr 2025). These results do not negate the model’s forecast skill, but they delimit the regimes in which its extrapolation and tail behavior remain problematic.

6. Applications, extensions, and derivative systems

Pangu-Weather has become a reusable backbone for multiple research programs. In probabilistic forecasting, uncertainty-quantification methods built around deterministic Pangu produced competitive or better CRPS than the raw ECMWF ensemble over Europe up to about 5 days, especially for 2 m temperature, using random-field perturbations and post-hoc methods such as EasyUQ and distributional regression networks (Bülte et al., 2024). In severe-convective guidance over CONUS, decoder-only transformer post-processing applied to Pangu environmental forecasts outperformed GFS-based counterparts on Days 3–6, and Pangu-HRES and Pangu-GFS initializations showed strong medium-range skill even without explicit convective parameters (Hua et al., 16 May 2025).

The model has also been extended beyond its original forecast horizon. A teleconnection-informed transformer kept the pretrained 24-hour Pangu backbone mostly frozen, updated only about 700k parameters—roughly 1.1% of the full model—via LoRA and an external teleconnection module, and improved predictability at a 14-day lead while preserving more spatial granularity than full fine-tuning (Zhao et al., 2024). In another direction, PW-FouCast used Pangu forecasts of geopotential, humidity, temperature, and winds as spectral priors for precipitation nowcasting, arguing that Pangu phases contain useful large-scale structural information for radar forecasting and reporting state-of-the-art results on SEVIR and MeteoNet (Qin et al., 23 Mar 2026).

Control-oriented and regional adaptations further broaden the model’s role. The typhoon-sensitivity testbed built around Pangu and PID control demonstrated controlled perturbation–response experiments on Super Typhoon Connie, including feasible track-angle responses and intensity modulation via velocity and temperature perturbations, while explicitly treating Pangu as a fast nonlinear atmospheric surrogate (Peng et al., 21 May 2026). A regional Indian forecasting system retained the Swin-transformer backbone of Pangu, removed Earth-specific modules, trained on IMDAA rather than ERA5, and reported robust 6-hour performance with hierarchical forecasting judged best among static, autoregressive, and hierarchical schemes (Choudhury et al., 17 Mar 2025).

Finally, Pangu-Weather has served as a direct architectural and geometric reference for successor models. ArchesWeather criticized its 3D local processing and proposed factorized 2D horizontal attention plus cross-level attention (Couairon et al., 2024), while PEAR recast a Pangu-style transformer on the equal-area HEALPix grid and reported better forecast skill than corresponding Driscoll–Healy implementations without computational overhead (Linander et al., 23 May 2025). This suggests that Pangu’s lasting influence may lie as much in its design vocabulary as in any single released model.

7. Limitations, controversies, and future directions

The literature converges on a mixed assessment. Pangu-Weather is clearly fast, skillful on standard deterministic metrics, and flexible enough to support ensemble generation, post-processing, control-theoretic experiments, and downstream hazard models (Bi et al., 2022). It is also sufficiently robust to accept multiple operational analyses as input and sufficiently modular to underpin regional or task-specific adaptations (Cheng et al., 2023).

At the same time, multiple studies emphasize that it should not be treated as a fully faithful digital twin of atmospheric dynamics. Documented concerns include spectral damping, reduced divergent flow and vertical motion, underestimation of strong tropical cyclones and record-breaking extremes, and biases in heat-wave temperature forecasts (Bonavita, 2023). A plausible implication is that deterministic score advantages can coexist with deficiencies in physical balance, tail risk, and downstream utility for applications that require realistic extremes or conservation-compatible fields.

Future work in the surrounding literature therefore points in several directions: probabilistic or generative formulations rather than purely deterministic outputs (Bülte et al., 2024); physics-informed or hybrid loss functions that explicitly constrain spectra, balances, and vertical motions (Bonavita, 2023); more advanced control frameworks such as model predictive control and multi-input multi-output perturbation analysis built on Pangu-like surrogates (Peng et al., 21 May 2026); and architectural revisions involving equal-area spherical grids, factorized vertical interactions, or parameter-efficient adaptation for new lead times and tasks (Linander et al., 23 May 2025). In that sense, Pangu-Weather marks both a milestone and a transitional design: it established that AI weather forecasting can be globally skill-competitive at operational resolutions, while also making clear that forecast accuracy, physical realism, and robustness to rare extremes remain distinct scientific objectives (Bi et al., 2022).