Accurate weather forecasting is crucial for various scientific and societal applications, particularly for anticipating and mitigating the impact of extreme weather events. Numerical weather prediction (NWP) methods, which solve partial differential equations representing atmospheric states, are currently the most accurate but computationally expensive. While AI-based methods offer significant speed advantages, their accuracy has lagged behind NWP. This research addresses this gap by presenting a novel AI-based approach that leverages three-dimensional deep learning to achieve superior accuracy in medium-range global weather forecasting, surpassing the performance of established NWP systems.

Traditional NWP methods rely on discretized atmospheric grid representations and numerical solutions of partial differential equations governing atmospheric transitions. However, these methods are computationally intensive and incorporate parameterizations that introduce approximation errors. Recent advancements in deep learning have sparked interest in AI-based weather forecasting, with methods like FourCastNet demonstrating faster processing times. Despite this speed improvement, the accuracy of AI-based methods has fallen short of NWP systems like ECMWF's operational integrated forecasting system (IFS). Existing research highlights the potential of AI but acknowledges the need for substantial breakthroughs to surpass the accuracy of NWP.

Pangu-Weather employs a three-dimensional (3D) deep learning architecture, the 3D Earth-specific transformer (3DEST). Unlike previous 2D models, 3DEST integrates height as a dimension, allowing the network to capture relationships between atmospheric states at different pressure levels. The architecture utilizes an encoder-decoder structure inspired by the Swin transformer, incorporating an Earth-specific positional bias to encode location information. This 3DEST model is trained on 39 years of global ERA5 reanalysis data (1979-2017), with a training set encompassing 341,880 one-hour time points. To mitigate overfitting, data is randomly permuted at the start of each epoch. Four deep networks are trained with lead times of 1, 3, 6, and 24 hours. Each network undergoes 100 epochs of training using a cluster of 192 NVIDIA Tesla V100 GPUs, taking approximately 16 days per network. A hierarchical temporal aggregation strategy is implemented for medium-range forecasting. This greedy algorithm iteratively utilizes the trained networks with the longest feasible lead times to minimize cumulative errors. For instance, a 56-hour forecast might involve two 24-hour forecasts, one 6-hour forecast, and two 1-hour forecasts. The spatial resolution of Pangu-Weather is 0.25° × 0.25°, comparable to FourCastNet and the ECMWF ENS control forecast. The forecast spacing is 1 hour, six times finer than FourCastNet. For ensemble forecasting, 99 random perturbations are added to the initial state to create a 100-member ensemble, the results of which are averaged.

Pangu-Weather demonstrates superior performance compared to both the ECMWF IFS and FourCastNet across all tested weather variables using ERA5 reanalysis data for validation (2019) and testing (2018). Specifically, Pangu-Weather achieves a lower root mean square error (RMSE) and higher anomaly correlation coefficient (ACC) than both competitors. For a 5-day Z500 forecast, Pangu-Weather achieves an RMSE of 296.7, significantly lower than the 333.7 of the ECMWF IFS and the 462.5 of FourCastNet. Inference time for Pangu-Weather is 1.4 seconds on a single GPU, more than 10,000 times faster than the ECMWF IFS and comparable to FourCastNet. The model's accuracy extends to extreme weather events, as evidenced by its superior performance in tropical cyclone tracking. In 2018, Pangu-Weather exhibited lower mean direct position errors than ECMWF-HRES for 88 named tropical cyclones, particularly for longer lead times. The ensemble forecasting approach, while slightly less accurate than the single-member forecast for short-range predictions, provides significant improvement for medium-range forecasts (5-7 days), particularly for non-smooth variables. The spread-skill ratio indicates some underdispersion in the ensemble method.

The results demonstrate that Pangu-Weather's novel 3D architecture and hierarchical temporal aggregation strategy lead to substantial improvements in medium-range global weather forecasting accuracy and speed compared to state-of-the-art NWP and AI-based methods. The superior performance in both deterministic and ensemble forecasting, along with accurate extreme weather event tracking, positions Pangu-Weather as a significant advance in the field. The high speed of the model allows for the practical application of large-ensemble forecasting, opening new avenues for improving weather prediction. This research contributes significantly to the ongoing effort of establishing AI as a valuable complement or even a potential replacement for traditional NWP methods.

Pangu-Weather offers a significant advancement in global weather forecasting, surpassing the accuracy of established NWP systems while maintaining remarkable computational efficiency. Future research should focus on adapting Pangu-Weather for use with observational data, incorporating additional variables (like precipitation), refining the ensemble forecasting methodology, and addressing temporal inconsistencies introduced by using multiple models with varying lead times. Further advancements could involve incorporating more vertical levels, using four-dimensional deep networks, and improving model training processes.

This study's primary limitation is the reliance on reanalysis data for both training and testing. Real-world forecasting systems use observational data, which may differ in ways that could impact Pangu-Weather's performance. Additionally, the model's evaluation did not include critical variables such as precipitation, potentially affecting its ability to predict small-scale extreme events. The inherent smoothing characteristic of AI-based methods could lead to an underestimation of extreme weather magnitudes. Finally, the use of models with different lead times can introduce temporal inconsistencies, requiring further investigation.