FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators (2202.11214v1)

Abstract: FourCastNet, short for Fourier Forecasting Neural Network, is a global data-driven weather forecasting model that provides accurate short to medium-range global predictions at $0.25^{\circ}$ resolution. FourCastNet accurately forecasts high-resolution, fast-timescale variables such as the surface wind speed, precipitation, and atmospheric water vapor. It has important implications for planning wind energy resources, predicting extreme weather events such as tropical cyclones, extra-tropical cyclones, and atmospheric rivers. FourCastNet matches the forecasting accuracy of the ECMWF Integrated Forecasting System (IFS), a state-of-the-art Numerical Weather Prediction (NWP) model, at short lead times for large-scale variables, while outperforming IFS for variables with complex fine-scale structure, including precipitation. FourCastNet generates a week-long forecast in less than 2 seconds, orders of magnitude faster than IFS. The speed of FourCastNet enables the creation of rapid and inexpensive large-ensemble forecasts with thousands of ensemble-members for improving probabilistic forecasting. We discuss how data-driven deep learning models such as FourCastNet are a valuable addition to the meteorology toolkit to aid and augment NWP models.

Overview of FourCastNet: A Global Data-Driven High-Resolution Weather Model using Adaptive Fourier Neural Operators

The paper "FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators" presents a novel approach to numerical weather prediction (NWP) that leverages deep learning techniques to provide highly accurate and computationally efficient weather forecasts. This work introduces FourCastNet, a Fourier-based neural network model designed to produce global weather forecasts with a spatial resolution of $0.25^{\circ}$ latitude-longitude.

FourCastNet's primary strength lies in its ability to achieve forecasting performance comparable to the ECMWF Integrated Forecasting System (IFS) while being several orders of magnitude faster, consequently enabling the generation of extensive ensemble forecasts at a fraction of the computational cost typically associated with traditional NWP models.

Technical Contributions and Methodology

FourCastNet employs a sophisticated architecture combining Adaptive Fourier Neural Operators (AFNO) with a vision transformer (ViT) backbone. This hybrid model benefits from resolution-invariant learning properties and the ability to model long-range dependencies effectively.

Key Features:

1. High-Resolution Forecasting: FourCastNet operates at $0.25^{\circ}$ resolution, capturing fine-scale atmospheric dynamics that are crucial for accurate weather prediction, especially for variables like surface wind speed ($U_{10}$, $V_{10}$) and precipitation.
2. Numerical Aspects:
   • Fourier Neural Operators (FNO): Utilized for efficient token mixing in the spectral domain, improving the spatial mixing complexity to $\mathcal{O}(N \log N)$.
   • Vision Transformers (ViT): Enable handling of large spatial dimensions, facilitating the learning of detailed spatiotemporal dependencies.
3. Training and Optimization: The training process is divided into pre-training and fine-tuning stages, leveraging a large dataset from ERA5. The model is fine-tuned to predict two consecutive time steps, thereby enhancing stability during auto-regressive inference.
4. Precipitation Modeling: The total precipitation (TP) is treated as a diagnostic variable. A separate AFNO model was trained specifically for TP to avoid the complexities involved in direct forecasting due to its highly intermittent nature.

Predictive Performance and Case Studies

Quantitative Metrics:

• Anomaly Correlation Coefficient (ACC) and Root Mean Squared Error (RMSE) were used to evaluate model performance.
• For key atmospheric variables (e.g., $U_{10}$, $T_{2m}$, $Z_{500}$), FourCastNet displayed significant short-term accuracy, outperforming IFS in some metrics, especially for short lead times.

Case Studies:

1. Hurricanes: FourCastNet's ability to predict the formation, intensification, and trajectory of hurricanes was demonstrated using Hurricane Michael (2018) as a case paper. The model effectively tracked the hurricane's eye and intensity changes over time.
2. Atmospheric Rivers: The model successfully captured the formation and evolution of atmospheric rivers, which are critical for water resource planning and flood prediction.

Computational Efficiency

A notable advantage of FourCastNet is its computational efficiency:

• The model can generate a week-long forecast in less than 2 seconds using modern GPU hardware.
• Energy consumption is dramatically reduced, with FourCastNet using about 12,000 times less energy than IFS for forecast generation.

Implications and Future Directions

FourCastNet's unprecedented speed and accuracy offer significant benefits, particularly in generating large ensembles for probabilistic forecasting. This capability is pivotal in estimating well-calibrated uncertainties, which improve the reliability of extreme weather event predictions and governmental disaster response initiatives.

Speculative Future Developments:

1. Higher Resolution Models: Training on even finer spatial resolutions could further enhance predictive accuracy for small-scale atmospheric phenomena.
2. Physics-Informed Models: Incorporating physical laws into the model training process could ensure physically consistent forecasts.
3. Sub-seasonal to Seasonal (S2S) Predictions: Combining FourCastNet with existing S2S models might extend the prediction horizon with improved skill.

Overall, the development of FourCastNet represents a significant advancement in the application of deep learning to atmospheric sciences, showcasing its potential to complement and possibly replace traditional NWP models in the coming years. As computational power and data availability continue to grow, leveraging such AI-driven techniques will become increasingly viable for operational weather forecasting.