2022

FourCastNet: A Global Data-Driven High-Resolution Weather Model Using Adaptive Fourier Neural Operators

J. Pathak, et al.

Observation and Context

Traditional weather forecasting relies on physical simulations called Numerical Weather Prediction (NWP) models, such as the Integrated Forecasting System (IFS). While highly accurate, these physics-based models require massive supercomputers and hours of computing time to generate global forecasts. Although data-driven Deep Learning (DL) models are much faster, prior versions were trained on low-resolution data that could not capture small-scale physical features. This grid coarseness made it impossible to resolve critical, high-impact events like local winds, severe rain, and tropical cyclones.

Hypothesis

The researchers hypothesized that by utilizing an Adaptive Fourier Neural Operator (AFNO) coupled with a vision transformer (ViT) architecture, they could build a global, data-driven weather model called FourCastNet. They predicted that this neural network could produce high-resolution global forecasts (0.25° grid spacing) that match the accuracy of physics-based NWP systems at short lead times, while generating forecasts thousands of times faster.

Experiment and Methodology

To test this, the research team designed and trained FourCastNet using the historical ERA5 atmospheric reanalysis dataset.

• Data Split: They trained the model on data spanning 1979 to 2015, validated it using 2016 to 2017 data, and evaluated it with untouched out-of-sample data from 2018 onward.
• Variables: The model tracks 20 key atmospheric variables across 5 vertical pressure levels.
• Model Mechanics: Utilizing the AFNO framework, the model translates spatial information into the frequency domain using Discrete Fourier Transforms. This reduces computational complexity to O(N log N), enabling highly efficient scaling on high-resolution data.
• Precipitation Diagnostic: Because precipitation is highly localized and complex, the team trained a separate diagnostic model to forecast 6-hourly accumulated rainfall using the backbone model’s outputs.

Results and Data

The experimental results demonstrated massive speed and energy improvements over traditional systems:

• Incredible Speed: FourCastNet produced a week-long global forecast in less than 2 seconds, making it approximately 45,000 times faster than physics-based IFS models on a node-hour basis.
• Competitive Accuracy: For short-term forecasts under 48 hours, FourCastNet outperformed the IFS model in both Anomaly Correlation Coefficient (ACC) and Root Mean Squared Error (RMSE) for surface winds and precipitation.
• Extreme Weather Tracking: The model successfully tracked the precise formation, rapid intensification, and paths of massive systems, such as Category 5 Hurricane Michael and atmospheric rivers bringing rain to California.
• Energy Efficiency: Generating a 100-member, 24-hour ensemble forecast took 7 seconds and consumed only 8 kJ of energy on a single GPU node. In contrast, the IFS model consumed 271,000 kJ of energy for the same forecast — meaning FourCastNet used over 12,000 times less energy.

Conclusion and Climate Impact

The study proved that high-resolution, purely data-driven neural networks can match the precision of traditional, physics-based weather systems.

Because FourCastNet runs incredibly fast on consumer-grade computer hardware, it democratizes weather forecasting. Meteorologists can now run massive ensemble forecasts with thousands of members in seconds rather than hours. This allows disaster response teams to obtain rapid, probabilistic warnings for extreme events like floods and cyclones, saving lives and dramatically improving operational planning for wind energy farms.