Source-linked AI summary
FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators
Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, Pedram Hassanzadeh, Karthik Kashinath, Animashree Anandkumar
TL;DR
High-resolution global weather forecasting must preserve fine-scale information while remaining computationally affordable for large ensembles. FourCastNet uses an AFNO-based data-driven model trained on ERA5 and autoregressive forecasts, achieving competitive short-range accuracy with major speed and ensemble-generation advantages. Its scope remains narrower than IFS in physical constraints, variables, and vertical levels.
Problem
Existing data-driven weather models commonly use coarse grids that lose fine-scale information, limiting forecasts of phenomena such as tropical cyclones and precipitation.
Method
FourCastNet uses an Adaptive Fourier Neural Operator trained on ERA5 atmospheric states, with autoregressive inference and a separate diagnostic model for precipitation.
Results
FourCastNet matches IFS on RMSE and ACC through three days, remains close behind through one week, and is about 45,000 times faster than traditional NWP models.
Takeaways & Limitations
Rapid, inexpensive inference enables very large ensembles for probabilistic forecasting of extreme weather and supports wind-energy resource planning.
Takeaways & Limitations
FourCastNet does not impose physics constraints and outputs far fewer variables and vertical levels than IFS; its speed assessments also omit common deep-learning inference optimizations.
Abstract
from arXiv · showhide
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.
1 Introduction
FourCastNet addresses the need for high-resolution global data-driven weather forecasts by producing 0.25° predictions that resolve fine-scale atmospheric features. It combines competitive accuracy with much faster inference, enabling large ensembles and applications to extreme weather and wind-energy planning.
- Motivation and approach: Prior data-driven weather models commonly used coarse 2°–5.625° grids that lost fine-scale physical information and could not resolve features smaller than approximately 500 km.High-resolution forecasts are needed to represent high-impact events such as tropical cyclones.
- Motivation and approach: 0.25° global forecasts provide roughly 30 km × 30 km resolution near the equator and enable direct comparison with high-resolution IFS forecasts.The model uses a 720 × 1440 global grid.
- Contributions: 45,000 times faster than traditional NWP models, FourCastNet enables inexpensive large-ensemble forecasts for extreme-event uncertainty and rapid assessment.A separate discussion reports that 1,000-member ensembles can be generated rapidly, while forecasts can support disaster response and wind-energy planning.
- Contributions: FourCastNet predicts challenging variables such as surface winds and precipitation with accuracy at forecast lead times of up to one week.The paper emphasizes implications for disaster mitigation and wind-energy resource planning.
- Contributions: FourCastNet has eight times greater resolution than state-of-the-art deep-learning global weather models and resolves extreme events such as tropical cyclones and atmospheric rivers.Prior models inadequately represented these events because of coarser grids.
- Contributions: FourCastNet predictions are comparable to IFS on RMSE and ACC through three days, while later forecasts remain close behind IFS through one week.FourCastNet models 20 variables at five vertical levels, whereas IFS contains more than 150 variables at over 50 vertical levels and is physics-guided.
2 Training Methods
FourCastNet combines ERA5 atmospheric data, an Adaptive Fourier Neural Operator architecture, and autoregressive training to forecast high-resolution atmospheric states. Its Fourier-domain token mixing is designed to scale efficiently with resolution, while precipitation is handled through a separate diagnostic model.
- Data and variables: ERA5 supplies hourly atmospheric reanalysis fields at 0.25° resolution from the surface to roughly 100 km altitude.ERA5 combines observations with numerical-model output through data assimilation.
- Data and variables: The model forecasts 20 atmospheric variables, including winds, temperature, pressure, humidity, geopotential height, and total-column water vapor.Near-surface wind and six-hourly precipitation are the primary challenging variables of interest.
- Model architecture: AFNO replaces quadratic self-attention token mixing with efficient Fourier-domain global convolution whose complexity scales as O(N log N).This design supports high-resolution inputs and models dependencies across spatial and channel dimensions.
- Model architecture: The architecture patches the 720 × 1440 input grid, embeds patches as tokens with positional encoding, applies AFNO layers, and decodes the next-frame patches.Each AFNO layer performs Fourier-domain spatial mixing followed by channel mixing.
- Training and inference: Training first learns a one-step mapping, then fine-tunes the model on two autoregressive steps using losses against both successive ground-truth states.Inference subsequently runs the trained model autoregressively from held-out initial conditions.
- Training and inference: Precipitation is excluded from the backbone’s 20 training variables and predicted by a separate AFNO diagnostic model using backbone outputs.This decouples precipitation modeling from forecasting the general atmospheric state.
3 Results
FourCastNet produces high-resolution forecasts of fine-scale weather phenomena and achieves competitive skill against IFS across several variables. Its speed enables large ensembles, while precipitation extremes and some hurricane-intensity changes remain challenging.
- High-resolution forecasts: 0.25° resolution allows FourCastNet to resolve and forecast global surface-wind structures, including cyclone formation, intensification, and tracks.The model qualitatively tracks Super Typhoon Mangkhut and several Atlantic hurricanes over a four-day forecast.
- High-resolution forecasts: FourCastNet diagnoses high-resolution precipitation from predicted prognostic variables and captures localized intense features associated with atmospheric rivers.The authors describe this as the first competitive global precipitation diagnosis by a deep-learning model at this scale.
- Extreme-weather applications: FourCastNet forecasts Hurricane Michael’s formation, intensification, and landfall using 100 perturbed initial conditions and related wind and pressure variables.The forecasts qualitatively match the ground truth over 72 hours, but do not fully capture the sharp pressure drop between 36 and 48 hours.
- Extreme-weather applications: FourCastNet predicts atmospheric-river evolution through total-column water vapor, with ACC above 0.6 beyond eight days.The example tracks an atmospheric river building and making landfall in Northern California; future work would add integrated vapor transport and liquid-water variables.
- Quantitative forecast skill: At lead times of approximately 48 hours or less, FourCastNet outperforms IFS in ACC and/or RMSE for key precipitation, wind, and temperature variables while remaining competitive over several days.This accuracy is achieved using only part of IFS’s full variable set and at a fraction of its computational cost.
- Ensemble forecasting: A 100-member ensemble mean improves ACC and RMSE over the unperturbed control at longer lead times but slightly degrades short-lead skill below 48 hours.The reported improvement is attributed to averaging over perturbed members, which can smooth relevant fine-scale features at short lead times.
4 Computational Cost of FourCastNet
FourCastNet offers substantially faster ensemble forecast generation than IFS, although the comparison is based on differing hardware and model capabilities. Its low infrastructure requirements make it attractive for limited-purpose forecasting.
- The speed comparison is not definitive because IFS runs on a CPU cluster whereas FourCastNet runs on one or several GPUs.The authors explicitly frame the estimate as a nuanced comparison rather than an apples-to-apples numerical advantage.
- FourCastNet computes a 100-member 24-hour forecast in 7 seconds on one Perlmutter node containing four A100 GPUs.Batched inference uses a batch size of 25 across the four GPUs.
- 45,000–145,000 times faster: FourCastNet generates 100-member ensemble forecasts using vastly fewer nodes than IFS, depending on resolution.The reported range compares hypothetical 18 km and 30 km FourCastNet inference with IFS on a node-to-node basis.
- IFS provides physical consistency and many more variables, while FourCastNet currently imposes no physics constraints and is generally less accurate.FourCastNet nevertheless approaches IFS accuracy for several variables and exceeds it for precipitation in some cases.
- FourCastNet can generate a 10-day global forecast at full ERA5 resolution on a single device, completing it in seconds.The authors present this as especially attractive for limited-purpose applications such as short-term surface-wind forecasting.
5 Comparison Against State-of-the-art DL Weather Prediction
FourCastNet is evaluated against the lower-resolution DLWP model by coarsening its outputs to comparable resolution. Despite that adjustment, it shows substantial improvement while retaining finer native resolution.
- 8 times greater resolution: FourCastNet predicts more variables than DLWP and resolves small-scale features such as hurricanes, extreme precipitation, and atmospheric rivers.DLWP operates at 2° resolution, whereas FourCastNet operates at 0.25° resolution.
- FourCastNet’s higher resolution enables small-scale weather features that would not be possible to represent adequately at DLWP’s 2° grid.The paper contrasts this with even coarser 5° resolution, described as almost entirely futile for such features.
- FourCastNet predictions show significant improvement over DLWP for ACC and RMSE on downsampled Z500 and T2m forecasts.The comparison uses downsampled FourCastNet and IFS predictions alongside the DLWP baseline.
- The authors caution that coarsening makes FourCastNet less effective to accommodate a lower-resolution benchmark, so the comparison is not fully favorable to FourCastNet.They provide the comparison for completeness rather than as an ideal evaluation of the model.
- The DLWP comparison coarsens FourCastNet predictions eight times in each direction using bilinear interpolation.ACC and RMSE are recomputed for Z500 and T2m, the variables reported by DLWP.
6 Implications, Discussion, and Future Work
FourCastNet combines high-resolution forecasting accuracy with exceptional speed, enabling large ensembles and practical applications while leaving clear avenues for future development and unresolved climate-change generalization questions.
- Implications: FourCastNet’s accuracy is comparable to IFS and better than state-of-the-art deep-learning weather models on short timescales, despite using fewer variables and vertical levels.The authors anticipate further development could extend capabilities across all timescales and atmospheric levels.
- Implications: Four to five orders of magnitude faster predictions enable thousands-member ensembles, improving uncertainty estimation for extreme events and supporting rapid hypothesis testing.The paper contrasts potential 1,000-member ensembles with current NWP ensembles of approximately 50 members.
- Implications: Accurate short-range precipitation forecasts could support rapid disaster responses and accept prognostic inputs from different models or observational sources.This flexibility is described for a diagnostic precipitation model.
- Implications: Rapid, reliable, high-resolution wind forecasts can support extreme-wind disaster mitigation, wind-power planning, and wind-farm layout optimization.The applications cover fluctuating wind output and diverse wind and weather conditions.
- Future Work: FourCastNet’s skill improves with more modeled variables, suggesting larger models could extend prediction horizons and produce fast forecasts of complete 3D atmospheric fields.The paper notes that training such models would be computationally expensive, while inference could remain rapid.
- Limitations and Future Work: Without data assimilation, FourCastNet cannot yet generate up-to-the-minute forecasts, although future observational initialization could enable real-time prediction.The paper proposes adding an ensemble-based data-assimilation component.
- Future Work: A sub-kilometer FourCastNet could offer more than six orders of magnitude speedup over sub-kilometer NWP, despite expensive training.The projection relies on rapidly performed inference for large deep-learning models.
- Future Work: A coupled two-timescale model could combine FourCastNet’s short-timescale skill with DLWP’s subseasonal-to-seasonal skill.The envisioned coupling includes two-way interactions across short-, medium-, and long-range forecasting.
Appendix A: Model Hyperparameters
The appendix identifies the tables containing FourCastNet’s model-training hyperparameters and the evaluation settings used for ACC and RMSE analyses.
- Model Hyperparameters: Table 3 lists the hyperparameters used in the FourCastNet model and training.The appendix refers readers to Guibas et al. for AFNO backbone parameter definitions.
- Evaluation Settings: Table 4 lists the number of initial conditions and assumed temporal decorrelation times used to compute ACC and RMSE plots.The listed variables include Z500, T850, T2m, U10, V10, and TP.
Appendix B: MSLP visualization for Hurricane Michael
The appendix visualizes FourCastNet’s MSLP forecasts for Hurricane Michael and contrasts high- and coarse-resolution representations of hurricane-scale structure.
- Hurricane Michael: FourCastNet’s MSLP forecast depicts Hurricane Michael intensifying from a tropical depression to a hurricane while moving toward Florida’s coast.The forecast is compared with target ERA5 snapshots at matching time steps.
- Resolution Comparison: The appendix argues that high-resolution training is essential because coarse-resolution data cannot capture important small-scale phenomena such as hurricanes.Prior work at roughly 2° resolution is used as the coarse-resolution reference.
- Hurricane Michael: Figure 11 compares FourCastNet MSLP predictions with ground truth through 72-hour lead times from an October 7, 2018 initialization.A land-sea mask sets MSLP to zero over land for visualization.
- Resolution Comparison: Figure 12 provides ERA5 MSLP ground truth downsampled by a factor of eight for coarse-resolution visual reference.The accompanying discussion uses this representation to highlight lost small-scale structure.
Appendix C: ACC and RMSE definitions
The appendix defines latitude-weighted ACC and RMSE for forecast evaluation, explaining anomaly construction, latitude weighting, and aggregation across initial conditions.
- ACC: Latitude-weighted ACC measures agreement between predicted and true anomaly fields for variable v at forecast time-step l.The anomalies subtract each variable’s long-term mean at every grid location.
- ACC: The long-term mean is computed from historical training samples, and anomaly fields exclude the component represented by that mean.The predicted and true fields are evaluated over grid coordinates m and n.
- Reporting: Reported ACC values are averaged over forecasts from different initial conditions, with first- and third-quartile variability shown in the plots unless stated otherwise.This summarizes both central performance and variation across initial conditions.
- RMSE: Latitude-weighted RMSE quantifies prediction error for variable v at forecast time-step l using the same latitude weighting factor as ACC.The definition uses predicted and true variable values at grid coordinates m and n.
Appendix D: Additional ACC and RMSE results
The appendix extends FourCastNet skill evaluation with latitude-weighted ACC and RMSE across key and backbone-modeled variables, and compares wind accuracy over land and sea. Forecast skill generally remains strong for 5–10 days, while relative humidity errors accumulate fastest and surface-wind accuracy is similar over land and ocean.
- Latitude-weighted ACC and RMSE are evaluated for FourCastNet forecasts alongside matched IFS forecasts across key weather variables.The evaluation uses forecasts initialized from the 2018 out-of-sample testing dataset.
- ACC generally stays above 0.6 for 5–10 days across the backbone model’s variable set.The variables are grouped into wind velocities, geopotentials, temperatures, and other variables.
- Relative humidity variables accumulate forecast errors faster than the other modeled variables.
- Land- and sea-specific ACC is computed for surface wind velocity using a land–sea mask derived from ERA5 grid-box coverage.Grid boxes with more than 50% land are classified as land; the remainder are classified as water.
- FourCastNet has very similar surface-wind ACC over landmass and ocean.The appendix notes implications for wind-energy resource planning.