FourCastNet: Global Weather Forecasting with Adaptive Fourier Neural Operators
New1. Overview
FourCastNet is a global data-driven weather forecasting model proposed by NVIDIA, NERSC, and partners in 2022, short for Fourier Forecasting Neural Network. Its preprint, "FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators", was posted on February 22, 2022 as arXiv 2202.11214 with thirteen authors and Jaideep Pathak listed first.
Two consequences made it influential. The first is accuracy placement: at short lead times on large-scale variables, FourCastNet matches the ECMWF Integrated Forecasting System, the operational numerical weather prediction system; on variables with complex fine-scale structure, notably precipitation, it outperforms IFS. The second is speed: a week-long global forecast is produced in under two seconds, orders of magnitude faster than IFS. Combined, these turn generating a thousand-member probabilistic ensemble from an expensive undertaking into a routine operation.
The downstream value of that speed is often underrated. The paper notes that operational numerical weather prediction ensembles are limited by compute cost to roughly fifty members, while data-driven models can cheaply produce thousand-member ensembles and thereby give better constrained uncertainty estimates for extreme events, improving the reliability of early warnings. This matters most for tropical cyclones, extra-tropical cyclones, and atmospheric rivers.
The accompanying open-source repository is NVlabs/FourCastNet, created on July 7, 2022, with master as the default branch and Python as the main language, described as the initial public release of FourCastNet code, data, and model weights. As of November 2026 it carries roughly 705 stars, 158 forks, and 19 open issues. Activity peaked between 2022 and 2023, after which the main branch has largely been static, which matters when planning use and citation.
One common confusion deserves early clarification. FourCastNet and the later FuXi, Pangu-Weather, GraphCast, GenCast, and NeuralGCM belong to different branches of the same technical lineage rather than a single succession. Its distinctive legacy is establishing the operator-learning plus vision-transformer route together with thousand-member ensembles, and it provided the direct starting point for the later spherical operator line.
2. How It Works
The central component is the Adaptive Fourier Neural Operator, which performs global token mixing, while the backbone is borrowed from the vision transformer. The combination traces to the AFNO paper by Guibas and colleagues from 2021 and to the official implementation repository NVlabs/AFNO-transformer.
AFNO rests on re-framing token mixing as a continuous global convolution that carries no dependence on input resolution, carried out in the Fourier domain to avoid the quadratic cost of self-attention as pixel counts grow. Compared with the original FNO, AFNO additionally addresses discontinuities and high-resolution inputs encountered in visual representation learning. This is what allows FourCastNet to run directly on a 0.25-degree global grid instead of a regularized Cartesian domain.
The data flow is specific. Input variables on a 720 by 1440 lat-lon grid are projected onto an (h by w) grid of patches, each patch a d-dimensional token with patch size p taken as 8; the token sequence then enters a series of AFNO layers together with a positional encoding, and a linear decoder reconstructs the patches of the next frame. The time step is fixed at six hours.
Model size and training hyperparameters appear in the paper tables. The number of AFNO blocks is 8, depth is 12, MLP ratio is 4, embedding dimension is 768, activation is GELU, dropout is 0, the sparsification threshold lambda is 1e-2, patch size is 8 by 8, and the learning rate is 2.5e-4 with cosine scheduling.
Training proceeds in two stages. Pre-training learns the mapping from X(k) to X(k+1) by supervised learning, and fine-tuning then starts from that pretrained model to further train the output layer. Precipitation is handled separately: the main model forecasts twenty variables while a secondary diagnostic model supplies precipitation from six-hour accumulated totals. Inference proceeds autoregressively, with the model initialized from a 2018 sample and running freely forward.
Twenty prognostic variables span five vertical levels. Surface quantities include the two components of 10-meter wind, 2-meter temperature, surface pressure, and mean sea level pressure. The 1000 hPa and 850 hPa levels carry geopotential height with related wind, temperature, and humidity components; the 500 hPa level carries temperature, wind, and relative humidity; the 50 hPa level carries geopotential height; and there is integrated total column water vapour. Relative humidity is among the variable families whose error accumulates fastest in the paper.
The paper also provides a magnitude comparison against IFS. IFS was developed over decades, contains more than 150 variables at more than 50 vertical levels, and is guided by physics; FourCastNet models 20 variables at five vertical levels and is purely data driven. The comparison illustrates the model's potential while marking the boundary of its coverage.
3. Scope and Assumptions
Several boundaries should be established before use.
The accuracy horizon is explicitly divided. Measured by RMSE and ACC, FourCastNet predictions are comparable to IFS for lead times up to three days. After that and out to a week, all modelled variables lag somewhat behind IFS. The appendix gives latitude-weighted ACC per variable family, with values generally above 0.6 across five to ten days, and the relative humidity family accumulating error fastest.
Ensemble figures require careful reading of the wording. The hurricane case uses a 100-member ensemble, produced by perturbing the initial condition with Gaussian noise into 100 perturbed initial conditions. The paper body also argues about the value of thousand-member ensembles, which serves as capability reasoning and outlook, while the published weights and official scripts correspond to the hundred-member configuration in practice. Conflating the hundred-member case with a thousand-member configuration overstates what is actually available.
Energy numbers also carry conditions. The paper states that the energy required to train FourCastNet is approximately equal to that required to generate one 10-day forecast with 50 IFS ensemble members, and that afterwards a single forecast consumes about 12000 times less energy than IFS. The 18-kilometer figures are extrapolations obtained by interpolating the 30-kilometer results, because no public data at that resolution existed, and should be cited separately from measured numbers. In the node-level comparison, IFS needs 3060 nodes to compute a 100-member 24-hour ensemble forecast, while FourCastNet needs one four-GPU A100 node at the measured 30-kilometer resolution and two nodes at the extrapolated 18-kilometer resolution.
Data and compute access form real constraints. The training data is hosted by NERSC and shared through Globus, requiring a Globus account and login; the official documentation puts the full training dataset for this version at roughly 5TB. The data is organized as hdf5 files per year, with the training set spanning 1979 to 2015, the test set covering 2016 and 2017, and 2018 held out as the out-of-sample set. Normalization statistics are provided separately, and the README explicitly warns that the statistics and the pretrained weights must be used together, because that is what the model was trained with.
On hardware and environment, the training scripts target a Slurm-based HPC cluster, with the official example being a distributed data parallel launch on NERSC's Perlmutter.
4. Getting Started
Inference is the actual entry point for most users, and the path is relatively clear.
Three things must be prepared first. The out-of-sample initial condition file, which may be the hosted out_of_sample/2018.h5 or a self-processed hdf5 downloaded from the Copernicus Climate Data Store; the model weights, comprising the backbone.ckpt and precip.ckpt files; and the normalization statistics in stats_v0, containing global means, global standard deviations, the land-sea mask, latitude, longitude, and time means.
Paths are then set in config/AFNO.yaml. The configuration needs the orography toggle and orography_path, the experiment output directory exp_dir, the train, valid, and inference data paths, and the three statistics paths for time means, global means, and global standard deviations.
There are two inference commands. Main variables use inference/inference.py and precipitation uses inference/inference_precip.py, each taking a configuration section name via --config, a weight file via --weights, and an output directory via --override_dir. The output is an hdf5 file containing the computed ACC and RMSE as configured, plus raw fields for visualization.
Separate scripts handle ensemble inference. inference_ensemble.py processes wind variables and inference_ensemble_precip.py processes total precipitation, with submit_batch_ensemble.sh used to submit ensemble members across different initial conditions in parallel to the cluster.
The README documents custom-interval forecasting as its own section. First use the copernicus scripts get_data_pl_short_length.py and get_data_sfc_short_length.py to download pressure-level and single-level variables, downloading consecutive days only and keeping snapshots at the 0, 6, 12, and 18 hour marks. Then preprocess with data_process/parallel_copy_small_set.py, which supports MPI for long intervals and runs in a single process for short ones. Finally run inference following the earlier steps.
Training configuration is centralized in config/AFNO.yaml, with interactive_train.sh and submit_batch.sh provided for interactive runs and cluster batch submission. A docker directory is included for containerized environments.
5. Comparison with Alternatives
Placing FourCastNet among contemporaneous and later data-driven forecasting models clarifies what each contributes.
Against Pangu-Weather, which follows a 3D Earth-specific Transformer route encoding three-dimensional vorticity and divergence fields into a latent space, the parameter scale and the treatment of physical constraints differ. Both reported beating ECMWF high-resolution forecasts in their own time, but the number of modelled variables and the physical-encoding approach are clearly distinct.
Against FuXi, which uses cubic convolutions and axial attention on a cubed-sphere grid and reported broader superiority over ECMWF for medium-range forecasting, the difference can be summarized as follows: FourCastNet mixes in the Fourier domain, Pangu uses a physically encoded three-dimensional transformer, FuXi works on a cubed-sphere grid with convolutions, and GraphCast operates a GNN over a graph mesh.
The comparison with GraphCast deserves separate mention. GraphCast uses multiscale rotation-equivariant graph networks on a 0.25-degree grid and reports significant improvement over the ECMWF operational deterministic system on 90 percent of 1380 verification targets. It reaches further on accuracy and initial-condition dependency, while FourCastNet compares against IFS on large-scale variables at short lead times. The temporal relation is that GraphCast drew on FourCastNet's evaluation framework and operator-learning perspective rather than deriving from its code.
The spherical line was advanced by the same team. Spherical Fourier Neural Operators generalize Fourier mixing to the sphere, and the Huge Ensembles series, with sixteen authors in the first part and fifteen in the second, systematize large-ensemble forecasting for studying the tail probabilities of tropical cyclones, extreme precipitation, and extreme temperature.
On the engineering side the continuation is FourCastNeXt, an optimization of FourCastNet that keeps comparable accuracy while reducing training compute to roughly five percent of the original, aimed at making neural Earth system modelling ablations affordable, described in arXiv 2401.05584.
6. Benchmarks and Evidence
The accuracy evaluation framework is among the best specified parts of the paper. The core metrics are RMSE and the anomaly correlation coefficient, computed per initial condition and per lead time, and evaluated separately for control and mean forecasts so that the additional skill from ensembling can be isolated. Ensemble evaluation further compares the FourCastNet ensemble mean, the unperturbed FourCastNet control, and the unperturbed IFS control.
Per-variable-family ACC results give a finer picture. Surface wind, geopotential height, and temperature differ markedly in behavior, and the relative humidity family accumulates error fastest. This is a practical point for users: workloads sensitive to humidity need shorter lead times.
The extreme-event case analysis gives concrete evidence. In the typhoon case the model starts near 10 degrees north and 210 degrees west and captures the storm's intensification and track over four days. Three named hurricanes form in the Atlantic during the case, namely Florence, Issac, and Helene, approaching the eastern coast of North America. Hurricane eye position is estimated from mean sea level pressure, with the minimum pressure taken at the eye.
The compute and efficiency comparison comes from the PASC conference paper. It reports that FourCastNet was optimized and scaled efficiently on three supercomputing systems, namely Selene, Perlmutter, and JUWELS Booster, using up to 3,808 NVIDIA A100 GPUs and reaching 140.8 petaFLOPS in mixed precision, about 11.9 percent of peak at that scale. The time-to-solution for training measured on JUWELS Booster with 3,072 GPUs is 67.4 minutes, roughly 80000 times faster than state-of-the-art numerical weather prediction in inference. The paper also notes that without public 18-kilometer data, those figures are extrapolated from the 30-kilometer results by interpolation.
Training resources amount to about 16 hours of wall-clock time for end-to-end training on a cluster of 64 NVIDIA A100 GPUs.
The data foundation is ERA5, which provides hourly estimates of several atmospheric variables at 0.25-degree latitude and longitude resolution from the surface to roughly 100 km altitude, from 1979 to the present. The model selects a 20-variable subset and subsamples at six-hour intervals, sampling each day at 0000, 0600, 1200, and 1800. Normalization applies specific dimensional transforms to quantities such as temperature and geopotential height, and precipitation is first converted to accumulation and then distribution-transformed to reduce skew; when compared against IFS or ERA5 ground truth, precipitation is transformed back to units of length.
Weights are offered through two download paths, Globus and the NERSC project portal, with a January 2023 commit adding the portal link for model weights.
7. Role in Physics-AI Workflows
FourCastNet can enter an actual workflow at three levels.
The first level is serving as a fast surrogate. Compressing a one-week forecast into seconds makes large-scale scanning over a design space possible. Wind energy resource planning is one case, where the reliability and economics of near-surface wind forecasts directly support siting of onshore and offshore wind farms. Another is generating many perturbed initial conditions across candidate schemes and making decisions on statistics rather than on a single trajectory.
The second level is the uncertainty expression that ensembles enable. A 100-member ensemble, with hurricane track and intensity estimated from mean sea level pressure, yields probability distributions over track and intensity that a single deterministic forecast cannot provide. For warning systems, the reliability of that distribution matters more directly to usefulness than point accuracy.
The third level is leaving the physics inside the pipeline. This model family does not write conservation laws into a loss function as physics-informed neural networks do; it learns the mapping directly from reanalysis data and relies on data assimilation for initial conditions. It occupies the same problem domain as discussions of physics foundation models and surrogate modeling, but within meteorology it developed along independent lines that produced GraphCast, Pangu-Weather, NeuralGCM, and GenCast among others.
A pragmatic combination is to use these fast models for screening and initial-condition generation, then confirm with first-principles or high-resolution numerical models. The speed of the learned model makes the number of candidates that reach the confirmation stage tractable.
Research uses also include dataset construction and reanalysis. The Huge Ensembles series uses large hindcast sets generated by SFNO as a basis for studying the statistical properties of extreme events, and a 2025 work further quantifies extreme precipitation and temperature distributions from machine-learning-based ensembles.
8. Limitations and Failure Modes
Four usage patterns produce unreliable results.
Lead times beyond three days require care. The paper itself states that between three days and one week the model's skill lags IFS. Using FourCastNet output for outlooks beyond ten days, or asserting parity with operational model accuracy, goes beyond its validated range.
Humidity-dependent quantities need their own horizon. Because the relative humidity family accumulates error fastest, products depending on humidity should be designed for shorter lead times.
Variable coverage is a hard boundary. The model covers 20 variables at five vertical levels and omits clouds, radiation, soil temperature and moisture, and boundary layer structure, all of which operational models require. IFS carries more than 150 variables at more than 50 vertical levels. Applications needing complete model fields cannot be served here. Precipitation comes from a secondary diagnostic model, so it is not a consistent output of the same model as the main variables and cross-variable consistency requires separate checking.
Differences in ensemble framing create misjudgments. The paper's ensemble perturbs the initial condition with Gaussian noise, a coarse characterization of initial-condition uncertainty compared with how operational systems treat multi-source observation error.
The engineering failure modes are equally concrete. Normalization statistics and weights must be used together, and substituting self-computed statistics directly corrupts the model output. The data interface imposes firm requirements on variable order, timestamps, and file organization, where misordering raises no error yet invalidates results. Weight download depends on Globus login, which is a practical barrier in restricted network environments. The main branch has had essentially no new commits since the initial release in August 2022 and the added weight link in January 2023, and expectations for issue response should be adjusted accordingly.
The ensemble method itself carries known limits. Because it perturbs initial conditions with Gaussian noise, that construction is a rough proxy for initial-condition uncertainty.
9. Version and Source Notes
A few version and framing details should be pinned down to avoid mixing them in citations.
The repository license field is inconsistent. The GitHub API reports NOASSERTION, while the LICENSE file at the repository root is the complete BSD 3-Clause text with copyright marked as Copyright (c) 2022, FourCastNet authors. Citations should follow the LICENSE file, since the API field is not a license conclusion.
On code status, the repository carries a single tag and release, both v0.0.0, published on August 6, 2022. There are two branches, master and feature-gfs-initialization, the latter exploring global forecast system initialization. The main-branch commit history has three entries, the most recent dated January 13, 2023 and adding a portal download link for the model weights. The repository push timestamp stops at October 2023 with no later update, and this state must be weighed when assessing usability in 2026.
The papers come in preprint and conference versions. The main paper is arXiv 2202.11214 with thirteen authors, Pathak first, posted February 22, 2022. The supercomputing and efficiency work appeared at the PASC 2023 conference as "FourCastNet: Accelerating Global High-Resolution Weather Forecasting Using Adaptive Fourier Neural Operators", with nine authors, Kurth first, published June 26, 2023, DOI 10.1145/3592979.3593412, with preprint arXiv 2208.05419.
The data version is a single v0. The README version table has one row, the initial release v0.0.0, pairing FCN_ERA5_data_v0 with FCN_weights_v0 and stats_v0. No other data or weight releases exist.
At least five sets of numbers require separation. The inference figure of a week-long forecast in under two seconds versus the roughly 80000-fold inference speedup in the PASC paper. The hundred-member ensemble used in the case versus scalability to a thousand members. The training energy equal to one 50-member 10-day IFS forecast versus inference energy at about 12000 times lower. The measured 30-kilometer resolution versus the extrapolated 18-kilometer estimate. And the 67.4-minute training time-to-solution on 3,072 GPUs versus 140.8 petaFLOPS on 3,808 A100 GPUs at 11.9 percent of peak. These come from different papers under different framings, and mixing them produces order-of-magnitude misunderstandings.
Downstream repository licenses also need individual confirmation. NVlabs/AFNO-transformer, the official PyTorch implementation of AFNO, likewise reports NOASSERTION through the GitHub API, with its last push in November 2022. nci/FourCastNeXt declares Apache-2.0 and was last pushed in March 2025.
10. References
- Pathak J., Subramanian S., Harrington P., et al. FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators. arXiv:2202.11214, 2022. https://arxiv.org/abs/2202.11214
- Kurth T., Subramanian S., Harrington P., et al. FourCastNet: Accelerating Global High-Resolution Weather Forecasting Using Adaptive Fourier Neural Operators. Proceedings of the Platform for Advanced Scientific Computing Conference, 2023. https://doi.org/10.1145/3592979.3593412
- Kurth T., Subramanian S., Harrington P., et al. FourCastNet: Accelerating Global High-Resolution Weather Forecasting using Adaptive Fourier Neural Operators (preprint). arXiv:2208.05419, 2022. https://arxiv.org/abs/2208.05419
- NVIDIA and NERSC. NVlabs/FourCastNet source repository (code, weights, and data access notes). https://github.com/NVlabs/FourCastNet
- Guibas J., Mardani M., Li Z., et al. Adaptive Fourier Neural Operators: Efficient Token Mixers for Transformers. arXiv:2111.13587, 2021. https://arxiv.org/abs/2111.13587
- NVlabs. AFNO-transformer official PyTorch implementation. https://github.com/NVlabs/AFNO-transformer
- Pathak J., Kurth T., Bonev B., et al. Spherical Fourier Neural Operators: Learning Stable Dynamics on the Sphere. arXiv:2306.03838, 2023. https://arxiv.org/abs/2306.03838
- Pathak J., Kurth T., Xu R., et al. Huge Ensembles Part I: Design of Ensemble Weather Forecasts using Spherical Fourier Neural Operators. arXiv:2408.03100, 2024. https://arxiv.org/abs/2408.03100
- Pathak J., Kurth T., Xu R., et al. Huge Ensembles Part II: Properties of a Huge Ensemble of Hindcasts Generated with Spherical Fourier Neural Operators. arXiv:2408.01581, 2024. https://arxiv.org/abs/2408.01581
- Guo E., Ahmed M., Sun Y., et al. FourCastNeXt: Optimizing FourCastNet Training for Limited Compute. arXiv:2401.05584, 2024. https://arxiv.org/abs/2401.05584
- Hersbach H., Bell B., Berrisford P., et al. ERA5 hourly data on pressure levels from 1959 to present. Copernicus Climate Change Service (C3S) Climate Data Store, 10.24381/cds.bd0915c6. https://doi.org/10.24381/cds.bd0915c6
- Hersbach H., Bell B., Berrisford P., et al. ERA5 hourly data on single levels from 1959 to present. Copernicus Climate Change Service (C3S) Climate Data Store, 10.24381/cds.adbb2d47. https://doi.org/10.24381/cds.adbb2d47
- Lam R., Sanchez-Gonzalez A., Willson J., et al. GraphCast: Learning skillful medium-range global weather forecasting. Science, 382(6677), 2023. https://doi.org/10.1126/science.adi2336
- Bi K., Chen Y., Wang F., et al. Accurate medium-range global weather forecasting with 3D neural networks. Nature, 619, 533-538, 2023. https://doi.org/10.1038/s41586-023-06185-3
- Chen L., Zhong X., Zhang F., et al. FuXi: a cascade machine learning forecasting system for 15-day global weather forecast. npj Computational Materials, 9, 2023. https://doi.org/10.1038/s41612-023-00512-1
- Chen L., Zhong X., Zhang F., et al. FuXi preprint. arXiv:2306.12873, 2023. https://arxiv.org/abs/2306.12873
- Zhang C., Yao Z., Yang H., et al. Pangu-Weather: A Global Data-driven High-resolution Weather Model. arXiv:2211.02556, 2022. https://arxiv.org/abs/2211.02556
- Kochkov A., Yuval J., Langmore I., et al. Neural general circulation models for weather and climate. Nature, 632, 1060-1066, 2024. https://doi.org/10.1038/s41586-024-07744-y
- Price I., Sanchez-Gonzalez A., Alet L., et al. GenCast: Probabilistic weather forecasting with machine learning. Nature, 637(8044), 84-90, 2024. https://doi.org/10.1038/s41586-024-08252-9
- Zhang Y., Pathak J., Wellbrock O., et al. Skilful nowcasting of extreme events with the neural operator. Nature, 619, 526-532, 2023. https://doi.org/10.1038/s41586-023-06184-4
- nci. FourCastNeXt training and inference code. https://github.com/nci/FourCastNeXt
- NERSC. FourCastNet pretrained model weights portal (FCN_weights_v0). https://portal.nersc.gov/project/m4134/FCN_weights_v0/