GraphCast: Graph Neural Networks for Global Medium-Range Weather Forecasting
New1. Overview
GraphCast is a global medium-range weather forecasting model from Google DeepMind, first posted as an arXiv preprint (2212.12794) in December 2022 and published on December 22, 2023 in Science, volume 382, issue 6677, pages 1416-1421, under the title "Learning skillful medium-range global weather forecasting". It has 18 authors, with Remi Lam first and Peter Battaglia last.
What it changes is the basic approach to medium-range forecasting. Traditional numerical weather prediction improves accuracy by adding compute resources, and the model itself does not directly use historical weather data. GraphCast takes the opposite route: it trains directly from reanalysis data, and one model emits hundreds of meteorological variables for the next 10 days globally at 0.25 degree resolution, with a full forecast completing in under one minute on a single accelerator.
The most cited claim, and the one that needs accurate reading, is the accuracy comparison. GraphCast significantly outperforms the most accurate operational deterministic system on 90 percent of 1380 verification targets. The comparison object is the ECMWF HRES operational deterministic system, scored separately per lead time and per variable, rather than a single aggregate accuracy claim. A vague statement about beating conventional numerical prediction on many metrics loses the two checkable numbers, 1380 and 90 percent.
Beyond average skill scores, the paper argues separately for improved severe event prediction, covering tropical cyclone tracks, atmospheric rivers, and extreme temperatures. This is the step that turns a lower average RMSE into a concrete statement about hazard scenarios, and it is what operational agencies actually evaluate.
The accompanying artifacts are complete. Model code and trained weights were first published in the deepmind/graphcast repository. That repository now redirects to google-deepmind/weathernext, which serves as the unified home for the WeatherNext family. GraphCast is retained there as WeatherNext 1 Graph, with file and checkpoint names keeping the original GraphCast naming for backward compatibility. This migration matters for users: old links still resolve, but documentation, installation, and weight licensing have all changed.
For engineers, GraphCast demonstrates a reproducible, verifiable pattern: treat reanalysis archives as supervised learning corpora, run autoregressive prediction on a non-structured spherical mesh with a graph neural network, and demote the numerical model to the role of initial-state provider. This "learning replaces integration" idea belongs to the same lineage as the surrogate modeling methodology discussed in the surrogate modeling and reduced-order modeling collection and the model taxonomy catalogued in the physics foundation models collection.
2. How It Works
The modeling target is the atmospheric state distributed over a sphere, discretized on an icosahedral mesh. This differs from the usual latitude-longitude grid: the icosahedral mesh approximates the sphere with triangular faces whose nodes are close to equal-area, avoiding the grid distortion that concentrates cells at high latitudes and giving more uniform sampling for the same node count.
Architecturally the model is a graph neural network. Nodes are grid points, edges connect neighbours, and messages propagate iteratively over the graph with the number of aggregation layers setting the receptive field. The model works autoregressively: given the atmospheric state at one time, it predicts the state 6 hours later, then feeds that prediction back as input for the next step and rolls forward to 10 days. This mirrors how a numerical model advances in time, so the model can be slotted into an existing forecasting pipeline.
Input variables cover the full ERA5 reanalysis state vector, including six scalar fields (geopotential, temperature, specific humidity, and the three wind components) plus atmospheric state variables describing the vertical structure, described in the paper as hundreds of meteorological variables. The number of verification targets reaches 1380 across multiple lead times and variables.
Training data is ERA5 reanalysis. The official documentation gives the exact scope: the main paper model trains on ERA5 from 1979 to 2017 at 0.25 degree resolution with 37 pressure levels. That training span bounds how well the model represents the atmosphere after 2018 and is the reason later fine-tuned variants exist.
Three pretrained checkpoints are provided, each with a clear purpose. GraphCast is the main paper model at 0.25 degree resolution, 37 pressure levels, trained on ERA5 from 1979 to 2017. GraphCast_small is a low-resolution version at 1 degree resolution, 13 pressure levels, and a smaller mesh, trained on ERA5 from 1979 to 2015, for running under tighter memory and compute constraints. GraphCast_operational is also 0.25 degree resolution but with 13 pressure levels, pretrained on ERA5 from 1979 to 2017 and fine-tuned on HRES data from 2016 to 2021, so it can be initialized directly from HRES and does not require precipitation inputs.
Several code details matter operationally. The file graphcast.py provides the GraphCast model architecture for a single prediction step. The file casting.py is a BFloat16 precision wrapper that lets GraphCast run at half precision to cut memory use. The file solar_radiation.py computes top-of-atmosphere incident solar radiation compatible with ERA5, which in an operational setting must be recomputed for each target lead time as a forcing variable. Shared library files covering autoregressive rollout, normalization, graph construction, loss computation, and JAX-compatible xarray utilities live in the utils directory and are reused across the model family.
The architectural trend is worth noting. GenCast and the later WeatherNext models all build on the graph mesh and autoregressive framework GraphCast established, treating GraphCast as the deterministic base case. Related entries in this catalogue include GenCast, Pangu-Weather, and FourCastNet, which differ in mesh type, attention mechanism, and whether they produce ensemble forecasts.
3. Scope and Assumptions
GraphCast's scope is bounded by five groups of premises.
First, it consumes reanalysis, not direct observation. ERA5 is a four-dimensional assimilation product built from observations and a numerical model, so it remains model output in substance. GraphCast learns the statistical regularities of that assimilation system, and its capability ceiling is bounded by reanalysis quality. Work that used GraphCast to identify errors in ERA5 2-metre temperature confirms that the model and its training data share the same origin.
Second, it performs medium-range deterministic forecasting, neither nowcasting nor climate prediction. Ten days at 0.25 degrees is its design envelope. Minute-to-hour precipitation nowcasting belongs to a different model class, as in the NowcastNet work. Seasonal and decadal scales are initial-value and coupling problems outside this model's reach, and the coupling route discussed in the NeuralGCM entry addresses a different physical layer.
Third, it does not issue operational warnings. The official disclaimer states plainly that GraphCast outputs have not been produced in collaboration with nor endorsed by any government meteorological agency or department, and in no way replace official alerts, warnings, or notices published by such agencies. This constraint must be respected: model output can feed internal decision pipelines but cannot serve as the published warning product.
Fourth, it is an atmosphere-only model without ocean, land surface, or chemistry. Operational systems such as HRES include the full coupled system, while GraphCast advances only the atmospheric state. On predictions that depend on air-sea coupling feedback, such as rapid intensification of tropical cyclones or storms over the open ocean, there is a structural gap relative to a complete model. Fine-tuning GraphCast_operational on HRES absorbs some operational-model information but does not change the physical scope.
Compute requirements form a fifth practical constraint. The official documentation states that non-Mini models require H100-class memory, Mini models can run inference on a P100, that the implementation is optimized for TPU, and that running on GPU requires switching the attention implementation. This directly determines whether a mid-sized team can run the model itself or must call a hosted service.
Licensing also needs a date qualifier. GraphCast weights were previously restricted to non-commercial use. On August 6, 2026 the repository updated the weight license to permit commercial use and replaced the previous terms with the new ones. Continuing to download or use the weights, including weights previously obtained from this repository, constitutes acceptance of the updated terms. Any statement about licensing status should carry that update date.
4. Getting Started
The easiest entry point is the graphcast_demo.ipynb notebook in the repository. It demonstrates the full sequence: loading initial weather data, generating random weights or loading a pretrained snapshot, running predictions, computing the loss, and computing gradients. Opening it directly in Google Colab avoids local environment setup.
Installation pins a released version from Git, and the official recommendation is to lock to a specific release:
pip install git+https://github.com/google-deepmind/weathernext.git@v0.3.0
The repository currently carries four releases: v0.1 from October 31, 2023, which is the original GraphCast release, v0.1.1 from October 9, 2024, v0.2 from March 30, 2026, and v0.3.0 from August 6, 2026. The README notes this is research code provided as is, with no guarantee of API stability and possible breaking changes in future updates, so pinning is the recommended practice.
Weights and sample data live in the Google Cloud Storage bucket dm_graphcast. The directory holds GraphCast data both at the bucket top level and under a graphcast subdirectory, the former kept for backward compatibility and slated for deletion in favour of the latter. Pointing at the graphcast subdirectory avoids future breakage.
Data preparation is on the user. Full training requires downloading ERA5 from ECMWF, best accessed as Zarr through WeatherBench2, and operational fine-tuning data is available from the WeatherBench2 HRES dataset. These datasets carry their own terms and conditions, and compliance must be confirmed before use.
Choosing among the three checkpoints: use GraphCast to reproduce the paper or for high-resolution research, GraphCast_small when compute is constrained or for a quick look before committing resources, and GraphCast_operational when initialization from the operational system and consistency with operational conventions matter. The fact that GraphCast_operational needs no precipitation input is significant in practice, because operational initial states generally do not carry precipitation.
For teams wanting to call the service rather than operate it themselves, the WeatherNext family offers direct forecast data feeds: on Google Cloud including Earth Engine, BigQuery, and Vertex AI, on the WeatherLab platform including cyclone tracks, and through OpenMeteo with an API and an interactive builder. That route removes both weight hosting and inference compute, which suits assessing business feasibility before deciding on self-hosting.
In broader physical AI terms, GraphCast is the canonical instance of the "replace the solver with a learned model" pattern. The fluid and structural analogues are discussed in the surrogate modeling and reduced-order modeling collection, and cross-scale model organization is covered in the neural operators collection.
5. Comparison with Alternatives
Placed within the AI weather model lineage, GraphCast clarifies the trade-offs each family makes.
The comparison with Pangu-Weather shows most clearly how architecture choice matters. Pangu-Weather (arXiv:2211.02556, November 2022) uses a 3D earth-specific autoregressive Transformer that factorizes spatial and temporal dimensions into windowed attention, achieving strong results at 0.25 degree resolution. GraphCast instead commits to graph message passing. The difference is not accuracy but how priors enter: the graph structure explicitly encodes adjacency on the spherical mesh, while a Transformer learns spatial association implicitly. Later work in the GraphCast lineage develops this into two distinct branches.
The relationship with GenCast is that of base and extension. GenCast (arXiv:2312.15796, later published in Nature volume 637, pages 84-90) inherits the graph mesh and autoregressive framework while replacing deterministic single-step prediction with diffusion-driven ensemble forecasting, which yields forecasts with uncertainty. GraphDOP (arXiv:2412.15687), also in this lineage, applies graph networks to coupled Earth system dynamics and brings ocean and atmosphere into the graph structure together.
FourCastNet (arXiv:2202.11214 and 2208.05419) represents another engineering line, replacing graph message passing with Fourier neural operators and emphasizing the operator-learning efficiency of adaptive Fourier transforms; its companion FourCastNeXt (arXiv:2401.05584) optimizes training for limited compute. The FuXi series (arXiv:2306.12873, published in npj Climate and Atmospheric Science volume 6) uses a cascade structure that decomposes global forecasting into multiple regional or multi-stage stages.
For precipitation nowcasting, DeepMind's NowcastNet and the generative radar extrapolation model target 0-to-6-hour heavy precipitation, so they do not compete with GraphCast at the 10-day scale. The original GenCast paper draws the distinction between deterministic and probabilistic forecasting clearly and also cautions about the difference in decision value between them.
Against traditional numerical models, a checkable reference is required. GraphCast compares against ECMWF HRES, and the paper cites the ECMWF evaluation literature for HRES as background. HRES itself evolves continuously, with ECMWF moving to a higher-resolution IFS configuration and introducing machine learning corrections after 2023, so the operational baseline has been rising throughout this period. Any claim of beating HRES therefore carries an explicit time coordinate, and cross-year citation must state the competing version.
It is worth stressing that AI weather models and numerical models are not in simple substitution. A growing body of work inserts AI models into data assimilation systems to improve analysis quality, and other work uses machine learning weather models to improve numerical prediction itself. Both directions indicate that AI models and physical models are forming complementary combinations rather than one-way replacement.
6. Benchmarks and Evidence
The paper's verifiable results fall into three groups: skill scores, event forecasting, and computational efficiency.
On skill scores, GraphCast significantly outperforms the most accurate operational deterministic system on 90 percent of 1380 verification targets. The essential element is that this is a proportion with a denominator, covering combinations of variables and lead times, rather than a single composite metric.
On event forecasting, the paper reports separate results for three severe event categories: tropical cyclone tracks, atmospheric rivers, and extreme temperatures, stating that its forecasts support better severe event prediction including these three. This part matters because it provides evidence sliced by event type, since a composite RMSE hides differences in capability across variables and regions.
On computational efficiency, a 10-day global 0.25 degree forecast takes under one minute on a single accelerator, against hours on a supercomputer for traditional numerical models. This comparison is among the most influential aspects of GraphCast and follows structurally from distilling 39 years of reanalysis into a single network rather than from constant-factor optimization.
Data scale is equally explicit: training uses ERA5 from 1979 to 2017 at 0.25 degree resolution with 37 pressure levels, and evaluation runs against HRES analysis and forecast data over periods from 2018 onward, which is the direct reason for choosing the 1979 to 2017 training window.
Compute requirements appear in the third section: official documentation states that non-Mini models need H100-class memory, Mini versions can run on P100, the implementation is optimized for TPU, and running on GPU requires switching the attention implementation.
Quantitative evidence in the ecosystem has accumulated heavily since publication. Downstream evaluation and reanalysis work includes a performance evaluation of GraphCast for medium-range forecasting (arXiv:2606.06348), a study of skill and systematic biases in Indian summer monsoon forecasts (arXiv:2607.11905), an evaluation of data-driven background ensemble covariances from GraphCast output (arXiv:2603.26703), a station-based evaluation of machine learning weather models (arXiv:2609.10564), and a mechanistic analysis of internal activations (arXiv:2512.24440). The UT-GraphCast hindcast dataset (arXiv:2506.17453) archives GraphCast hindcasts as a global AI forecast archive so third parties can run independent checks.
Derived directions include Solarcast-ML, a per-node GraphCast extension for solar energy production (arXiv:2406.13559), mechanistic interpretability of atmospheric rivers in GraphCast (arXiv:2610.07583), and GraphCast-Mamba, which learns non-Markovian corrections for weather forecasting (arXiv:2610.28512). Work combining GraphCast with operational systems includes efficient fine-tuning of the 37-level model onto the Canadian global deterministic prediction system (arXiv:2408.14587, published in Artificial Intelligence for the Earth Systems, DOI 10.1175/AIES-D-24-0101.1) and leveraging data-driven weather models for improving numerical weather prediction (arXiv:2407.06100, DOI 10.1175/WAF-D-24-0139.1). On data assimilation, there is a systematic exploration of machine learning weather models in that context (arXiv:2411.14677).
What these downstream efforts share is that they re-test GraphCast under third-party conventions. This layer of evidence says more about the model's position in a real operational chain than the paper's own scores do.
7. Role in Physics-AI Workflows
GraphCast's uses in a physics AI workflow fall into four classes.
The first is as a source of high-quality initial conditions and background fields. Numerical simulation and engineering analysis often need a high-quality atmospheric state. A fast high-resolution GraphCast forecast can serve as a rapid preprocessing step, or as a tool for generating ensemble perturbations that feed downstream CFD, transient analysis, and load assessment. This usage is not unusual in power and computing infrastructure contexts, where extreme weather impact assessment on power facilities requires high-resolution near-surface wind and temperature fields.
The second is as a surrogate model baseline. GraphCast demonstrates a complete methodological template: use physical model output as training data, train a learned model, and place the learned model back into the loop to accelerate iteration. Convergence of that loop depends entirely on the quality of the physical solver output and the model's tolerance for error. The same loop transfers to fluid and structural problems, as discussed in the surrogate modeling and reduced-order modeling collection.
The third is as a diagnostic probe for data quality. Work that used GraphCast to find errors in ERA5 2-metre temperature shows the model can serve as a probe for auditing reanalysis quality: when the learned model departs systematically from observations in a given region or variable, the signal likely points at a local defect in the training data or the assimilation system. This use of "inferring data from the model" tends to be underestimated in the early stages of data-driven methods.
The fourth is as a member of a multi-model ensemble. Deterministic GraphCast output can be combined with probabilistic GenCast output and conventional numerical forecasts to form multi-source ensembles for risk quantification. In decision settings that require probabilistic products, such as power dispatch, flood control, and shipping, such a combination delivers more decision value than a single deterministic product.
For concrete engineering decisions, several practical points apply. When evaluating, compare GraphCast against the same version of the ECMWF operational system and state both the competing version and the evaluation window, otherwise the conclusion is not reproducible. Prefer skill scores sliced by variable, region, and lead time over a single composite metric, since differences between model capability in tropical cyclones and heavy precipitation are far larger than average RMSE suggests. Remember that the output is a forecast field rather than a diagnostic quantity, and that physical consistency is inherited from the nature of the training data, so a downstream consumer needing strictly conservative fields may require extra handling. And for anything involving publicly issued warnings, go through official channels, treating model output strictly as internal input.
8. Limitations and Failure Modes
GraphCast's failure modes correspond directly to the sources of its capability, and listing them is more useful than a general caveat.
First, extrapolation beyond the training period. The model trains on ERA5 from 1979 to 2017, and there is no guarantee that later atmospheric states are well covered. Climate warming trends, changes in the observing system, and revisions to the reanalysis assimilation system can all induce drift in the model's behaviour on recent data. A study from January 2026 specifically uses GraphCast to identify a systematic error in ERA5 2-metre temperature, which is precisely the kind of risk being addressed.
Second, uneven capability across regions and variables. Global average skill hides significant local gaps. The Indian summer monsoon, a strongly seasonal regional system, was singled out for a dedicated bias study (arXiv:2607.11905), indicating identifiable systematic problems in that region. Station-based evaluations (arXiv:2609.10564) using near-surface observation conventions likewise often reveal conclusions that differ from gridded scores.
Third, initial-state dependence. GraphCast advances the state and nothing more, so initial-state error accumulates over autoregressive steps. Initial states typically come from operational analyses or reanalysis, so the model's ceiling is set jointly by initial-state quality and model drift. It is not an end-to-end forecasting system, and there is no shortcut around that.
Fourth, limited physical process coverage. Pure atmospheric stepping means ocean, land surface, snow and ice, and air-sea coupling feedback are all absent. Rapid intensification of tropical cyclones depends heavily on sea surface state, and that structural gap cannot be closed by fine-tuning. GraphCast_operational absorbs some operational-model information through HRES fine-tuning without changing the physical scope.
Fifth, boundary handling. The autoregressive rollout is valid within 10 days, but extending it to several weeks or seasonal horizons runs into out-of-distribution drift. The better-suited model for that regime is the coupling route discussed in the NeuralGCM entry.
Sixth, a comparison trap when benchmarking directly against operational models. HRES itself keeps evolving, with ECMWF moving to a higher-resolution IFS configuration and introducing machine learning corrections. There is also a subtlety in that GraphCast trains on ERA5 while being evaluated and initialized from HRES, and this training-versus-evaluation data combination is itself a detail that must be stated.
Seventh, precipitation and forcing details. GraphCast_operational explicitly requires no precipitation input, and the existence of solar_radiation.py is explained by the need to recompute top-of-atmosphere incident solar radiation for each target lead time in operational settings. These two details indicate that forcing-variable recomputation and input-field availability are genuine engineering problems in operational deployment, and that copying the paper configuration verbatim will fail.
Eighth, terms of use and endorsement limits. The weight license was updated on August 6, 2026 to permit commercial use, and previously downloaded weights are covered by the new terms, which needs to be communicated clearly within a team. The disclaimer also makes clear that model output carries no endorsement from any government meteorological agency and does not replace official warnings, a constraint that is easily overlooked in external communication.
9. Version and Source Notes
Several conventions should be checked before citing GraphCast.
The relationship between preprint and journal version must be kept straight. The arXiv preprint 2212.12794 was first submitted on December 24, 2022 and updated to v2 on August 4, 2023, titled "GraphCast: Learning skillful medium-range global weather forecasting". The Science version appeared on December 22, 2023 with the title "Learning skillful medium-range global weather forecasting", volume 382, issue 6677, pages 1416-1421, DOI 10.1126/science.adi2336, with 18 authors. Both refer to the same work under different titles, and the citation should pick whichever is appropriate.
Accuracy statements need exact numbers. "Significantly outperforms the most accurate operational deterministic system on 90 percent of 1380 verification targets" is the complete checkable phrasing, while "most variables and lead times" is a blurred version that should be replaced with the proportion and its denominator.
The repository migration path should be stated. Code was first hosted at deepmind/graphcast and now lives in google-deepmind/weathernext, where GraphCast is retained as WeatherNext 1 Graph with GraphCast naming preserved for backward compatibility. The old address redirects to the new one, so giving the new address directly is more accurate.
Model version conventions matter. The repository carries four releases: v0.1 from 2023-10-31, v0.1.1 from 2024-10-09, v0.2 from 2026-03-30, and v0.3.0 from 2026-08-06. The original paper version is v0.1, and subsequent releases are mainly adaptation and maintenance for GraphCast itself while introducing newer models such as WeatherNext 2 and WeatherNext Cyclones. Reproducing the paper means locking to v0.1 or the matching checkpoint.
The three checkpoints must be kept distinct. GraphCast at 0.25 degree with 37 pressure levels trained on 1979 to 2017 ERA5 is the paper model, GraphCast_small at 1 degree with 13 pressure levels and a smaller mesh trained on 1979 to 2015 ERA5 is the low-compute version, and GraphCast_operational at 0.25 degree with 13 pressure levels, pretrained on 1979 to 2017 ERA5 and fine-tuned on 2016 to 2021 HRES, initializable from HRES and needing no precipitation input, is the operational version. Their training data, resolution, level count, and initialization differ, and mixing them causes reproduction failures.
Licensing has an explicit date. Code and notebooks are Apache-2.0, other materials are CC BY 4.0, and the model weight license was updated on August 6, 2026 to permit commercial use, replacing the earlier non-commercial terms; continuing to download or use the weights, including ones previously obtained from this repository, constitutes acceptance. Any licensing statement should carry that update.
Storage path conventions: weights and sample data are in the GCS bucket dm_graphcast, where GraphCast data exists both at the bucket top level and under the graphcast subdirectory. The top level exists for backward compatibility and will be deleted, so later work should target the graphcast subdirectory.
Training and evaluation data conventions: the main model trains on ERA5 from 1979 to 2017 at 0.25 degree with 37 pressure levels, and evaluation runs on HRES data over periods from 2018 onward. The combination of reanalysis training and numerical-model evaluation is a precondition for reading the paper's conclusions.
Attribution conventions: the paper has 18 authors, Remi Lam first and Peter Battaglia last, with the team at Google DeepMind and Google Research. Official documentation acknowledges data relationships with external bodies such as ECMWF and states that those bodies bear no responsibility for the work and that users must confirm the applicable terms themselves.
10. References
- GraphCast code repository, now the unified WeatherNext family repository. https://github.com/google-deepmind/weathernext
- Lam R., Sanchez-Gonzalez A., Willson M., et al. GraphCast: Learning skillful medium-range global weather forecasting. arXiv:2212.12794, 2022. DOI: https://doi.org/10.48550/arXiv.2212.12794
- Lam R., Sanchez-Gonzalez A., Willson M., et al. Learning skillful medium-range global weather forecasting. Science, 382(6677), 1416-1421, 2023. DOI: https://doi.org/10.1126/science.adi2336
- WeatherNext 1 Graph (GraphCast) model documentation and checkpoint notes. https://github.com/google-deepmind/weathernext/tree/main/docs/weathernext1_graph
- GraphCast demonstration Colab notebook. https://github.com/google-deepmind/weathernext/blob/main/docs/weathernext1_graph/graphcast_demo.ipynb
- WeatherNext 2 technical report: Skillful joint probabilistic weather forecasting from marginals. arXiv:2506.10772, 2025. DOI: https://doi.org/10.48550/arXiv.2506.10772
- Alet F., Andersson T.R., Price I., et al. Operational tropical cyclone forecasting with AI. Nature, 2026. DOI: https://doi.org/10.1038/s41586-026-10953-2
- Price I., Sanchez-Gonzalez A., Alet F., et al. Probabilistic weather forecasting with machine learning. Nature, 637(8044), 84-90, 2024. DOI: https://doi.org/10.1038/s41586-024-08252-9
- Bi K., Xie H., Zhang H., et al. Pangu-Weather: A 3D high-resolution model for fast and accurate global weather forecast. arXiv:2211.02556, 2022. DOI: https://doi.org/10.48550/arXiv.2211.02556
- Chen L., Zhong X., Zhang F., et al. Accurate medium-range global weather forecasting with 3D neural networks. Nature, 619, 533-538, 2023. DOI: https://doi.org/10.1038/s41586-023-06185-3
- Pathak J., Subramanian P., Harrington P., et al. Learning to simulate complex physics with graph networks. arXiv:2202.07575, 2022. DOI: https://doi.org/10.48550/arXiv.2202.07575
- Kurth T., Subramanian P., Harrington J., et al. FourCastNet: Accelerating Global High-Resolution Weather Forecasting Using Adaptive Fourier Neural Operators. Proceedings of the Platform for Advanced Scientific Computing Conference, 2023. DOI: https://doi.org/10.1145/3592979.3593412
- Chen L., Zhong X., Zhang F., et al. FuXi: a cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6, 2023. DOI: https://doi.org/10.1038/s41612-023-00512-1
- Alexe M., Boucher E., et al. GraphDOP: Towards skilful data-driven medium-range weather forecasts learnt and initialised directly from observations. arXiv:2412.15687, 2024. DOI: https://doi.org/10.48550/arXiv.2412.15687
- Price I., Alet F., Sanchez-Gonzalez A., et al. GenCast: Diffusion-based ensemble forecasting for medium-range weather. arXiv:2312.15796, 2023. DOI: https://doi.org/10.48550/arXiv.2312.15796
- Kochkov A., Yuval J., Langmore I., et al. Neural general circulation models for weather and climate. Nature, 632, 1060-1066, 2024. DOI: https://doi.org/10.1038/s41586-024-07744-y
- Zhang S., Li Y., Liu X., et al. Skilful nowcasting of extreme precipitation with NowcastNet. Nature, 619, 526-532, 2023. DOI: https://doi.org/10.1038/s41586-023-06184-4
- Seasonal forecasting using the GenCast probabilistic machine learning model. Climate Dynamics, 64, 2026. DOI: https://doi.org/10.1007/s00382-026-08077-4
- Efficient fine-tuning of 37-level GraphCast with the Canadian global deterministic prediction system. Artificial Intelligence for the Earth Systems, 2024. DOI: https://doi.org/10.1175/AIES-D-24-0101.1
- Leveraging data-driven weather models for improving numerical weather prediction. Weather and Forecasting, 2024. DOI: https://doi.org/10.1175/WAF-D-24-0139.1
- UT-GraphCast Hindcast Dataset: A Global AI Forecast Archive from UT Austin. arXiv:2506.17453, 2025. DOI: https://doi.org/10.48550/arXiv.2506.17453
- Error in ERA5 2m Temperature identified using GraphCast. arXiv:2601.04701, 2026. DOI: https://doi.org/10.48550/arXiv.2601.04701
- GraphCast Skill and Systematic Biases in Indian Summer Monsoon Forecasts. arXiv:2607.11905, 2026. DOI: https://doi.org/10.48550/arXiv.2607.11905
- Mechanistic Interpretability of Atmospheric Rivers in GraphCast. arXiv:2610.07583, 2026. DOI: https://doi.org/10.48550/arXiv.2610.07583
- WeatherBench2 data guide covering ERA5 and HRES access. https://weatherbench2.readthedocs.io/en/latest/data-guide.html