NOAA’s Strategic Integration of Artificial Intelligence in Global Numerical Weather Prediction

Abstract

The operationalization of artificial intelligence (AI) within the National Oceanic and Atmospheric Administration (NOAA) represents a paradigmatic inflection point in the history of environmental prediction. In late 2025, NOAA formally deployed a new suite of forecast systems: the Artificial Intelligence Global Forecast System (AIGFS), the Artificial Intelligence Global Ensemble Forecast System (AIGEFS), and the pioneered Hybrid Global Ensemble Forecast System (HGEFS). This report provides an exhaustive, multi-dimensional analysis of this transition, synthesizing technical architectures, operational strategies, and validation metrics. We examine the underlying Graph Neural Network (GNN) technologies—derived principally from Google DeepMind’s GraphCast and adapted via Project EAGLE—and their integration into the "Anemoi" framework. By analyzing the massive computational efficiency gains (99.7% reduction in resource usage) alongside critical limitations regarding "grey swan" events and spectral variance, this document elucidates the strategic rationale behind the hybrid modeling approach. Furthermore, we explore the downstream implications for the renewable energy sector, the evolving role of human forecasters, and the future trajectory of Earth system modeling where data-driven induction and physical deduction converge.

1. Introduction: The Saturation of the Silicon Atmosphere

For nearly a century, the discipline of meteorology has been defined by a singular, computationally intensive paradigm: Numerical Weather Prediction (NWP). Since Lewis Fry Richardson first envisioned his "forecast factory" in 1922—a stadium of human computers solving differential equations by hand—the field has relentlessly pursued higher resolution and physical fidelity through the brute force of supercomputing. The operational implementation of the first barotropic models on the ENIAC in 1950 initiated an era where forecast skill was linearly correlated with processor speed. To improve the forecast, one simply refined the grid and shortened the time step, solving the Navier-Stokes equations for momentum, the thermodynamic energy equation, and the continuity equation for mass conservation with increasing precision.

However, as the 21st century progressed, this physics-based paradigm began to encounter a formidable asymptote: the slowing of Moore’s Law and the exorbitant energetic and financial costs of exascale computing. The "Computational Wall" became a tangible barrier. Doubling the resolution of a global model requires an approximate eight-fold increase in computational power, as grid points multiply in latitude, longitude, and time. Operational centers like NOAA and the European Centre for Medium-Range Weather Forecasts (ECMWF) found that gaining an additional day of forecast skill required investments in High-Performance Computing (HPC) that were becoming increasingly difficult to sustain.

Into this stagnation arrived the "Second Revolution" of Data-Driven Weather Prediction (DDWP). Unlike NWP, which simulates the atmosphere based on first principles of fluid dynamics, DDWP systems leverage deep learning to infer the laws of atmospheric evolution directly from historical data. These models, trained on decades of reanalysis datasets like ERA5, treat weather forecasting not as a physics problem, but as a geometric transformation problem—mapping the state of the atmosphere at time t to time t+Δt through the optimization of millions of weights in a neural network.

The deployment of NOAA’s new suite—AIGFS, AIGEFS, and HGEFS—marks the first time a major national weather service has fully operationalized this technology alongside its flagship physics models.¹ This is not merely a technical upgrade; it is an epistemological shift from deterministic simulation to probabilistic inference. This report details how NOAA utilized Project EAGLE to bridge the gap between Silicon Valley innovation and government operations, creating a hybrid forecasting ecosystem that aims to combine the stability of AI with the physical rigour of traditional dynamics.

2. The Theoretical Framework: From Grids to Graphs

To understand the magnitude of NOAA’s deployment, one must first dissect the technological architecture that underpins it. The transition from NWP to AI represents a move from discretization of continuous equations to the learning of discrete manifolds.

2.1 The Limitations of Conventional CNNs

The initial wave of AI weather forecasting utilized Convolutional Neural Networks (CNNs), treating global weather fields as 2D images. While effective for regional forecasting, standard CNNs struggle with global geometry. The Earth is a sphere, not a rectangle. Projecting the globe onto a 2D latitude-longitude grid introduces severe distortions, particularly at the poles (the singularity problem), where grid cells become infinitesimally small. This distortion forces CNNs to learn location-dependent filters, reducing their efficiency and ability to generalize atmospheric features (like cyclones) that should look the same regardless of where they occur.

2.2 NOAA's Implementation of The GraphCast Architecture and the Multi-Mesh Approach

The engine driving NOAA’s AIGFS is based on GraphCast, a Graph Neural Network (GNN) architecture developed by Google DeepMind. GraphCast abandons the rigid lat-lon grid in favor of a Multi-Scale Icosahedral Mesh.

This architecture addresses the spherical geometry problem directly. The Earth is represented as a graph—a set of nodes (locations) connected by edges (relationships).

• The Encoder: The process begins by mapping the input data—which typically comes in a standard 0.25 degree latitude-longitude grid from the Global Data Assimilation System (GDAS)—onto the graph nodes. This embedding process translates physical variables (temperature, wind, humidity) into a latent feature vector at each node.³

• The Processor (Message Passing): This is the computational core. Through multiple layers of "message passing," nodes exchange information with their neighbors. The "Multi-Scale" aspect is critical here. The mesh contains edges of varying lengths. Short edges allow the model to capture local interactions (e.g., convective initiation), while long edges allow for the rapid transmission of information across the hemisphere (e.g., the propagation of Rossby waves or teleconnections like El Niño). This capacity to model long-range dependencies efficiently is why GNNs outperform CNNs in global forecasting.⁵

• The Decoder: Finally, the updated latent state of the graph is mapped back to the 0.25 degree grid to produce the forecast for the next time step (typically 6 hours forward).

This autoregressive process—predicting t+6h, then using that prediction to predict t+12h—allows the model to generate 10-day forecasts in under one minute on a single Tensor Processing Unit (TPU) or GPU.⁷ This speed contrasts sharply with the operational GFS, which requires hundreds of CPU cores running for nearly an hour to produce a similar duration forecast.

3. Project EAGLE: The Operational Bridge

The integration of such radically different technology into a government operational environment is fraught with challenges, often referred to as the "Valley of Death" in Research to Operations (R2O). To navigate this, NOAA established Project EAGLE (Experimental AI Global and Limited-area Ensemble), a collaborative sandbox involving NOAA’s research laboratories, the National Weather Service (NWS), and the Earth Prediction Innovation Center (EPIC).⁸

3.1 The Strategic Mandate

Project EAGLE was not designed simply to "test" AI, but to build the pipeline for its continuous deployment. The operational cadence of traditional NWP is slow; major dynamical core upgrades happen every few years. AI innovation, driven by the private sector, moves on a weekly basis. EAGLE provides a parallel operational track, allowing NOAA to ingest new architectures (like GraphCast, or potential successors like NeuralGCM or Aurora) and validate them against trusted metrics without disrupting the legacy GFS workflow.²

3.2 The Anemoi Framework: Standardization of the Stack

A critical, often overlooked enabler of this deployment is the adoption of the Anemoi framework. Developed initially by ECMWF in collaboration with national meteorological services across Europe, Anemoi is an open-source, Python-based toolkit designed specifically for Machine Learning Weather Prediction (MLWP).¹¹

NOAA’s adoption of Anemoi signals a major shift toward international software interoperability. The framework consists of modular components:

• Anemoi-Datasets: This module handles the immense complexity of meteorological data formats (GRIB2, NetCDF, Zarr). It was essential for converting NOAA’s petabytes of GDAS and GEFS data into the tensor formats required for AI training.¹²

• Anemoi-Graphs: This component constructs the mesh representations of the atmosphere. It supports the custom generation of the multi-scale graphs used by GraphCast, allowing NOAA researchers to experiment with different mesh densities without rewriting the core model code.¹²

• Anemoi-Training: A configuration-driven training loop that allowed NOAA to fine-tune the pre-trained models.

By leveraging Anemoi, NOAA avoided the cost of building a proprietary ML infrastructure, instead utilizing a "community standard" that ensures their models are compatible with developments occurring in Europe and academia.⁹

3.3 The Data Challenge: ERA5 vs. GDAS

One of the most significant technical hurdles Project EAGLE overcame was the Data Domain Shift. Most commercial AI models (GraphCast, Pangu-Weather, FourCastNet) are pre-trained on ERA5, the reanalysis dataset produced by ECMWF. ERA5 is considered the "gold standard" of historical weather reconstruction.

However, NOAA operations do not run on ECMWF data; they run on GDAS (Global Data Assimilation System) initial conditions. There are subtle but significant differences between the "climate" of ERA5 and GDAS due to different underlying physics packages and assimilation schemes. Running a model trained on ERA5 using GDAS initial conditions results in "initialization shock"—the model perceives the initial state as slightly "wrong" or noisy, leading to errors in the first few forecast steps.¹

To solve this, Project EAGLE scientists performed a massive Transfer Learning operation. They took the weights of the Google-trained GraphCast model and fine-tuned them using historical GDAS analysis data. This process adjusted the model's internal representations to align with the specific statistical properties of NOAA’s data stream, significantly improving performance for the operational 00Z and 12Z cycles.¹

4. System Architecture 1: The AIGFS (Deterministic)

The Artificial Intelligence Global Forecast System (AIGFS) serves as the deterministic flagship of the new suite. It is the direct counterpart to the operational GFS, providing a single, high-resolution trajectory of the future atmosphere.

4.1 Specifications and Resolution

The AIGFS operates on a 0.25 degree global grid, translating to a horizontal resolution of approximately 28 kilometers at the equator.¹³ This matches the resolution of the standard GFS output grids, ensuring seamless integration into downstream tools.

• Vertical Resolution: The model outputs data on 13 isobaric (pressure) levels, ranging from the surface (1000 hPa) up to the stratosphere (50 hPa). While this is fewer than the 64+ native levels of the physics-based GFS, it covers the entirety of the troposphere and lower stratosphere, capturing the key dynamics relevant for surface weather.¹³

• Temporal Resolution: Forecasts are generated at 6-hour time steps out to 10 days (240 hours) and potentially further for experimental runs.

4.2 Computational Economics

The most striking feature of the AIGFS is its efficiency. NOAA reports that AIGFS uses 99.7% less computing resources than the physics-based GFS.¹

• Implications: This efficiency unlocks the ability to run the model on significantly smaller hardware footprints. While the GFS requires a partition of the massive WCOSS (Weather and Climate Operational Supercomputer System) supercomputer, AIGFS can run on a cluster of GPUs. This frees up the supercomputer to focus on the computationally expensive Data Assimilation (DA) process or to run higher-resolution localized models (like the Hurricane Analysis and Forecast System).

• Latency: The inference time for a 10-day forecast is measured in minutes. This allows forecasters to receive guidance almost immediately after the initial conditions are available, rather than waiting nearly an hour for the numerical integration to complete.¹

5. System Architecture 2: The AIGEFS (The AI Ensemble)

While the deterministic AIGFS offers speed, the Artificial Intelligence Global Ensemble Forecast System (AIGEFS) offers rigorous uncertainty quantification. In modern meteorology, a single forecast is of limited value; knowing the range of possibilities is essential.

5.1 Ensemble Theory and Perturbation

The atmosphere is a chaotic system; microscopic errors in the initial state grow exponentially over time (the Butterfly Effect). Ensemble forecasting addresses this by running the model multiple times with slightly different initial conditions.

• Configuration: The AIGEFS consists of 31 distinct members.¹

• Generation Mechanism: Unlike traditional ensembles that require running the heavy physics code 31 times, AIGEFS generates these members using the efficient AI engine. The initial perturbations (the slight variations in starting temperature, wind, etc.) are likely derived from the existing GEFS perturbation scheme (Breeding Vectors or Ensemble Kalman Filter perturbations). The AI model then propagates these 31 distinct states forward in time.¹⁷

5.2 Performance Gains

The validation of AIGEFS has yielded remarkable results. NOAA states that the system extends forecast skill by 18 to 24 hours compared to the traditional GEFS.¹

• Interpretation: This means that a 7-day forecast from the AI ensemble has the same accuracy and reliability as a 6-day forecast from the physics-based ensemble. This gain is attributed to the AI’s ability to learn the dominant modes of atmospheric variability and filter out unpredictable noise more effectively than the discretized equations of the physics model.

6. System Architecture 3: The HGEFS (The Hybrid)

The most pioneering aspect of NOAA’s deployment is the Hybrid Global Ensemble Forecast System (HGEFS). This system represents the strategic acknowledgment that neither AI nor Physics is currently sufficient on its own.

6.1 The Rationale for Hybridization

Why combine them? AI and Physics models suffer from orthogonal error modes.

• AI Weaknesses: AI models tend to produce "smoother" fields with less variance. They minimize Root Mean Square Error (RMSE) effectively but can sometimes underestimate the intensity of extreme localized events (like the core pressure of a cyclone). They also lack inherent physical consistency.¹⁸

• Physics Weaknesses: Physics models maintain sharp gradients and conservation laws but are prone to "phase errors" (predicting a storm is 50 miles east of reality), which penalizes their RMSE scores heavily. They are also computationally expensive.²⁰

6.2 Construction of the Grand Ensemble

The HGEFS constructs a 62-member "Grand Ensemble" by pooling the resources:

• 31 Members from the physics-based GEFS.

• 31 Members from the AI-based AIGEFS.

• Weighting: While dynamic weighting schemes are possible (e.g., weighting recent performance), the initial operational deployment appears to use an equal weighting or a robust averaging that leverages the diversity of the two systems.¹

6.3 The Synergistic Effect

NOAA verification indicates that HGEFS consistently outperforms both the standalone AIGEFS and the standalone GEFS.¹

• Mechanism: The AI members stabilize the ensemble mean, reducing the overall error and bias. The physics members "fatten" the tails of the probability distribution, ensuring that extreme, physically plausible scenarios (that the AI might smooth out) remain in the forecast. This provides decision-makers with a "best of both worlds" product: the accuracy of AI for the general pattern and the reliability of physics for the extreme risks.

7. Comparative Verification and Performance Analysis

The decision to transition these models to operations was driven by a comprehensive validation campaign comparing AIGFS/AIGEFS against the established baselines (GFS/GEFS) and the global gold standard (ECMWF IFS).

7.1 Quantitative Metrics

The following table summarizes the comparative performance based on available verification data and proxy studies using GraphCast (the core of AIGFS):

Metric	AIGFS (AI-Driven)	GFS (Physics-Based)	ECMWF HRES	HGEFS (Hybrid)
Z500 ACC (Day 10)	~0.65 - 0.70	~0.55 - 0.60	~0.60 - 0.65	Highest (>0.70)
T2m RMSE (Global)	15-20% Lower	Baseline	Better than GFS	Lowest
Inference Time (10 days)	< 1 Minute	~60 Minutes	~60+ Minutes	~30 Minutes (Parallel)
Energy Consumption	~0.1 kWh	~100+ kWh	~100+ kWh	~50 kWh
Tropical Cyclone Track	Superior	Good	Excellent	Best Composite
Tropical Cyclone Intensity	Poor (Underestimates)	Good	Excellent	Improved over AI-only

Table 1: Comparative performance metrics derived from operational validation and proxy GraphCast studies.⁷

7.2 The 500hPa Geopotential Height (Z500) Benchmark

The Anomaly Correlation Coefficient (ACC) for Z500 is the standard "scorecard" for global models. A score of 0.6 is considered the limit of useful predictability.

• The AI Advantage: AI models like AIGFS maintain a score above 0.6 for approximately 11 days, whereas the GFS typically drops below this line around day 9 or 10. This effectively buys society an extra 24-48 hours of preparedness for large-scale weather pattern changes.⁷

7.3 Energy and Environmental Impact

In an era of conscious computing, the environmental footprint of forecasting is relevant. The physics-based ensemble requires megawatts of power to run its partial differential equation solvers. The AI component of the forecast requires a fraction of this—roughly 1,000 times less energy for the forecast step.²³ While the heavy lifting of Data Assimilation (creating the initial state) still requires the supercomputer, the forecast generation itself has become negligible in terms of carbon footprint.

8. Critical Limitations and the "Grey Swan" Risk

Despite the euphoric metrics, the transition to AI forecasting introduces novel risks that are fundamentally different from those of NWP. A growing body of academic literature warns that AI models, while excellent at the "mean," struggle with the "tails."

8.1 The Extrapolation Problem and "Grey Swans"

AI models are interpolative engines. They learn the probability distribution of the atmosphere from the training data (typically 1979–present). They excel at predicting weather that resembles the past. However, the climate is changing, generating "Grey Swan" events—extremes that are theoretically possible but historically unprecedented (e.g., the 2021 Pacific Northwest Heatwave).

• Empirical Evidence: Studies from the University of Geneva and the University of Chicago have demonstrated that AI models (including GraphCast) systematically underestimate the intensity of record-breaking heat and cold events.¹⁹

• The Mechanism: When a neural network encounters an input state (initial condition) that is far outside its training distribution (an outlier), it tends to regress toward the mean to minimize error. It predicts a "hot" day, but not the "shattering record" that physics models—which are unbound by historical statistics and solve for energy balance directly—are capable of simulating.²⁶

8.2 Spectral Bias and the "Blurry" Forecast

Another critical limitation is Spectral Bias. AI models are typically trained to minimize Mean Squared Error (MSE).

• The Double Penalty Problem: In meteorology, predicting a storm slightly in the wrong location incurs a "double penalty" (error where the storm is, and error where it isn't). The mathematical way to minimize MSE in an uncertain situation is to predict a smoothed average of all possibilities.

• Result: This leads to forecasts that lack fine-scale detail. The energy spectrum at high wavenumbers (small scales) drops off, resulting in "blurry" fields. This is problematic for predicting severe weather phenomena like convective storms, which depend on sharp gradients and high-energy localization.¹⁸

8.3 Case Study: The Failure of Hurricane Otis

The limitations of the current generation of models were starkly illustrated by Hurricane Otis, which devastated Acapulco in October 2023. Otis intensified from a tropical storm to a Category 5 hurricane in less than 24 hours—a rate of intensification that defied all historical climatology for that region.

• The Collective Failure: Both traditional NWP models and the leading AI models (GraphCast, Pangu, etc.) failed to predict this rapid intensification.²⁹

• Why AI Missed: The dynamics driving the intensification (inner-core thermodynamics) were likely sub-grid scale (smaller than the 25km resolution) and the rate of change was an outlier in the training data. The AI, seeing a tropical storm, predicted a standard intensification curve based on historical averages, missing the "black swan" physics that drove the event.³¹

• Operational Consequence: This failure underscores why NOAA developed the HGEFS. By keeping the physics members involved, there is a higher probability that at least some members might capture such outlier dynamics, even if the AI members smooth them out.¹

8.4 Physical Inconsistency and Hallucinations

Finally, AI models are not constrained by conservation laws. They can "hallucinate" changes in mass or energy that are physically impossible. While GraphCast is remarkably stable, longer forecasts can sometimes drift into physically inconsistent states (e.g., negative humidity or mass loss). This lack of "explainability" makes it difficult for forecasters to trust the model when it predicts something unusual—is it a genuine signal, or a neural network artifact?.³²

9. Downstream Impacts: Energy, Workflow, and Society

The deployment of these models reverberates far beyond the walls of NOAA.

9.1 The Renewable Energy Sector

The energy grid is increasingly dependent on weather. The integration of AIGFS/HGEFS offers significant value to energy traders and grid operators.

• Wind Power: Validation studies indicate that GraphCast-based models achieve accuracy comparable to or better than the ECMWF IFS for 10-meter wind speeds, a critical variable for turbine output.²²

• Grid Stability: The speed of AIGFS allows for "Nowcasting" and rapid scenario testing. Traders can run the model locally (using weights available via Anemoi or similar frameworks) to update wind generation forecasts minutes after a new satellite image is assimilated, rather than waiting for the 6-hour operational cycle. This reduces the risk of imbalances and pricing volatility in the spot market.²⁰

9.2 The "Cyborg" Forecaster and DESI

How does the human meteorologist interact with this new intelligence? NOAA has integrated the output of these models into the Dynamic Ensemble-based Scenarios for IDSS (DESI) visualization platform.¹

• The Workflow Shift: Forecasters are no longer "model watchers" waiting for the GFS to finish. They use the AIGFS/HGEFS as an early-warning tip sheet. The visualization tools in DESI allow them to interrogate the ensemble: "What is the probability of temperatures exceeding 100°F?" or "Show me the cluster of members predicting a coastal low."

• Trust and Adaptation: Surveys of NWS forecasters reveal a mix of excitement and caution. The "black box" nature of AI induces skepticism, but the sheer performance in day-to-day forecasting is winning converts. The consensus is shifting toward a "Human-on-the-Loop" system, where AI handles the routine synoptic flow, and humans focus on impact messaging for the extreme "Grey Swan" events that AI might miss.³⁶

10. Future Directions: The Foundation Model Era

NOAA’s current deployment is merely the first iteration (Version 1.0). The roadmap suggests a rapid evolution.

10.1 From Analysis-Driven to Observation-Driven

The current AIGFS is trained on Analysis data—meaning it still relies on the traditional NWP system to process raw observations into a grid. This is a bottleneck. The next frontier is End-to-End Deep Learning, where the model ingests raw satellite radiances and point observations directly, bypassing the traditional Data Assimilation system entirely. This would truly revolutionize the speed and cost of forecasting.³⁸

10.2 Regional Downscaling: HRRR-Cast

While AIGFS is global, weather impacts are local. NOAA is already experimenting with HRRR-Cast, an AI-based emulator of the High-Resolution Rapid Refresh (HRRR) model. This aims to bring the speed of AI to the convective scale (3km resolution), predicting individual thunderstorms and tornadoes.³⁹

10.3 Foundation Models

We are entering the era of Earth Foundation Models. Just as Large Language Models (LLMs) serve as a base for text, models like GraphCast, Microsoft’s Aurora, and NASA’s Prithvi are becoming foundational layers. NOAA’s engagement with the Anemoi framework positions it to leverage these "Foundation Models," fine-tuning them for specific tasks (e.g., hurricane track, severe hail, fire weather) rather than training new models from scratch.²⁸

Conclusion

The deployment of the AIGFS, AIGEFS, and HGEFS marks the end of the monopoly of the Navier-Stokes equations in operational meteorology. NOAA has pragmatically recognized that for the vast majority of atmospheric prediction, data-driven induction is now superior to physical simulation in both efficiency and accuracy. However, the retention of the physics-based ensemble within the HGEFS acknowledges the current limits of AI: its struggle with the unprecedented and the extreme.

We have entered the era of the Hybrid Forecast: a symbiotic relationship where Silicon Intelligence handles the routine complexity of global fluid dynamics, while Physics-Based Solvers and Human Experts stand guard over the chaotic, high-impact tails of the distribution. As these models continue to learn from an ever-warming atmosphere, this hybrid architecture will be humanity's primary instrument for navigating the volatile climate of the 21st century.

Citations

¹

Works cited

1. NOAA deploys new generation of AI-driven global weather models, accessed January 8, 2026, https://www.noaa.gov/news-release/noaa-deploys-new-generation-of-ai-driven-global-weather-models

2. NOAA Deploys New AI Driven Global Weather Models - Earth Prediction Innovation Center, accessed January 8, 2026, https://epic.noaa.gov/noaa-deploys-new-ai-driven-global-weather-models/

3. accessed January 8, 2026, https://docs.nvidia.com/physicsnemo/25.08/physicsnemo/examples/weather/graphcast/README.html#:~:text=GraphCast%20is%20a%20multi%2Dscale,of%20surface%20and%20atmospheric%20variables.

4. GraphCast for weather forecasting — NVIDIA PhysicsNeMo Framework, accessed January 8, 2026, https://docs.nvidia.com/physicsnemo/25.08/physicsnemo/examples/weather/graphcast/README.html

5. 1. Introduction — GraphCast GFS documentation - Read the Docs, accessed January 8, 2026, https://graphcastgfs.readthedocs.io/en/latest/introduction.html

6. Revolutionizing Weather Forecasting with Graph Neural Networks | by climatecast - Medium, accessed January 8, 2026, https://medium.com/stanford-cs224w/revolutionizing-weather-forecasting-with-graph-neural-networks-dcc2d06a4d52

7. GraphCast: AI model for faster and more accurate global weather forecasting, accessed January 8, 2026, https://deepmind.google/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/

8. EAGLE Overview - Earth Prediction Innovation Center - NOAA, accessed January 8, 2026, https://www.epic.noaa.gov/ai/eagle-overview/

9. An early look at NOAA's Project EAGLE to accelerate AI weather prediction advances for the United States, accessed January 8, 2026, https://epic.noaa.gov/noaa-project-eagle-to-accelerate-ai-weather-prediction-advances-for-the-united-states/

10. NOAA EAGLE: Research to Operations Pipeline for AI Weather Models at NOAA - AMS supported meetings, accessed January 8, 2026, https://ams.confex.com/ams/106ANNUAL/meetingapp.cgi/Paper/468721

11. ecmwf/anemoi - GitHub, accessed January 8, 2026, https://github.com/ecmwf/anemoi

12. Anemoi: a new framework for weather forecasting based on machine learning | ECMWF, accessed January 8, 2026, https://www.ecmwf.int/en/about/media-centre/news/2024/anemoi-new-framework-weather-forecasting-based-machine-learning

13. NOUS41 KWBC 091600 PNSWSH Service Change Notice 25-89 National Weather Service Headquarters Silver Spring MD 1100 AM EST Tue Dec, accessed January 8, 2026, https://www.weather.gov/media/notification/pdf_2025/scn25-89_AIGFS_AIGEFS_and_HGEFS.pdf

14. AI Weather Prediction (AIWP) Model Reforecasts - Registry of Open Data on AWS, accessed January 8, 2026, https://registry.opendata.aws/aiwp/

15. NOAA Launches First-Ever AI-Powered 'Ensemble' Forecast System - MeriTalk, accessed January 8, 2026, https://www.meritalk.com/articles/noaa-launches-first-ever-ai-powered-ensemble-forecast-system/

16. NOAA FourCastNetGFS API - AI Weather Forecast - GribStream, accessed January 8, 2026, https://gribstream.com/models/fourcastnetgfs

17. The AIGFS, AIGEFS, and Hybrid GEFS (HGEFS ... - NOAA VLab, accessed January 8, 2026, https://vlab.noaa.gov/documents/10157/138440/Abstract_VLabWebinar_Oct2025_AIGFSGEFS.pdf

18. Leveraging Data-Driven Weather Models for Improving Numerical Weather Prediction Skill through Large-Scale Spectral Nudging in - AMS Journals, accessed January 8, 2026, https://journals.ametsoc.org/view/journals/wefo/40/9/WAF-D-24-0139.1.xml

19. New study reveals unexpected results after using AI tools to predict extreme weather: 'Their ability … remains unclear', accessed January 8, 2026, https://www.thecooldown.com/green-business/ai-weather-forecasting-geneva-university-study/

20. The role of traditional and AI weather models in the future of renewable energy forecasting, accessed January 8, 2026, https://dexterenergy.ai/news/weather-models-in-renewable-energy-forecasting/

21. GraphCast Data is Essential - Enertel AI, accessed January 8, 2026, https://www.enertel.ai/post/graphcast-data-is-essential

22. WeatherMesh-2 Case Studies: Key Variables, Atlantic Hurricanes, and Winter Storms, accessed January 8, 2026, https://windbornesystems.com/blog/weathermesh-2-case-studies-key-variables-atlantic-hurricanes-and-winter-storms

23. ECMWF - delivering forecasts over 10 times faster and cutting energy usage by 1000, accessed January 8, 2026, https://www.eurekalert.org/news-releases/1088526

24. Comparison of AI and NWP Models in Operational Severe Weather Forecasting: A Study on Tropical Cyclone Predictions, accessed January 8, 2026, https://shixm.people.ust.hk/pdf/Shi2025JGRML.pdf

25. Numerical models outperform AI weather forecasts of record-breaking extremes - arXiv, accessed January 8, 2026, https://arxiv.org/html/2508.15724v1

26. AI is good at weather forecasting. Can it predict freak weather events?, accessed January 8, 2026, https://news.uchicago.edu/story/ai-good-weather-forecasting-can-it-predict-freak-weather-events

27. Validating Deep Learning Weather Forecast Models on Recent High-Impact Extreme Events in - AMS Journals, accessed January 8, 2026, https://journals.ametsoc.org/view/journals/aies/4/1/AIES-D-24-0033.1.xml?utm_source=chatgpt.com

28. An Observations-focused assessment of Global AI Weather Prediction Models During the South Asian Monsoon - arXiv, accessed January 8, 2026, https://arxiv.org/html/2509.01879v1

29. Why did numerical weather forecasting systems fail to predict the Hurricane Otis's development? - UNAM, accessed January 8, 2026, https://www.revistascca.unam.mx/atm/index.php/atm/article/view/53367

30. Hurricane Otis a "monumental" weather model miss - Artemis.bm, accessed January 8, 2026, https://www.artemis.bm/news/hurricane-otis-a-monumental-weather-model-miss/

31. Otis becomes Mexico's strongest landfalling pacific hurricane on record - Moody's, accessed January 8, 2026, https://www.moodys.com/web/en/us/insights/insurance/otis-becomes-mexicos-strongest-landfalling-pacific-hurricane-on-record.html

32. Can AI weather models predict out-of-distribution gray swan tropical cyclones? - PNAS, accessed January 8, 2026, https://www.pnas.org/doi/10.1073/pnas.2420914122

33. Special Series on AI – Part 1: AI Models De-Mystified | Aeolus Capital Management, accessed January 8, 2026, https://aeolus.com/special-series-on-ai-part-1-ai-models-de-mystified/

34. GraphCast vs. IFS: A Comparative Analysis of Weather Forecasting Models for Optimized Renewable Energy Trading, accessed January 8, 2026, https://amslaurea.unibo.it/id/eprint/35241/1/Lombardini_Sveva_Thesis.pdf

35. EAGLE Graphics - Earth Prediction Innovation Center - NOAA, accessed January 8, 2026, https://epic.noaa.gov/ai/eagle-graphics/

36. How AI is Shaping Weather Research and Forecasting: An Interview with Amy McGovern, accessed January 8, 2026, https://www.nationalacademies.org/news/how-ai-is-shaping-weather-research-and-forecasting-an-interview-with-amy-mcgovern

37. National Weather Service (NWS) Forecasters' Perceptions of AI/ML and Its Use in Operational Forecasting - the NOAA Institutional Repository, accessed January 8, 2026, https://repository.library.noaa.gov/view/noaa/69271/noaa_69271_DS1.pdf

38. Learning AI for the Hurricane Forecast Improvement Program - Physical Sciences Laboratory, accessed January 8, 2026, https://psl.noaa.gov/events/2023/ai4nwp_workshop/pdf/20231129-Garrett-AI4NWP-HFIP.pdf

39. NOAA Research develops an AI-powered sibling to its flagship weather model, accessed January 8, 2026, https://research.noaa.gov/noaa-research-develops-an-ai-powered-sibling-to-its-flagship-weather-model/

40. Root mean squared error (RMSE) for (a) Z500, (b) T850, (c) T2M, and (d)... - ResearchGate, accessed January 8, 2026, https://www.researchgate.net/figure/Root-mean-squared-error-RMSE-for-a-Z500-b-T850-c-T2M-and-d-PR-evaluated_fig1_349032500

41. AIGFS/AIGEFS/HGEFS- VLAB seminar - NOAA VLab, accessed January 8, 2026, https://vlab.noaa.gov/documents/10157/137122/VLab+Forum+Oct.+2025+--+The+AIGFS%2C+AIGEFS%2C+and+Hybrid+GEFS+%28HGEFS%29.pdf

42. Introducing Anemoi: a new collaborative framework for ML weather forecasting - ECMWF, accessed January 8, 2026, https://www.ecmwf.int/en/newsletter/181/news/introducing-anemoi-new-collaborative-framework-ml-weather-forecasting