Key NOAA Presentations on Artificial Intelligence
NOAA offered several Artificial Intelligence (AI) focused presentations, which highlighted a growing ecosystem designed to support the development, verification, and operational transition of machine learning–based weather prediction models. Together, the key presentations below demonstrate frameworks such as Anemoi, wxvx, and NOAA’s Project Experimental AI Global and Limited-area Ensemble forecast system (EAGLE). The presentations highlight scalable infrastructure, flexible verification tools, and research-to-operations pipelines enabling the community to build, test, and deploy next-generation forecasting systems with greater efficiency and reliability.
The NOAA Anemoi Experience: Scalable and User-Friendly Tools for Training AI Weather Prediction Models
Contributors: Mariah Pope(1), Timothy A. Smith(2), Sergey Frolov(2), Daniel Abdi(3), Isidora Jankov(3), D. Alex Burrows(1), Anil Kumar(1), Paul Madden(3)
1 Earth Prediction Innovation Center (EPIC), 2 Physical Sciences Laboratory (PSL), 3 Global Systems Laboratory (GSL)
This session introduced a community framework combining:
ufs2arco: NOAA Physical Sciences Laboratory (PSL)
Anemoi: European Centre for Medium-Range Weather Forecasting ECMWF’s open-source graph-based AI framework
The rapid growth of data-driven weather prediction models has produced systems that rival traditional numerical weather prediction (NWP) skill at a fraction of the computational cost. However, retraining these models using NOAA-specific datasets can be cumbersome due to hardware setup, data staging, and scaling challenges.
This framework addresses those barriers by:
Enabling streamlined cloud access to NOAA datasets
Providing customizable graph-based ML model development
Supporting cloud and on-prem execution
Guiding users from environment setup to inference
Benchmarking computational efficiency and cost
Together, ufs2arco and Anemoi provide a scalable pathway for community-driven AI weather model development and experimentation.
Supporting AI/ML Weather Prediction in NOAA: A Flexible Verification Pipeline with wxvx
Contributors: D. Alex Burrows (1, 2), Anil Kumar (1, 3), Mariah Pope (1, 3), Tim Smith (4), Sergey Frolov (4), Paul Madden (5), Isidora Jankov (5), Jacob Carley (6), Jun Wang (6)
1 EPIC, 2 RedLine, 3 Tomrrow.io, 4 PSL, 5 GSL, and 6 Environmental Modeling Center (EMC)
As AI forecast systems mature, rigorous verification becomes essential.
This presentation introduced wxvx, a flexible verification framework designed to support both physics-based and AI-based forecast systems.
wxvx offers:
YAML-driven configuration
Portability across NOAA, non-NOAA, and cloud systems
Pre-built MET/METplus executables via met2go
Adjustable forecast cycle and length configurations
Support for global and regional grid-to-grid verification
Active development includes:
Grid-to-observation verification
Hosting verification statistics via the EPIC web portal
Full integration with Anemoi forecast outputs
wxvx ensures AI models meet operational verification standards before transition.
Ocean and Coupled Earth System Emulators: Challenges, Solutions and Opportunities
Contributors: Niraj Agarwal (3), Tim Smith (1), Sergey Frolov (1), Laura Slivinski (1)
1 NOAA/OAR/Physical Sciences Laboratory (PSL), 2 National Oceanic and Atmospheric Administration (NOAA), 3 Cooperative Institute for Research in Environmental Sciences (CIRES)
Researchers have developed a flexible machine learning framework based on Google DeepMind’s GraphCast architecture to emulate components of the Earth system for medium-range weather forecasting. The framework allows models to represent either single components (such as the atmosphere) or multiple coupled components (such as atmosphere and ocean) by modifying the propagating state vector, enabling stronger coupling and potential support for strongly coupled data assimilation. Using NOAA’s UFS-Replay dataset, the team has successfully trained an atmosphere-only emulator and is now developing atmosphere–surface ocean and ocean-only models. Early results show that coupling atmospheric and ocean components meaningfully influences atmospheric predictions, while ocean variable errors remain comparable to other ocean emulators. Ongoing work focuses on overcoming challenges in building a 3D ocean emulator, including handling slow ocean dynamics, vertical correlations, and model design considerations.
Just Add DA
Contributors: Sergey Frolov (1), Laura Slivinski (1), Tim Smith (1), Chong-Chi Tong (3), Matt Bender (4), Kelsey Lieberman (4), Josh DaRosa (4), Nicholas Silverman (4), Chris Miller (4), Mohammad Alam (4), Alex Philp (4), Noah D. Brenowitz (5)
1 NOAA/OAR/Physical Sciences Laboratory (PSL), 2 NOAA, 3 Cooperative Institute for Research in Environmental Sciences (CIRES), 4 The MITRE Corporation, 5 NVIDIA Research
Machine learning (ML) is increasingly being explored for weather prediction, but current ML approaches are not yet capable of fully replacing traditional operational data assimilation (DA) systems. This work introduces add-DA, a framework that integrates machine learning with existing DA systems to improve forecast initialization. Instead of replacing DA, the approach uses ML to better leverage observational data, particularly observations that are currently underused—to correct operational initial conditions. Using NOAA’s UFS-Replay dataset, the prototype system learns key components including the observation operator, forecast model, and a DA operator that converts observation innovations into improved analysis states. Early results demonstrate how combining these learned components can enhance forecast initialization and overall prediction performance.
Nested-EAGLE: A Data Driven, Global Weather Model with High Resolution over the Contiguous US
Contributors: Tim Smith (1), Mariah Pope (2), Sergey Frolov (1), Brett M. Basarab (1,3), Paul Madden (3,4), Daniel Abdi (3,4), Isidora Jankov (4)
1 National Oceanic and Atmospheric Administration Physical Sciences Laboratory (NOAA-PSL), 2 EPIC, 3 Cooperative Institute for Research in Environmental Sciences (CIRES), 4 National Oceanic and Atmospheric Administration Global Systems Laboratory (GSL)
NOAA’s PSL is working to improve forecasts of precipitation extremes, particularly Atmospheric Rivers (ARs) that impact the western United States. Because ARs span thousands of kilometers but interact with complex local terrain after landfall, they present a challenging multiscale forecasting problem. To address this, researchers developed a graph-based neural network weather model trained on archived GFS and HRRR datasets. The model provides global forecasts at ~28 km resolution with refinements to under 10 km across the continental United States. Initial evaluations assess the model’s ability to predict precipitation across short- to medium-range forecasts and during AR events, comparing performance with operational GFS and HRRR models using the Fractions Skill Score and the Analysis of Record for Calibration (AORC) dataset, while highlighting progress and remaining challenges in high-resolution data-driven forecasting.
ML Weather Prediction Development for EAGLE Global Forecast
Contributors: Jun Wang (1), L. Cui (2,1), J. Liu (2,1), X. Luo (2,1), B. Cui (2,1), R. Manser (3), C. Diaz (3), M. Curtis (3), S. Earle (3), J. Cooke (3), W. Li (2,1), B. Fu (1), Z. Zhang (1), L. Zhu (2,1), J. Peng (2,1), B. Liu (2,1), S. Frolov (4), T. Smith (4), K. Blackman (5), M. Pope (5), J. Carley (1)
1 Environmental Modeling Center (EMC), 2 Lynker, 3 NCEP Central Operations, 4 OAR Physical Sciences Laboratory, 5 OAR Earth Prediction Innovation Center
As machine learning (ML) weather models rapidly mature, NOAA is developing an operational pathway to integrate AI-driven global forecasts into its production ecosystem.
This presentation outlines the ML development efforts supporting NOAA’s Project EAGLE GFS.
The work focuses on:
Adapting state-of-the-art ML architectures for NOAA operational requirements
Training and fine-tuning AI models using GDAS-derived atmospheric analyses
Producing near-real-time deterministic and ensemble global forecasts
Ensuring compatibility with NOAA verification standards and infrastructure
Building scalable workflows suitable for HPC and cloud environments
Key development components include:
Data preprocessing pipelines aligned with operational formats
Model optimization for forecast skill, stability, and computational efficiency
Integration pathways toward EMC production systems
Evaluation against baseline physics-based models
This effort represents a structured research-to-operations bridge, accelerating AI experimentation while maintaining NOAA’s reliability, transparency, and verification rigor. The EAGLE framework enables rapid testing, benchmarking, and transition of ML-based global forecast systems into NOAA’s operational suite.
NOAA EAGLE: Research-to-Operations Pipeline for AI Weather Models at NOAA
Contributors: Sergey Frolov (1), Jun Wang (2), Isidora Jankov (3), Jacob Carley (1), Keven Blackman (4), Daryl Kleist (2), Travis Wilson (3), Linlin Cui (2), John Ten Hoeve (5), Maoyi Huang (5)
1 PSL, 2 EMC, 3 GSL, 4 EPIC, 5 Weather Program Office (WPO)
NOAA’s weather forecasting enterprise is building a dedicated pipeline — Project EAGLE forecast system) — to accelerate the evaluation and operational transition of AI forecast models.
Key points from the presentation:
Purpose of EAGLE: Provide an open, near-real-time environment where AI forecasting models can be tested, compared, and verified against NOAA’s trusted operational metrics and systems (like Global Forecast System (GFS), Global Ensemble Forecast System (GEFS)).
Current capability: The early version produces daily deterministic and ensemble forecasts using a NOAA-tuned version of GraphCast (from Google DeepMind), fine-trained on inputs from NOAA’s Global Data Assimilation System (GDAS).
Research-to-Operations focus: EAGLE is structured as a pipeline—research → demonstration → operations—to help identify promising AI innovations, support near-real-time experimentation, and expedite adoption of successful models into future operational workflows.
Community leverage: The system is expected to build on common AI infrastructure (e.g., the open-source Anemoi framework) so that NOAA, academic, and enterprise partners can contribute and benefit from shared tools and best practices.
Long-term outlook: By integrating AI models into a structured evaluation environment, EAGLE aims to shorten development cycles and improve forecast accuracy while aligning new techniques with operational expectations.
Warn-on-Forecast System Cast (WoFSCast): Results from the 2025 HWT Spring Forecasting Experiment and improvements from a probabilistic loss function
Contributors: Corey K. Potvin (1), Montgomery L. Flora (1,2), Joshua Martin (2), Adam J. Clark (1), Miranda K. Silcott (2), Michael E. Baldwin (2,3), Patrick C. Burke (1), Tim A. Supinie (3)
1 NOAA National Severe Storms Laboratory (NSSL), 2 Cooperative Institute for Severe and High-Impact Weather Research and Operations (CIWRO), 3 NOAA/NWS Storm Prediction Center (SPC)
Traditional AI weather prediction models trained with standard loss functions often produce forecasts that are overly smooth and ensembles that lack sufficient spread. To address this, researchers adapted a probabilistic training approach, similar to that used in ECMWF’s AI Forecasting System, by implementing a modified Continuous Ranked Probability Score (CRPS) loss and noise injection during training. Applied to the WoFSCast 3-km AI model, this new WoFSCast-CRPS framework reduces forecast smoothing and improves ensemble calibration, particularly for storm-related variables like composite reflectivity. Initial evaluations, including feedback from participants in the 2025 Hazardous Weather Testbed Spring Forecasting Experiment, indicate that the updated model produces sharper, better-calibrated forecasts and represents a promising direction for future WoFSCast development.
HRRRCast: a data-driven model for regional weather forecasting at convection allowing scales
Contributors: Isidora Jankov (1) and Daniel Shawol Abdi (2): A Cross NOAA Collaboration through AI4NWP Tiger Team (GSL, EMC, PSL, NSSL, WPO, OSTI)
1 GSL, 2 CIRES
HRRR-Cast is an experimental AI-based regional forecasting system developed by NOAA’s Global Systems Laboratory to complement the operational High-Resolution Rapid Refresh (HRRR) model. Trained on three years of HRRR analysis data, the model uses a ResNet-based architecture (ResHRRR) with probabilistic forecasting capabilities and is being evaluated by the National Weather Service as part of Project EAGLE, NOAA’s initiative to accelerate AI weather prediction research. HRRR-Cast builds on earlier StormCast work with improvements including full CONUS coverage, multi-lead-time training, direct use of analysis data, incorporation of future GFS forecasts, and enhanced downscaling for higher-resolution detail. Recent updates further expand vertical resolution, add additional meteorological variables, and improve ensemble performance.
