CREDIT

Community AI Modeling: Community Research Earth Digital Intelligence Twin (CREDIT)

NSF NCAR fully embraces community-driven open-source development—a collaborative model in which software is created, maintained, and advanced by a broad community united by a shared scientific vision. Flagship community codes such as the Community Earth System Model (CESM), the Community Atmosphere Model (CAM), and the Model for Prediction Across Scales (MPAS) are developed through open repositories, transparent governance, and community review.

We are extending that same approach to AI/ML by treating data-driven models, training frameworks, and deployment services as community scientific infrastructure: transparent in architecture, training data, evaluation, and version history; independently verifiable; and sustainably maintained.

Learn more about NSF NCAR’s open development practices and how you can get involved with these packages:

• CREDIT
• CESM-MLe

Our commitment to responsible and ethical AI requires transparency in model architectures, training datasets, preprocessing pipelines, evaluation metrics, uncertainty characterization, and version histories. Open development enables independent validation, facilitates benchmarking and intercomparison, and supports reproducibility in a domain where model behavior can be sensitive to training data and implementation choices. By treating AI/ML models as community scientific infrastructure—rather than isolated research artifacts—we aim to ensure they are trustworthy, well-documented, rigorously evaluated, and sustainably maintained for the benefit of the broader scientific community.

Community Research Earth Digital Intelligence Twin (CREDIT): A community framework for training and deploying AI-based Earth System Science models

What is CREDIT?

CREDIT is an open, Python-based platform for training and evaluating AI Earth system models, also known as emulators (Schreck et al. 2025). Emulators learn from physics-based model outputs. They run in hours or days, often on far smaller hardware, where comparable physics-based simulations can take weeks or months on a supercomputer. Emulators don’t replace physics-based models. Instead, they extend those models’ value to researchers and students without supercomputer access.

CREDIT enables users to:

• Load large datasets,
• Train customizable AI models at scale,
• Incorporate physical constraints and transforms, and
• Deploy models for forecasting, hindcasting, or projections.

CREDIT’s full pipeline is open source and being developed openly with and for the community.

Who it’s for

CREDIT is built so that researchers can focus on science questions rather than machine learning code. Our target user is a graduate student with compelling Earth system data and an advisor who may not be an AI expert. That student should be able to build, train, and test a model without becoming a machine learning specialist first.

Links and Resources

• Learn more about and start using CREDIT.
• MILES project page
• GitHub repository

Underpins models:

• WXFormer
• CAMulator: emulates CAM6 for fast climate-scale simulation, including subseasonal predictability research.
  • Paper: https://arxiv.org/abs/2504.06007
  • Model weights: https://huggingface.co/willychap/camulator
• NCAR-FuXi
• subCESMulator: emulates CESM with interactive land and ocean, aimed at subseasonal prediction.

Papers:

• Schreck et al. 2025, npj Climate and Atmospheric Science: the flagship CREDIT paper. https://doi.org/10.1038/s41612-025-01125-6
• Sha et al. 2025, on mass and energy conservation schemes. https://arxiv.org/abs/2501.05648
• Sha et al. 2025, on terrain-following coordinates in precipitation forecasts. https://arxiv.org/abs/2503.00332
• Chapman et al. 2025, CAMulator. https://arxiv.org/abs/2504.06007

News:

• Chasing Subseasonal Forecasts with AI
• Scientists build up CREDIT for AI research
  

2025 CESM Workshop slides:

• DJ Gagne
• Will Chapman

For more information or for partnership opportunities, please contact John Clyne.