The U.S. National Science Foundation National Center for Atmospheric Research (NSF NCAR) has released the next generation of its CREDIT platform, an open-source tool designed to let researchers build AI models that emulate traditional Earth system simulations at a fraction of the computational cost. The update is aimed at scientists and graduate students who want to work with AI weather models without first becoming machine learning experts.
CREDIT, which stands for Community Research Earth Digital Intelligence Twin, uses machine learning to approximate the behavior of physics-based models. These emulators are trained on the datasets produced by complex simulations, then can mimic the results in hours or days on a laptop instead of weeks or months on a supercomputer.
"Our target user profile is a graduate student working with an Earth system science professor of some sort," said NSF NCAR scientist David John Gagne, one of CREDIT's primary developers. "The professor may not know anything about AI, but they have cool data and the graduate student wants to build an AI model to work with that data. We want the graduate student to jump through the minimum number of hoops to get that model up and running, and they shouldn't have to be an expert in machine learning."
From scratch to building blocks
The NSF NCAR team spent years building custom AI emulators for researchers, coding each one from scratch with custom data pipelines. The first version of CREDIT, released in December 2024, served as a proof of concept. The updated platform rebuilds the data pipeline to reduce delays and adds cloud access to training datasets.
The original approach required scientists to duplicate work across projects. Gagne and his colleagues built CREDIT to make the process modular, so researchers can assemble components instead of starting over with every model.
The new version includes better user documentation and a more modular structure, which lowers the barrier for researchers who have not worked with machine learning before. "We're really excited to see how far we've come and to let people know this resource exists," Gagne said. "We plan to keep working on improving both the models and the support structures that the community will need to truly take advantage of these new advances."
Physics guardrails keep AI results grounded
AI models in their simplest form reproduce patterns from their training data without regard for the laws of physics. The CREDIT team has added "physics checks" that evaluate emulator output and push results back toward physically consistent solutions. Adjustments can be propagated backwards through the model, keeping results realistic.
The new platform has already been used to build an emulator of the Community Atmosphere Model (CAM), which scientists are using to study predictability at subseasonal timescales - two weeks to two months. The team is also developing emulators for an ocean model and for parts of the Community Earth System Model.
The goal is to cover the full range of research needs, from high-resolution models that run for short periods to low-resolution models that simulate centuries. Researchers can access CREDIT as an open-source package and download it directly.
Why this matters for research scientists
For scientists without supercomputing access, AI emulators from CREDIT can make physics-based simulations practical for teaching and exploratory research. The tools spread the value of expensive, equation-based models to a much wider audience - including students who need to test ideas quickly. Researchers interested in building AI models for their own work can start with CREDIT's building blocks instead of writing machine learning code from the ground up. For those looking for foundational skills before diving into modeling, AI for Research Scientists covers machine learning basics for scientific applications, while resources across AI for Science & Research cover practical use cases for labs and field work.
Your membership also unlocks:
