An updated tool developed by machine learning scientists and engineers at the U.S. National Science Foundation National Center for Atmospheric Research (NSF NCAR) aims to make it much easier for researchers to build artificial intelligence (AI) models that can emulate traditional Earth system models at a fraction of the computational cost — even if those researchers don’t have AI expertise.

The next generation of NSF NCAR’s Community Research Earth Digital Intelligence Twin (CREDIT) improves on the original platform in significant ways: it streamlines the user experience with a more modular, building-blocks approach that includes better user documentation; it reduces time delays in running the system; and it introduces more physics to keep the AI model outputs as realistic as possible.

AI is accelerating solutions to longstanding scientific challenges that have been very difficult to make progress on but have important implications for society, including more accurate forecasts for seasonal droughts, hyperlocal winds, hurricane intensity, urban flooding, and many more. CREDIT is a key part of NSF NCAR’s effort to build and provide powerful AI capabilities to support the research community in this work.

The team’s ultimate goal is to make a platform that will help researchers and students focus their time on answering scientific questions rather than learning how to write the machine learning code needed to build, train, and test AI models. 

“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.”

Eliminating the need to build from scratch

Gagne and his team have already spent years working with Earth system scientists across NSF NCAR to build custom AI model emulators to help them with their research. NSF NCAR is home to world-class, physics-based models that can simulate phenomena across the Earth system, from global atmospheric, ocean, and land models to models of individual tornadoes, hurricanes, and wildfires. Those models are built on deep understandings of how these systems work, and they solve mathematical equations to represent the physical processes that unfold. But the mathematics used in physics-based models often require the use of a supercomputer to run, and it can take weeks or months to produce the simulations needed by researchers.

AI emulators work by approximating the behaviors of the more complex physics-based models. They are trained on the vast datasets produced by the physics-based model, and use machine learning techniques to mimic the results. While emulators, by definition, only approach the accuracy of the original, they are suitable for helping answer a wide range of scientific questions that could benefit society — and they can do so quickly and with much fewer computational demands.

A set of simulations that might take weeks or months on a supercomputer could be done in a matter of hours or days using an AI emulator on a laptop. Emulators do not negate the need for physics-based models. In fact, improvements in those models will also improve emulators, but they do spread the value of physics-based models to a much larger audience that includes students and others who may not have access to supercomputers.

When they first started building AI emulators, they found themselves having to code each one from scratch, building custom pipelines for the data used to train the emulators. So they turned their attention to finding a way to build new emulators without reinventing the wheel every time. 

The result was the first generation of CREDIT, which was released in December 2024 and was essentially proof of concept to show that it could be done. Satisfied that it could, NSF NCAR machine learning scientists began working on upgrades that move CREDIT toward a platform that is truly accessible and easy to use, while adding improvements to its performance along the way. 

The next generation CREDIT platform has already been used to build an emulator of the NSF NCAR-based Community Atmosphere Model (CAM). Scientists are interested in using the CAMulator, as it’s known, to explore predictability at subseasonal timescales (two weeks to two months). Now they’re working on a number of other emulators, including one that can mimic an ocean model and one that can mimic aspects of the Community Earth System Model, but this is just the start. The goal is to be able to use CREDIT to build a wide range of emulators, from models that run at high resolution for very short time scales to those that run at low resolution to simulate centuries and everything in between. 

The new update also addresses some latencies related to how the first version of CREDIT pulled down and processed data. They have now rebuilt the entire data pipeline to make it more efficient, and CREDIT can now access datasets in the cloud.

And finally, the team has added more physics to CREDIT. While AI models in their simplest form reproduce patterns based on the data they are trained on — without regard for the laws of physics — NSF NCAR scientists have found ways to introduce guardrails. CREDIT has a number of “physics checks” that can be performed on the output of the emulators, and adjustments can be propagated backwards through the model, nudging results back towards physically consistent solutions.

CREDIT’s development is ongoing, but the release of the next generation of the platform marks a significant milestone toward empowering AI use across the Earth system science community. 

“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.”

CREDIT is open source, and anyone can download and run the package. Visit the CREDIT website to learn more. 

See all News