Artificial intelligence is driving a seismic shift in how we approach weather forecasting, with a flurry of new AI weather prediction models debuting in recent years that have a number of potential advantages over traditional models. These include faster speeds, reduced demand for computing resources, and improved forecast performance for some weather phenomena, especially over longer time periods.

Despite the excitement around these new models, they currently have some significant limitations, and they are also not easy for the broader research community to adapt and use to make further scientific progress. Now, the U.S. National Science Foundation National Center for Atmospheric Research (NSF NCAR) has released a new platform that aims to make AI weather models more accessible. The goals are to make it easier to use AI to address science questions at the frontiers of weather research and to allow more researchers to contribute back to expanding the capabilities of AI weather modeling.

"When our team started to build our first AI weather model, we found little publicly available guidance on how to train these models at full scale. We had to learn the hard way through trial and error and many discussions with other experts,” said David John Gagne, who leads NSF NCAR’s machine learning science efforts. “Our goal is to provide a platform that allows our research community to easily train models on particular datasets and then configure them to run on our available computing resources, all without having to be an expert in artificial intelligence or supercomputing. We want to lower the barriers to entry and get this potentially transformative technology into the hands of many more interested researchers.”

The new platform is called the Community Research Earth Digital Intelligence Twin (CREDIT), and it has three important parts: a library of AI models, multiple pre-prepared, high-quality datasets to train those models, and access to high-performance computing. Ultimately, it allows researchers to choose between different model architectures and databases and then helps them train the model without having extensive knowledge of high-performance computing.

First steps toward a new kind of community model

NSF NCAR has a rich legacy of providing top-flight, open-source weather models to the community and providing training and support. Ultimately, the more people use these community models, the better they get because users contribute their own improvements and upgrades to the model code.

To date, these community models have been traditional computer models, meaning the computer solves equations that represent physical processes in the atmosphere — from the small-scale cycle of updrafts and downdrafts that create a thunderstorm to the large-scale movement of air masses that collide, creating warm fronts and cold fronts. These models are built on an understanding of atmospheric physics, and they are capable of providing extremely realistic simulations of how weather unfolds on Earth.

But these models have challenges. For one, increasing model resolution to study smaller-scale, localized weather phenomena requires massive increases in computational resources. For example, doubling a model’s resolution (both horizontally and vertically) requires a 16-fold increase in computing power. The intense computational demand of these models also limits the amount of times a model can be run. Running multiple simulations over the same time period — known as an ensemble — can give forecasters crucial information about the amount of uncertainty in the forecast, but it also requires vast amounts of computing power.

Traditional models also struggle in areas where our scientific understanding of how a phenomenon forms or behaves is incomplete and therefore cannot be as readily expressed using mathematical equations. These tend to be the weather events that are stubbornly hard to accurately predict, such as the rapid intensification of hurricanes or the growth of damaging hail.

AI weather models have the potential to address these shortcomings. Unlike traditional models, AI models do not solve individual equations. Instead they look for patterns and associations in the data that give insight into what’s likely to happen next. This approach significantly reduces computing demand and also allows the models to find ways to predict phenomena we can’t easily write equations for.

This potential and the rapid advancement of AI in general has led to a profusion of new models developed by tech companies, nonprofits, government agencies, and universities. But these models have their own drawbacks, including a tendency for small errors to grow into large errors quickly. And AI weather models are only as good as the data used to train them. Different questions may require using different training data, which itself can be difficult to source and prepare for use by AI models.

CREDIT is NSF NCAR’s first foray into building a community AI modelling platform to help facilitate progress in Earth system science research. There is still work to be done before CREDIT is a robust community modeling resource. In particular, the success of NSF NCAR’s community models relies heavily on the thorough documentation, training, and support infrastructure NSF NCAR provides, and this ecosystem of user support is not yet built out for CREDIT. Still, the CREDIT software, which is publicly available to download, is a significant step toward putting the power of AI in the hands of more Earth system researchers.

“We want scientists to be able to focus on their research goals, rather than the technical details,” Gagne said. “This framework makes AI accessible to a broad range of users, from experienced atmospheric researchers to students just beginning their journeys.”

Competitive with traditional weather models

In the process of building and testing CREDIT, the NSF NCAR team also built WXFormer, a new AI weather model specifically designed to be used for weather research questions and to address some of the shortfalls of existing AI weather modeling, including new strategies for limiting error growth and allowing for shorter time steps. Most existing AI weather models provide information in six-hour intervals, but WXFormer can provide forecasts hourly. WXFormer is now one of the models researchers can choose from within CREDIT’s library, along with others that have their codes published publicly.

Using CREDIT to train and set up the models, the NSF NCAR team tested the performance of WXFormer and FuXi, another AI model available in CREDIT’s library, against forecasts from the High Resolution Integrated Forecast System (HRES-IFS), a standard weather model developed by the European Centre for Medium-Range Weather Forecasts that is widely considered to be one of the best performing weather models in the world.

Both WXFormer and FuXi demonstrated they could provide the same or better forecasts than HRES-IFS at extended lead times for most atmospheric variables. For example, scientists tested how well the models predicted the track and intensity of Hurricane Laura, which struck western Louisiana as a category 4 storm in 2020. They found that WXFormer did the best job of predicting the hurricane’s intensity at 5 days out, though its predicted track was too far south. A version of WXFormer that produces hourly forecasts captured the hurricane’s track more accurately, but predicted a much weaker storm at five days out. The FuXi model predicted a weak storm at five days out and had similar track errors to WXFormer. The HRES-IFS model also predicted a very weak storm at five days out and had some track errors in the opposite direction of the AI models.

In all, the results underscore the potential of CREDIT and point to directions for future work. Already, the team has produced an updated release of CREDIT software that increases the platform’s user-friendliness and scalability. They are also working on another AI model that can emulate the NSF NCAR based Community Atmosphere Model (CAM), which is a component of the larger Community Earth System Model, which connects models of ocean, atmosphere, land, and ice together. The “CAMulator” would make it easier for AI modeling to be integrated into global Earth system simulations.  

CREDIT’s development is also being underpinned by a parallel effort at NSF NCAR to build an “integrated data commons” that would make the organization's vast datasets more accessible for training AI models. Having high-quality data is the bedrock of AI and the new data commons will serve as a foundation for future AI model development.

“We’re excited to see what problems CREDIT can help solve and where this can go with the help and participation of the broader community,” Gagne said. “Our aim is for CREDIT to be an open and collaborative environment where researchers at various levels of expertise can contribute to the framework’s evolution.”

See all News