Ask the Experts: what does the future of AI weather forecasting look like?
By Theresa Barosh | Feb. 2025 | Originally published in SOURCE
Colorado State University researchers are working collaboratively to evaluate emerging AI weather forecast models and provide resources to help other researchers investigate their usefulness.
“I’m definitely enthusiastic about AI [weather forecast] models and think they’re the future,” said Jacob Radford, researcher at CSU’s Cooperative Institute for Research in the Atmosphere. “As scientists, we must have some healthy skepticism. That’s ultimately how we’ll keep improving these models: identifying the weaknesses and then we’ll fix them.”
Multiple artificial intelligence weather forecasting tools have emerged from the private sector in the last few years, including Google’s latest GenCast released near the end of 2024.
Radford and a team of NOAA and CSU researchers recently published a research article in the Bulletin of the American Meteorological Society on two new resources meant to facilitate access to AI-based weather prediction model output data. The publication team included researchers from NOAA’s Global Systems Laboratory and CIRA, NOAA’s collaboration with CSU’s Walter Scott, Jr. College of Engineering.
CIRA’s real-time visualizations for purely AI-based weather models, developed by Radford, can be found online.
Radford said they hope providing archives of AI weather forecast model output will help the broader atmospheric science community to investigate AI weather forecast models. Radford pointed out that not all researchers have access to the computing power that he and his team have at CIRA. Radford and his team can save other researchers from expending energy and computing resources to run the models; they can skip straight to analysis.
Student Research
Researchers from CSU’s Atmospheric Science department have already started using the CIRA team’s archive. With CIRA funding and support from CIRA researchers Ryan Lagerquist and Imme Ebert-Uphoff, recent graduate Allie Mazurek used the output for part of her doctoral graduate research advised by Atmospheric Science Professor Russ Schumacher.
“It was really a mutually beneficial thing where we have resources that we can provide the students, and the students are doing really quality research,” said Radford, describing collaborations between CIRA and Atmospheric Science, “And investigating these things that we haven’t been able to get to or don’t have experience in.”
Radford said he was excited that Mazurek put her expertise toward the project.
Mazurek explored parameters, or values, that are relevant to predicting severe convective weather, such as thunderstorms.
“We’re not predicting those parameters directly from the models. We’re just taking the model output and then making some calculations to compute these parameters. Ryan was super helpful – he was the one responsible for making all those calculations,” said Mazurek. “My role has been analyzing those data and seeing what they look like: do they make sense physically and how do they compare to our current numerical weather prediction models?”
Numerical weather prediction models represent conventional weather-prediction methods and use physical equations making predictions on the basis of an understanding of physics. In contrast, generative AI models forecast weather patterns that are plausible given past measurements, or statistical likelihoods. Specifically, the models use machine learning, one method of training an AI by using past data.
Mazurek found that the moisture in the AI models tended to be too low throughout most of the atmosphere, while the temperature tended to be slightly too cool in the mid-levels of the atmosphere. Warm, moist air must rise and run into cool air to form thunderstorms as far as real-world physics are concerned.
Limitations and Future Prospects
Both Mazurek and Radford said it isn’t time to retire traditional forecasting methods. For one thing, AI weather forecast models are often trained on datasets from traditional forecasting models. Additionally, AI models can have a low resolution compared to traditional forecasting methods.
Extreme events, such as category 5 hurricanes that occur infrequently, might be under sampled depending on the duration of the data records used to train the AI model. AI systems can be unpredictable when operating under conditions that have never been encountered. When faced with an extreme weather event, AI systems might make highly erratic predictions.
Many AI weather forecast models leave out factors that especially impact people, like wind gusts and precipitation, Radford said. None of the current AI weather forecast models predict types of precipitation, such as snow versus rain.
“The traditional physics-based models are still needed to fill a lot of those gaps,” said Radford. He said forecasters may use multiple tools, leaning on each for their strengths. “I think there are a lot of future avenues or potential avenues for AI weather prediction models.”
Mazurek also looks forward to the future of AI weather models.
“It would be super helpful for these developers to explicitly predict more complex atmospheric variables,” said Mazurek.
Considering future research, Radford said he is most enthusiastic for more rigorous verification testing on AI weather forecast models. Some tests that have been used to compare traditional forecast methods with new AI weather forecast models do not take into account a spatial component.
“So, you forecast precipitation and it’s just a little bit to the left or a little bit to the right of where it’s observed, you can get really high errors,” said Radford. There are tools, such as neighborhood verification, that take spatial offsets into account.
Radford said it helps forecasters and researchers to know specifics about when each type of forecast method works best. They want to further explore how the model performs in certain regions, what situations they perform best in, or how the models might struggle.