Hilburn, a scientist with the Cooperative Institute for Research in the Atmosphere and Ebert-Uphoff, who is in the Department of Electrical and Computer Engineering and CIRA at CSU, have begun to explore how weather forecasters can adapt to – or anticipate – these errors.

Predicting the weather has long been a challenge for computer forecast models. The weather is governed by complex equations that computers struggle to replicate, and feeding the models we have with enough good data is a permanent challenge. Even the best forecasts are only accurate for a few days before data and equation challenges make the forecast inaccurate.

Generative artificial intelligence such as ChatGPT can can create exciting new opportunities, but science ethicists are also worried that it can be misused, making further study of this technology more important than ever.

Applied to weather forecasting, generative AI can be trained to replicate forecasts on a timescale similar to current numerical weather prediction forecasts, but at a much faster rate – up to a thousand times faster, in fact. This increase in speed means you could potentially also use the technique to make forecasts at much higher resolution, including short-term forecasts for rain and snow, for example.

The challenge, however, is that researchers still don’t understand the pitfalls, the CSU scientists say. If AI models are trained on bad data, or even incomplete data, they won’t be able to predict an accurate future – and knowing what is bad or incomplete data is the hardest thing to understand.

Enter Ebert-Uphoff and Hilburn, both at CSU’s Walter Scott, Jr. College of Engineering. Ebert-Uphoff has long been looking at the responsible use of AI for weather forecasting, in collaboration with Hilburn, a research scientist and doctoral candidate. Coming from a satellite remote sensing background and now using AI for a variety of purposes, Ebert-Uphoff’s work includes adapting AI to such topics as fire weather and smoke prediction.

Among these issues are how exactly to evaluate AI models before they can be used safely for the daily weather forecast.

“AI models have the potential to improve our forecasting abilities for severe weather events in terms of both accuracy and lead time,” said Ebert-Uphoff. “However, AI models have their own failure modes and risks, and we need to understand those better before we can safely use AI forecasting models in operations.”

As somebody with a meteorology background who’s now fully versed in AI, Hilburn has close knowledge of both worlds.

“AI has the promise to reduce the persistent biases that we see in weather models, but how well AI can handle unlikely or extreme events remains an open question,” said Hilburn. “One thing we are seeing is that AI-based predictions have different characteristics from physics-based predictions, so meteorologists need to learn about these differences to correctly interpret AI output. CSU and CIRA are in an excellent position to take a leading role in the evaluation of AI forecasting models, based on their extensive experience developing AI models for meteorological applications and their close collaborative ties with NOAA.”

To this end, Ebert-Uphoff and Hilburn have published an article in the ‘News and Views’ segment of Nature, the preeminent science journal internationally. Titled “The Outlook for AI Weather Prediction,” they cover the benefits and potential issues with AI forecasting. The article in Nature can be found online at col.st/6ptB8.

With these better forecasts just around the corner, CSU research will continue to guide and inform the use of this technology to the benefit of all.