Hurricane Lee was minding its business in the vast Atlantic in early September, but meteorologists at the European Centre for Medium-Range Weather Forecasts (ECMWF) kept a keen eye on it. A major concern was its trajectory – would it hit Scotland or pose a threat to the US Northeast?
Traditionally, these weather wizards would rely on models rooted in atmospheric physics. But this time, they had cutting-edge allies: AI weather models, developed by tech behemoths like Nvidia, Huawei, and Google’s DeepMind. The AI’s projection for Lee? A potential hit between Rhode Island and Nova Scotia, aligning with the official physics-based models. But as with all forecasts, the devil lay in the details.
Mark DeMaria, once an atmospheric scientist at the US National Hurricane Center, described the AI models’ entrance into the weather world as almost sudden and unexpected. Initially skeptical about their capabilities, DeMaria’s views transformed when AI predictions showcased their prowess. As it turned out, Hurricane Lee’s landfall in Nova Scotia resonated with AI’s forecast.
So, how do these AI models compare to conventional methods? Traditional models lean on complex atmospheric equations. Over time, with more refined data and enhanced understanding, they’ve improved. But the unpredictable nature of weather, as described by the chaos theory proposed by meteorologist Edward Lorenz in the 1960s, sets a challenging two-week predictive limit.
The AI-powered models, on the other hand, adopt a pattern-based approach, akin to ChatGPT’s text-generation. Rather than embedding the physics, these models discern patterns from historic atmospheric data, such as the ECMWF’s ERA5 dataset. Despite their success, they’re not flawless. Their propensity to focus on more common patterns often undermines outlier events, like severe heatwaves or storms.
DeepMind’s Shakir Mohamed acknowledges the challenges in predicting rainfall and extreme events using AI. Combining radar-based predictions, like DeepMind’s NowCasting, with these models is not straightforward. However, more granular data in upcoming ECMWF datasets may bring rain predictions closer to reality.
One distinct advantage of AI models? Efficiency. While traditional ensemble models, which outline multiple possible weather outcomes, can be time-consuming, AI models churn out several projections rapidly. Mark DeMaria highlights that their FourCastNet model can generate results in mere seconds, even on outdated hardware. This speed makes vast ensemble predictions more feasible.
Yet, AI models aren’t without their blind spots. A critical challenge is the ‘black box’ nature of many machine-learning systems. For weather predictions, understanding the reliability of your model is crucial. Lingxi Xie, a senior AI researcher at Huawei, emphasizes the meteorologists’ demand for clearer explanations from AI forecasts, which currently remains unmet.
While AI models promise broader accessibility for accurate forecasts, leveraging them universally isn’t immediate. Quality predictions require top-tier weather observations from sources like satellites and buoys, processed into datasets by agencies like NOAA and ECMWF. Despite the potential hurdles involving intellectual property and national security, the momentum is undeniable. The traditional and conservative meteorological realm sees the promise AI holds, and DeMaria believes AI might have a formal role in weather forecasts in the coming hurricane seasons.
