How top AI models pulled off Hurricane Lee predictions

How top AI models pulled off Hurricane Lee predictions

As Hurricane Lee was slowly moving westward in the mid-Atlantic on September 10, three new privately developed weather models predicted that the storm would make landfall in Nova Scotia about a week later. With the storm still thousands of miles away from North America, the forecast turned out to be a surprisingly accurate accomplishment of a technology that not long ago was considered in its infancy.

These AI-generated models are faster and cheaper to run than traditional government-run weather models. While AI models do not yet provide all the capabilities needed for practical forecasting, their emergence heralds a potentially radical change in how weather forecasts are made, and could signal a new chapter in the competition for weather forecasting between the United States and Europe.

“There is an amazing story that has emerged about the role of AI weather forecasting,” Daniel Rotenberg, an atmospheric scientist at Google’s sister company Waymo, said in an email. “This is a glimpse into the future of meteorology, perhaps on much faster time scales than most weather professionals expect.”

AI weather models have made rapid advances in the past 18 months. China-based Google, Microsoft, Nvidia, and Huawei have published academic articles claiming that their AI models perform at least as well as the “European model,” widely considered the gold standard in weather modeling. These claims were recently confirmed by scientists at the European Center for Medium-Range Weather Forecasts, which runs the European model. Startups including Atmo, Excarta and Zurich-based Jua are also building AI-powered weather models.

The European Center began exploring the potential of artificial intelligence to improve its forecasts several years ago. Earlier this month, just one day after Tropical Storm Li developed in the Atlantic, the European Center began publishing forecasts from Google, NVIDIA, and Huawei on its website. The models use current conditions from the European model as a starting point to produce 10-day forecasts at six-hour intervals of approximately one minute, according to the European Center.

After predicting on September 10 that Lee would make landfall in Nova Scotia, AI models fluctuated slightly in the following days, but were consistent in predicting landfall between Cape Cod, Massachusetts, and eastern Nova Scotia. Rotenberg said the AI ​​models were “just as good” as the European and American models, and were the first to accurately suggest that Lee could veer off near New England.

AI weather forecasting “suddenly emerged as a legitimate competitor to (traditional models),” Richard James, a meteorologist at Prescient Weather, which provides weather forecasting tools to the energy and agriculture industries, wrote in an analysis of AI forecasts for Hurricane Lee.

While James cautions that one storm is too small in the sample to prove that AI models are better than traditional models, “given the pace of innovation observed in just the last few years… it is not hard to imagine that[AI]will be able “Replace (traditional) models for at least some applications in the relatively near future,” he wrote.

Improving the performance of AI models has attracted the attention of not only the European Center, but also the National Oceanic and Atmospheric Administration, which runs the US Global Forecast System model, also known as GFS. The two agencies have long competed in computer modeling, with the European model showing greater overall accuracy.

NOAA’s Center for Artificial Intelligence was established in 2021, and this week the agency held its fifth annual Artificial Intelligence Workshop. NOAA’s Cooperative Institute for Research in the Atmosphere (CIRA) at Colorado State University will soon launch a website similar to the European center, which will feature forecasts from artificial intelligence models that start with current conditions from the U.S. model, according to Amy Ebert. – Uvoff, Machine Learning Lead at CIRA, who pointed out the importance of evaluating models.

“It is a matter of public safety to carefully evaluate these AI-based models. On the one hand, we want to ensure that we use all available tools to improve forecasts of severe weather events, and AI-based models have the potential to “On the other hand, we also need to make sure we don’t jump to the latest models too quickly. AI models can contain risks that do not exist in (traditional) models.

The difference between traditional weather models and artificial intelligence models

Traditional weather models and AI models use current weather conditions as a starting point for forecasts, but that’s where their similarities end.

Computer models programmed with complex mathematical equations and run by the world’s leading government meteorological agencies have long served as the backbone of forecasts and warnings. The accuracy of these traditional models has improved steadily over many decades, but they are expensive to run because they require enormous computing power to perform trillions of calculations to run a single model.

AI models are first trained to recognize patterns in vast amounts of historical weather data. They generate forecasts by assimilating current conditions and applying what they have learned from the past, a process that is less computationally intensive and can be completed in minutes or even seconds on a desktop computer, compared to more than an hour on the large supercomputers of traditional models.

Many experts say that AI models probably won’t make traditional models obsolete. This is because traditional models are essential for training AI models, and, at least for now, they also feed AI models with information about the initial state of the atmosphere. However, the speed and efficiency of AI models could change the way weather forecasts are made and allow for more accurate and detailed forecasts, especially for severe weather events.

Neil Jacobs, former acting head of the National Oceanic and Atmospheric Administration (NOAA) and chief science advisor for the agency’s next-generation modeling effort, envisions a day when AI models generate predictions, and traditional models are used only for training. Jacobs points out that AI models can be run more frequently and with greater accuracy without having to worry about draining computer resources.

“It’s crazy to think about what you could do with this once you take the limitations of HPC off the table,” Jacobs said in an interview. “NOAA can’t afford a system large enough to run the model at the current (higher) resolution you can configure it to. Well, that problem goes away if you use an AI-based system.

Benefits and limitations of artificial intelligence

One of the most promising applications of AI for weather forecasting is ensemble modeling, which is when the same model is run multiple times, each time starting with initial weather conditions slightly modified to represent the uncertainties and approximations provided by the model. The result is a set of possible outcomes, rather than a single forecast, that meteorologists use to determine the most likely forecast and assess confidence.

The ensemble forecasts from traditional models can miss extreme events, such as heavy rainfall or heat, because they are limited to about 50 simulations, due to the time and cost generated. AI can enable much larger collections to be generated in less than a few minutes, which could lead to more useful forecasts and assessments for emergency managers, the general public, and many industries.

“Our hypothesis is that we can now easily scale AI models to thousands or tens of thousands of group members,” Anima Anandkumar, senior director of AI research at NVIDIA, said in an interview.

The European Center says it believes the ensembles are “essential for providing valuable forecasts over medium-term timescales,” and has begun a project to create its own AI-based system.

AI models have limitations despite their recent progress. For example, not all of them have yet produced forecasts for a number of key factors, such as precipitation and clouds. They will also need to gain the trust and understanding of forecasters who have spent their careers working with traditional models. But the rapid pace of innovation has meteorologists excited about the possibilities.

“I think it’s the future, especially in terms of operational forecasting,” Jacobs said.

You may also like...

Leave a Reply

Your email address will not be published. Required fields are marked *