How a Big Tech AI model predicted Hurricane Lee

On September 10, as Hurricane Lee moved slowly westward through the middle of the Atlantic Ocean, three new weather models developed in the private sector predicted that the hurricane would make landfall in Nova Scotia in about a week. Because the storm was still thousands of miles from North America, this prediction turned out to be a surprisingly accurate feat for a technology that was considered to be in its infancy not so long ago.

Generated by artificial intelligence, this model is orders of magnitude faster and cheaper to operate than traditional government-run weather models. Although AI models do not yet provide all the capabilities needed for operational forecasting, their emergence portends the potential for major changes in the way weather forecasts are made, and the could signal the beginning of a new chapter in the weather forecasting race.

“Incredible stories are emerging about the role of AI weather forecasting,” Daniel Rothenberg, an atmospheric scientist at Google’s sister company Waymo, said in an email. “This is a glimpse into the future of meteorology, probably on a much faster timescale than most people in the weather industry expect.”

AI weather models have advanced rapidly over the past 18 months. Google, Microsoft, NVIDIA, and China-based Huawei have all published academic papers claiming their AI models perform at least as well as the “European model,” widely considered the gold standard for weather modeling. Announced. These claims were recently corroborated by scientists at the European Center for Medium-Range Weather Forecasts, which operates the European model. Startups like Atmo, Excarta, and Zurich-based Jua are also building AI weather models.

Several years ago, the European Center began exploring the potential of AI to further improve predictions. Earlier this month, the day after Tropical Cyclone Lee developed in the Atlantic Ocean, the European Center began publishing forecasts from Google, NVIDIA, and Huawei on his website. The model uses the current situation in the European model as a starting point and generates a 10-day forecast at six-hour intervals in about a minute, according to the European Center.

After predicting that Lee would make landfall in Nova Scotia on September 10, the AI ​​model fluctuated slightly over the next few days, but was consistent in predicting landfall between Cape Cod, Massachusetts and eastern Nova Scotia. The AI ​​model is “probably as good” as European and American models, Rosenberg said, and was the first to accurately suggest that Lee could approach New England.

AI weather forecasting “suddenly emerged as a legitimate contender for the world.” [conventional models]” wrote Richard James, a meteorologist at Prescient Weather, which provides weather forecasting tools to the energy and agricultural industries, in an analysis of the AI ​​forecast for Hurricane Lee.

James cautions that one storm is too small a sample to prove that AI models are better than traditional models, but “given the impressive pace of technological innovation in recent years, …It’s not hard to imagine.” [AI] would be able to replace [conventional] “Models for at least some applications will be developed in the relatively near future,” he wrote.

The improved performance of AI models has drawn attention not only from the European Center but also from the National Oceanic and Atmospheric Administration, which operates the U.S. Global Forecast System model, also known as GFS. The two institutions have long been competitors in computer modeling, with the European model proving more accurate overall.

NOAA’s Center for Artificial Intelligence was established in 2021, and this week the agency held its fifth annual AI Workshop. According to Ime Evert, the Collaborative Institute for Atmospheric Research (CIRA) at Colorado State University, part of NOAA, will soon launch a website similar to the European Center, which will begin with the current state of the American model and make predictions using AI models. is displayed. – Uphoff, head of machine learning at CIRA, points out the importance of evaluating models.

“It is a matter of public safety to carefully evaluate these AI-based models. First, we want to ensure that we are using all the tools available to improve predictions of severe weather events, and we believe that AI The base model could be very useful for that,” Evert-Aphoff said in an email. “On the other hand, we also need to be careful not to jump too quickly to the latest models. AI models can have pitfalls, including: [conventional] The model doesn’t have that. ”

Differences between traditional weather models and AI weather models

Both traditional and AI weather models use current atmospheric conditions as a starting point for their predictions, and that’s where the similarities end.

Computer models, programmed with complex formulas and operated by the world’s major government weather agencies, have long served as the backbone of forecasts and warnings. Although the accuracy of these traditional models has steadily improved over the decades, the trillions of calculations required to run a single model require enormous computational power; Operational costs are higher.

AI models are first trained to recognize patterns in vast amounts of historical weather data. Generate predictions by taking the current situation and applying what has been learned from the past. This process is much less computationally intensive and can be completed in minutes or seconds on a desktop computer, compared to an hour or more on a traditional model large supercomputer.

Many experts say AI models will probably never make traditional models obsolete. This is because traditional models are required to train AI models, and, at least for now, information about the initial state of the atmosphere is also fed into the AI ​​models. However, the speed and efficiency of AI models could change the way weather forecasts are made, allowing for more accurate and detailed predictions, especially in the case of extreme weather events.

Neil Jacobs, former NOAA acting administrator and chief scientific advisor for NOAA’s Next Generation Modeling effort, envisions a day when AI models generate predictions and traditional models are used only for training. Jacobs points to the potential for computers to run AI models more frequently and at higher resolutions without worrying about straining resources.

“It’s crazy to think what you could do with this if you took the limitations of high-performance computing off the table,” Jacobs said in an interview. “NOAA cannot afford to buy a system large enough to run the current model. [highest] Resolutions that can be set. Using AI-based systems solves that problem. ”

Advantages and limitations of AI

One of the most promising applications of AI in weather forecasting is ensemble modeling. In ensemble modeling, the same model is run multiple times, each time starting with slightly adjusted initial atmospheric conditions to represent the uncertainties and approximations made by the model. The result is a range of possible outcomes rather than a single forecast, which meteorologists use to identify the most likely forecast and assess its reliability.

Ensemble forecasts from traditional models are limited to about 50 simulations due to the time and cost of generation, and can miss extreme events such as excessive rainfall or heat. AI allows much larger ensembles to be generated in just minutes, potentially enabling more useful predictions and risk assessments for emergency managers, the general public, and numerous industries.

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

The European Center says it believes ensembles are “the key to delivering valuable predictions on medium time scales” and has launched a project to create its own AI-based system.

Despite recent advances, AI models have limitations. For example, it is not possible to create all forecasts for many important parameters, such as precipitation and cloud cover. They 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 this is the future, especially when it comes to operational forecasting,” Jacobs said.

Source link

Related Articles

Leave a Reply

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

Back to top button