Every day, meteorologist Hannah Wangari gets free graphs and maps created by the five predictive models she subscribes to and interprets what she sees. “What are the chances of it raining in different parts of the country?” she might wonder. “How much is likely to fall in the next 24 hours?” Answering questions like this quickly and accurately can save lives, as she and others at the Kenya Meteorological Department do. essential for sexual work.
As climate change causes extreme weather events to become more frequent and intense, the need for faster and more accurate forecasts will only increase. Heavy rains and floods have wreaked havoc this year, killing hundreds and displacing countless more, affecting more than 7.2 million people across vast swathes of the United States, Spain, Central Europe and Africa. . Ann An estimated 267 people have died in Kenya. Last year, 42 of the country’s 47 counties were flooded, and an additional 278,000 people were evacuated. The storm is expected to become even more intense in the future. 7 percent for every 1.8 degrees Fahrenheit Given the effects of global warming, accurately predicting when and where such events will occur is key to saving lives and livelihoods.
However, it can be a time-consuming and expensive task. Traditional forecasting relies on a method called numerical weather forecasting. Developed in the 1950s, this physics-based technology includes millions of–dollar A supercomputer that can solve complex equations that mimic atmospheric processes. Intensive numerical calculations can take hours to produce a single forecast, making it out of reach for many forecasters, especially in developing countries, and relying on data produced by others. It will depend on you.
Tools powered by artificial intelligence are becoming faster, often more accurate, and are becoming an alternative that can be easily created on a laptop. they use machine learning It leverages 40 years of open source weather data to identify patterns and trends that can help predict what will happen in the future. “They’re basically using the past to train a model to learn physics,” said John, who heads the NSF AI Institute for Trustworthy AI in Weather, Climate, and Coastal Oceanography at the University of Oklahoma. said computer scientist Amy McGovern.
AI-powered methods developed by Google, the University of Oxford, NVIDIA and others can provide accurate predictions within minutes, giving governments more time to prepare and respond. “More frequent updates will help government agencies monitor rapidly changing conditions, such as the path of a storm,” Dion Harris, who leads NVIDIA’s Accelerated Data Center, told Grist. “This will improve decision-making regarding evacuation planning, infrastructure protection, and resource allocation.”
As commercial weather forecasting becomes more widespread, who will be left behind?
Users like government meteorologists in Nairobi can enhance these models with local data such as surface temperature and humidity, as well as free satellite data, to tailor forecasts to specific geographic areas. The Kenya Meteorological Department is working with Oxford, the European Center for Medium-Range Weather Forecasts, Google and the World Food Program to develop the AI model. Improves accuracy of rainfall prediction.
Of the five traditional models used by the Kenya Meteorological Department, four offer only free charts and maps, which Wangari has studied closely. To access predictive data, you need to pay a license fee or own a supercomputer to run the model. Instead, she and her colleagues are analyzing the open source data they receive to see what happens next. Machine learning models developed in collaboration with Oxford can be used to evaluate real forecast data to determine the likelihood of extreme weather events. “For the first time, we were able to generate what are called probabilistic predictions,” she said. “When you give people a probability that something will happen, they are more likely to take action.”
“We can now say things like, ‘This area is going to get 2 inches of rain in the next 24 hours, and there’s a 75% chance that we’ll exceed this threshold,'” she says.
AI models take only minutes to generate predictions and provide the ability to make and explore more predictions. wider range of possible outcomes. This leads authorities to play what McGovern calls a “what-if game” and say, “If this happens, we need to evacuate this area,” or, “If this happens, we might want to take this action.” You can say “No.” They can predict the most likely scenarios and prepare for the worst, such as preemptively evacuating people with disabilities.
The machine learning techniques developed by Oxford and used by Wangari are proven. more effective than other methods Regarding rainfall prediction. That’s not unusual. Google’s GenCast, announced last monthoutperformed traditional prediction models. 97% of 1,320 indicators. Its predecessor, GraphCast, has proven that. more accurate It is better than the world’s best conventional tool, run by the European Center for Medium-Range Weather Forecasts. “AI produces much better results than physics-based models,” said Florian Pappenberger, deputy director of the European Center, which plans to launch its own AI model this year. It also happens more quickly. GenCast generates a 15-day weather forecast in less than eight minutes, and NVIDIA claims its FourCastNet is 45,000 times faster than numerical weather forecasts.
AI has also been proven to more accurately predict the path of hurricanes. Google scientists say GraphCast accurately predicted where Hurricane Lee would make landfall nine days before it hit Nova Scotia as it barreled through the North Atlantic in September 2023, and three days earlier than traditional forecasting methods. . told the Financial Times. Shruti Nath, a climate researcher at the Oxford Project, said two machine learning models accurately predicted Hurricane Milton’s path across the Gulf of Mexico, but underestimated the hurricane’s wind gusts and pressure. However, these tools are expected to improve as errors are corrected and models are fine-tuned.
Of course, predictions are only as useful as the predictive actions they lead to. Researchers developing them need to work with local meteorologists and others with local expertise to understand what it means for local communities and respond accordingly. Mr. Nass said that there is.
Questions remain about how accurately machine learning can predict edge cases like a 100-year flood beyond the datasets used for training. But “it actually represents a much better picture of the extremes than many of us initially expected,” Pappenberger said. “Maybe they’re learning more physics than we ever imagined.” These tools can handle all the things forecasters typically use, such as cloudiness, fog, and snowfall. It hasn’t produced any output yet, but Pappenberger is confident it will be done in time.
The overlooked impact of AI on the climate
Users may also benefit from hybrid models like Google’s NeuralGCM, which combines machine learning and physics. This is an approach that offers the benefits of AI, such as speed, as well as long-term predictive ability and other strengths of numerical weather forecasting.
Improving forecasts is aimed at addressing climate change, but there is also the risk of contributing to climate change. The data centers required to run AI are extremely energy intensive. Google and Microsoft rely on nuclear power The plants that provide it. Still, the supercomputers needed to run numerical weather forecasts are energy-intensive as well. Could be 1,000x cheaper In terms of energy consumption.
McGovern believes cross-functional collaboration is key to realizing the potential of AI models to democratize prediction. While the computing power needed to train models resides primarily in industry, academia, which writes much of the code and makes it available on the public software platform GitHub, has no need to provide quarterly reports. There are luxuries and so is the government. The ultimate end user knows what it takes to save a life, she explained.
For now, researchers and the private sector are working closely to improve the technology. “There’s a lot of collaboration going on, copying each other and trying to improve based on what others have created,” Pappenberger says. Many of these tools are available for free to researchers, but access to others depends on the situation. free They range in price from low to low, depending on the features you use and the specific hardware you purchase. Still, the models are cheaper than supercomputers, allowing organizations like the Kenya Meteorological Department to quickly and easily create forecasts tailored to local needs at a fraction of the cost of physically-based models. It will be.
Using traditional tools to create predictions relevant to people in Nairobi or Mombasa, for example, requires zooming in on a global map to get more detail and then manually analyzing large amounts of data. “With machine learning, you can generate a prediction for a specific point, given the precise coordinates,” she said. Doing so will make it much easier for her and others in similar jobs to stay on top of upcoming weather conditions and ultimately save lives.
