DeepMind 的飓风突破令气象科学家感到意外
2025 年 10 月,一场风暴在加勒比海上空酝酿。各气象模型对其路径的预测各不相同。它会保持弱势并最终抵达海地,还是会增强并扑向牙买加?由 Google 的 DeepMind 与 Google Research 共同开发的人工智能模型 WeatherNext 选择了后者。在登陆前五天,它预测这场风暴系统将以 5 级飓风的强度袭击牙买加,置信度为 80%。
飓风 Melissa 造成了灾难性破坏,在牙买加各地引发洪水和山体滑坡。但这一 AI 模型帮助预报人员更早地向其路径上的社区发出预警,使他们能够更好地做好准备。
在周四发表于 Nature 的一篇论文中,研究人员表明,WeatherNext AI 模型能够以前所未有的精度预测气旋。平均而言,它比现有模型为预报人员多提供一天的预警提前量;这意味着它提前三天的预测,与以往模型提前两天的预测一样准确。在实际生活中,这多出的一天可能意义重大。
“哪怕只有几个小时也能带来改变,”美国国家飓风中心主任 Mike Brennan 说。组织疏散、调配物资、调动资源以应对飓风风险,这些都是时间敏感的任务——而做出错误决定可能带来严重后果。“在这类决策上,时间真的非常宝贵,所以能够把预报准确度比我们以往所能达到的水平再提前多达一天,真的非常有价值,”他说。
研究人员表示,从历史上看,把预报提前一天需要花费十年的工作。
对 AI 而言,建模极端事件可能颇具挑战。机器学习需要充足的训练数据才能对未来做出预测,但极端事件本质上就是罕见发生的。“我们没有那么多气旋数据,但我们有大量天气数据,”Google DeepMind 研究科学家、该论文主要作者之一 Ferran Alet 说。“所以我们所做的是训练一个模型,让它既擅长天气,也擅长气旋。”
飓风尤其难以预测,因为它们同时在多个空间尺度上运作,合作大气研究所热带气旋组负责人、该论文作者 Kate Musgrave 说。预测风暴的路径——即它行进的方向——需要全球尺度的天气数据,纳入冷锋位置和盛行风等信息。然而,预测风暴的强度则需要小得多的尺度数据,专门聚焦于当地的大气和海洋状况。
“这正是我们从这些全球模型中无法获得的东西,”Musgrave 说。尽管此前的 AI 模型在预测风暴路径方面表现良好,“但在强度方面它们完全做不好。”
同时预测这两者至关重要:强度的变化可能意味着一个相对较弱的风暴与一场大型飓风之间的差别。有时——就像飓风 Melissa 的情况——一个风暴系统可能迅速增强,在一夜之间发展为紧急状况。Melissa 标志着美国国家飓风中心首次能够在风暴仅处于 1 级阶段时预测出其将达到 5 级。
在 WeatherNext 模型被用于实时预报之前,研究人员在回顾性数据上对其进行了测试。“结果太好了,以至于我们怀疑在实时演示中是否真的能看到这样的表现,”Musgrave 说。但当预报员开始将该模型纳入日常业务时,这一性能表现得到了验证。“我想每个人都对它表现得如此出色感到惊讶,”Musgrave 说。
即便是参与该模型研发的 DeepMind 研究人员,也并不完全理解这个 AI 模型为何能产生如此准确的预测,因为它所使用的大气数据分辨率远低于传统模型预测风暴强度所需的分辨率。“当我们告诉业界我们的模型只使用了相对粗糙的分辨率时,他们都很震惊,因为这意味着低分辨率输入所捕捉到的关于未来走向的信号比此前认为的更多,”Alet 说。
这个 AI 模型一定是从较低分辨率的数据中捕捉到了某些信息,从而能够对风暴强度做出预测,但研究人员并不知道那究竟是什么。“归根结底它是个黑箱,但这给了物理学家一个信号:有某种此前未被理解的现象正在发生,”Alet 说。
该模型不只是给出一个预测结果,而是会为一个正在发展的风暴生成一系列可能的情景。Alet 说,这有助于捕捉任何潜在的“蝴蝶效应”——即对某一趋势的微小偏离,可能在后续引发大得多的变化。预报员可以将这些输出与其他模型的输出结合使用,为他们关于风暴系统可能如何演变的预测提供依据。去年,该 AI 模型对每个风暴生成 50 种情景;如今,它会生成 1,000 种。
“以我们现有的算力,用现有的数值模型根本做不到这一点,”Musgrave 说。
Brennan 说,DeepMind 的模型是预报员工具箱中一件很棒的新工具,但他强调这只是众多工具之一。“不能保证某一个模型因为去年表现好,或者在某一场特定风暴中表现非常出色,就一定会在下一个季节或下一场风暴中成为最好的模型,”他说。他还说,人的因素仍然至关重要。“飓风不仅仅是一个路径或强度预报,”他说。“它需要专家将其转化为实际影响——而正是这些影响夺去了人们的生命。”
Google DeepMind 还宣布,将开源飓风季期间使用的 WeatherNext 模型,以便研究人员使用并加以改进。Alet 希望,向研究社区开放这些模型有助于揭示关于气旋运作机制的新见解。
“我对科学发现感到非常兴奋,”他说。“我认为 AI 正在为我们提供新的工具,去探究宇宙的法则。”
这篇报道最初发表于 wired.com。
DeepMind’s hurricane breakthrough has surprised weather scientists
In October 2025, a storm brewed over the Caribbean Sea. Weather models differed on its trajectory. Would it remain weak and end up in Haiti, or would it intensify and head to Jamaica? Artificial intelligence model WeatherNext, developed by Google’s DeepMind and Google Research, went with the latter. Five days before landfall, it predicted with 80 percent confidence that the storm system would hit Jamaica as a Category 5 hurricane.
Hurricane Melissa was catastrophic, causing flooding and landslides across Jamaica. But the AI model helped forecasters give an earlier warning to communities in its path, so they could better prepare.
In a paper published on Thursday in Nature, researchers show that the WeatherNext AI model can predict cyclones with unprecedented accuracy. On average, it gives forecasters a day more lead time than existing models; this means its predictions three days out are as accurate as previous models’ predictions two days out. On the ground, that extra day can mean a lot.
“Even a few hours can make a difference,” says Mike Brennan, director of the US National Hurricane Center. Organizing evacuations, staging supplies, and moving resources to respond to a hurricane risk are all time-sensitive tasks—and making the wrong decision can have big consequences. “Time is really golden when it comes to those types of decisions, so the ability to push forecast accuracy out as much as a day beyond what we’ve previously been able to do is really valuable,” he says.
Historically, bringing forecasts forward by a day would take a decade of work, the researchers say.
Modeling extreme events can be challenging for AI. Machine learning requires ample training data in order to make future predictions, but extreme events are by nature rare occurrences. “We don’t have that much cyclone data, but we have a lot of weather data,” says Ferran Alet, a research scientist at Google DeepMind and one of the paper’s lead authors. “So what we did was train a model to be both good at weather as well as cyclones.”
Hurricanes are particularly difficult to predict because they operate at multiple spatial scales, says Kate Musgrave, tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere, and an author on the paper. Predicting a storm’s track—which direction it’s traveling—requires data about weather on a global scale, taking in information such as the location of cold fronts and prevailing winds. Predicting a storm’s intensity, however, requires much smaller-scale data focused specifically on the local atmospheric and ocean conditions.
“That’s something we just don’t get from these global models,” Musgrave says. While earlier AI models have done well at predicting a storm’s track, “intensity they could not do well at all.”
It’s critical to predict both: A change in intensity can mean the difference between a relatively weak storm and a major hurricane. Sometimes—as in the case of Hurricane Melissa—a storm system can intensify rapidly, developing into an emergency situation overnight. Melissa marked the first time the National Hurricane Centre was able to predict a Category 5 hurricane when the storm was only at a Category 1 stage.
Before the WeatherNext model was used in live forecasts, researchers tested it on retrospective data. “The results were so good that we were skeptical that we would actually see that in the real-time demonstration,” Musgrave says. But when forecasters started adopting the model into their operations, this performance held true. “I think everybody was surprised at just how well it did,” Musgrave says.
Even the DeepMind researchers working on the model don’t fully understand how the AI model produces such accurate predictions, given that it uses much lower-resolution atmospheric data than traditional models require to forecast storm intensity. “When we told the community that our model was only using relatively coarse resolution, they were shocked, because that means that the lower-resolution inputs capture more signal about what’s going to happen than previously believed,” Alet says.
The AI model must be picking up on something in the lower-resolution data that allows it to make predictions about storm intensity, but the researchers don’t know what. “It’s a black box at the end of the day, but that gives physicists a signal that something is happening that was not previously understood,” Alet says.
The model doesn’t just spit out one prediction; it produces a range of potential scenarios for a developing storm. This helps to capture any potential “butterfly effect,” says Alet, where a small deviation from a trend could lead to much bigger changes down the line. Forecasters can use these outputs, alongside those of other models, to inform their predictions about how a storm system will likely unfold. Last year, the AI model created 50 scenarios per storm; now, it generates 1,000.
“That’s something that, with our computing power, we simply can’t do with our existing numerical models,” Musgrave says.
Brennan says DeepMind’s model is a great new tool in forecasters’ toolbox but emphasizes that it’s one of many. “There’s no guarantee that one model, because it did well last year or really did well for this particular storm, is necessarily going to be the best model for the next season or the next storm,” he says. The human element, he adds, is still critical. “A hurricane is not just a track or an intensity forecast,” he says. “It requires experts to translate that into what the impacts are going to be—and it’s the impacts that kill people.”
Google DeepMind also announced that it is open-sourcing the WeatherNext models used during hurricane season so that researchers can use and improve on them. Alet is hopeful that opening the models up to the research community could help uncover fresh insights into how cyclones work.
“I’m very excited about scientific discovery,” he says. “I think AI is giving us new tools to poke into the laws of the universe.”
This story originally appeared on wired.com.