Why the National Hurricane Center bet big on AI to forecast Isaias

From the time Hurricane Isaias first formed, meteorologists at the National Hurricane Center made a big bet. They shaped their predictions to closely match an AI hurricane model developed by Google DeepMind, rather than leaning on the classic physics-based models many Americans are more familiar with, such as the European and GFS. So far, the bet has paid off.

Forecasts for this storm reflect the improving reliability of AI weather models overall, and signal an astonishingly rapid shift to routine use of such products by human forecasters.

Relying heavily on DeepMind meant NHC forecast Isaias would become a Category 2 from the first advisory on Tuesday, when it didn’t even have a name yet. At that point, many of the physics-based computer models, which rely on mathematical equations describing how the atmosphere works to simulate future weather conditions, were vacillating between different projections but generally forecasting a weaker storm that would take a different track than the DeepMind and other AI models were showing.

If the forecast is accurate, it will mark a huge victory for Google and help the people in the path of the storm, who will have had more time to prepare for a significantly stronger blow. Hurricane Isaias is the first hurricane in the tropical Atlantic season, due largely to the effects of a super El Niño in the Pacific , setting a record for the ocean basin’s latest first hurricane of the season.

The confidence of NHC forecasters in Google DeepMind model in particular — the newest iteration of which is WeatherNext 3 — stems largely from experience. Last season the Hurricane Center correctly anticipated the rapid intensification of Hurricane Melissa almost three days before its devastating arrival in Jamaica based largely on the projections from DeepMind. It also stems from the fact that the Hurricane Center partnered with Google to help train the model and learn more about its capabilities.

As is occurring in many fields right now, the world of meteorology is undergoing rapid advancements due to AI , although these models are used in conjunction with the traditional physics models, which still have their advantages. AI models are trained on past weather events and, in the case of DeepMind’s, are also given current weather data.

⛅ Get CNN Weather in your inbox

The forecast is just the beginning. We’ll send you expert coverage and the stories behind the weather — so you always know more than just the number. Sign up for the newsletter

DeepMind performed extremely well last season, according to NHC director Michael Brennan, and has been accurate in anticipating the intensity of many Pacific Ocean storms this season.

“The AI models have shown more consistency” from one model run to the next, Brennan said regarding the reliance on DeepMind for Hurricane Isaias.

“It’s a combination of probably what we were seeing from looking at these physics-based models and some of the inconsistency there, and then just the performance of the AI guidance so far, and the fact that it has proven to be pretty reliable,” Brennan said.

Part of that proof came from a gutsy call that the Hurricane Center forecasters made last year during Hurricane Melissa.

Last season, in a first, Hurricane Center forecasters explicitly forecast that Hurricane Melissa would reach Category 5 intensity when it was just an 80-mph Category 1 hurricane, largely by relying on the clear projections from DeepMind. In fact, by some measures, DeepMind even outperformed human forecasters during the 2025 Atlantic hurricane season overall.

With global climate change enabling more tropical cyclones to rapidly intensify, the Hurricane Center has been making more bullish first forecasts for certain storms, including Isaias. Before AI stepped onto the forecasting stage, the NHC had employed more of a stair-step approach, gradually raising the intensity ceiling of a storm as it revved up, which often put the official forecasts behind the curve.

That practice seems to have ended. Rapidly intensifying storms are now more common, and the latest models are better able to anticipate it.

“We’ve definitely been making more aggressive intensity forecasts overall for the last several years, and we certainly do have more confidence to try to make some of these more aggressive intensity forecasts early on, like we did for Melissa last year, but we’ve also done for Helene and Milton,” Brennan said.

He cited physical computer model support for rapid intensification plus weather forecasters’ understanding of the environment in which a storm is forming as additional factors allowing early predictions of rapid intensification.

“It’s not just one thing, but certainly the AI models have done pretty well in some of those cases,” Brennan said. “So it’s another tool in the toolbox, so to speak, or another piece of evidence that can help support those more aggressive intensity forecasts,” he said of the AI models.

Brennan said forecasters are not skittish about relying on DeepMind in part because they know the training data that went into it.

“I think the most important thing is to make sure that the models are trained on the most accurate data we have,” he said, “and we’ve worked with Google to help them train on our best track analysis, and that’s one reason why I think their performance has improved in terms of intensity forecasting relative to some of the earlier AI models that did well for track, but did not do well for intensity.” Those models had been trained on large-scale data sets that weren’t hurricane-specific, he said.

Meanwhile, Hurricane Isaias is churning its way toward the northeastern Gulf Coast and is likely to make landfall as a hurricane, according to a consensus of most computer models — not just DeepMind.

从飓风伊萨亚斯形成之初，美国国家飓风中心的气象学家们就做出了一个大胆的尝试。他们没有依赖许多美国人更熟悉的传统物理模型（例如欧洲预报中心和GFS模型），而是调整了预测结果，使其与谷歌DeepMind开发的AI飓风模型高度吻合。到目前为止，这个尝试已经取得了成功。

此次风暴的预测反映了人工智能天气模型整体可靠性的提高，也标志着人类预报员对这类产品的常规使用正以惊人的速度转变。

由于严重依赖DeepMind的预测，美国国家飓风中心（NHC）从周二发布的第一份预报开始就预测伊萨亚斯飓风将增强为二级飓风，当时它甚至还没有正式命名。与此同时，许多基于物理的计算机模型（这些模型依靠描述大气运行规律的数学方程来模拟未来的天气状况）的预测结果在不同的预测之间摇摆不定，但总体上预测的风暴强度较弱，路径也与DeepMind和其他人工智能模型的预测不同。

如果预测准确，这将是谷歌的一次巨大胜利，并能帮助风暴路径上的人们，让他们有更多时间为更强烈的风暴做好准备。飓风伊萨亚斯是热带大西洋飓风季的首个飓风，这主要是由于太平洋超级厄尔尼诺现象的影响，也创下了该海域飓风季首个飓风出现时间最晚的纪录。

美国国家飓风中心（NHC）预报员对谷歌DeepMind模型（尤其是其最新版本WeatherNext 3）的信心主要源于经验。上个飓风季，飓风中心主要依靠DeepMind的预测，在飓风梅丽莎袭击牙买加造成毁灭性破坏前近三天，就准确预测了其快速增强的情况。此外，飓风中心与谷歌合作训练该模型并深入了解其性能，也是其信心的重要来源。

正如目前许多领域正在发生的那样，气象学领域也因人工智能而迅速发展。尽管这些模型通常与传统的物理模型结合使用，而传统物理模型仍然具有其优势。人工智能模型通过对过往天气事件的训练，而DeepMind的模型还会使用当前的天气数据进行训练。

⛅ 在您的收件箱中获取 CNN 天气预报

天气预报只是个开始。我们会为您推送专家报道和天气背后的故事——让您了解的不仅仅是数字。订阅我们的新闻简报。

据美国国家飓风中心主任迈克尔·布伦南称，DeepMind 在上个赛季表现非常出色，并且在本赛季对太平洋许多风暴的强度预测也相当准确。

布伦南在谈到对 DeepMind 在飓风伊萨亚斯预测中的依赖时表示，“人工智能模型在一次又一次的运行中表现出了更高的一致性”。

布伦南说：“这可能是我们从观察这些基于物理的模型中看到的一些不一致之处，以及人工智能指导迄今为止的表现，还有它已被证明相当可靠这一事实的综合结果。”

部分证据来自飓风中心预报员去年在飓风梅丽莎期间做出的一个大胆预测。

上个飓风季，飓风中心预报员首次明确预测飓风梅丽莎会达到五级强度，而当时它仅仅是一个风速80英里/小时的一级飓风，这主要得益于DeepMind提供的清晰预测。事实上，从某些指标来看，DeepMind在2025年大西洋飓风季的整体表现甚至超过了人类预报员。

由于全球气候变化导致更多热带气旋迅速增强，飓风中心对某些风暴（包括伊萨亚斯）的首次预测更为乐观。在人工智能应用于预测之前，美国国家飓风中心（NHC）采用的是阶梯式预测方法，随着风暴强度的增强逐步提高其强度上限，这往往导致官方预测滞后于实际情况。

这种做法似乎已经结束了。如今，风暴迅速增强的情况更为常见，而最新的模型能够更好地预测这种情况。

布伦南说：“过去几年，我们一直在做出更积极的强度预测，而且我们当然更有信心尽早做出一些更积极的强度预测，就像我们去年对梅丽莎飓风所做的那样，我们也对海伦和米尔顿飓风做过类似的预测。”

他指出，物理计算机模型对快速增强的支持，以及天气预报员对风暴形成环境的了解，都是能够及早预测快速增强的额外因素。

布伦南说：“这并非单一因素造成的，但人工智能模型在某些情况下确实表现出色。” 他谈到人工智能模型时说：“所以，可以说，它是工具箱里的又一件工具，或者说是能帮助支持更积极的强度预测的另一项证据。”

布伦南表示，预测者并不担心依赖 DeepMind，部分原因是他们了解 DeepMind 所使用的训练数据。

“我认为最重要的是确保模型使用我们掌握的最准确的数据进行训练，”他说，“我们与谷歌合作，帮助他们利用我们最佳的路径分析数据进行训练，这也是为什么我认为他们的模型在强度预测方面比之前一些在路径预测方面表现良好但在强度预测方面表现不佳的AI模型有了显著提升的原因之一。” 他表示，那些模型是基于并非专门针对飓风的大规模数据集进行训练的。

与此同时，飓风伊萨亚斯正向墨西哥湾东北沿岸移动，并可能以飓风的强度登陆，这是大多数计算机模型（不仅仅是DeepMind）的共识。