---
格式版本: 2
标题: "Ask a Scientist: How do researchers use AI to predict a cyclone?"
原文链接: "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/weathernext-extreme-weather-cyclone-predictions/"
发布日期: "2026-09-01"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:scrape:strict_html_metadata"
发布时间证据: "article:published_time: 2026-09-01"
发布时间校准原因: "规则确认唯一严格发布时间，来源 scrape:strict_html_metadata"
发布时间校准置信度: "high"
发布时间候选数量: 11
发布时间严格候选数量: 3
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-09-02T14:47:22+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-09-02T14:46:59+08:00"
入库时间: "2026-09-02T06:47:23.152Z"
来源平台: "固定入口"
搜索渠道: "fixed_url"
搜索词: "https://blog.google/"
匹配关键词:
  - "AI"
相关厂家:
  - "Google"
相关专家:
  []
内容类型: "网页"
抓取工具: "Free Fetch + Defuddle"
清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取"
原始附件:
  []
AI优质: "否"
AI打分: 33
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "正文主线是Google DeepMind利用WeatherNext预测气旋，属于天气模型及行业应用，并非超节点或机架级AI基础设施。来源为Google官方博客且正文完整；相较知识库中的TPU 8t/8i及AI Hypercomputer内容，本文仅提到模型可在单颗TPU运行，没有TPU代际、机架拓扑、互连、功耗或部署架构增量。可抽取事实包括使用50年天气数据、与美国国家飓风中心合作、模型开源及单TPU推理，但主要是既有成果访谈式回顾，且无基础设施量产、订单或规模部署信号。命中“应用与模型效率”硬否决项。"
AI质检模型: "gpt-5.6-sol"
AI质检时间: "2026-09-02T14:47:48+08:00"
AI主题相关性: 1
AI来源权威性: 13
AI新颖性: 5
AI技术细节: 3
AI商业部署信号: 2
AI完整性: 9
AI评分提示词版本: "v17-精简生产版"
AI评分提示词SHA256: "48fb9777f386026761b4873eaff30807694fb11e9b352d7c69bf2dfde750cc7d"
AI评分知识库版本: "knowledge_base_v1-20260819+runtime.81"
AI评分知识库SHA256: "3b93d12e47749b3512f545f51c44c011bdc0931677c2cfe61e4df59a2b1a5a48"
AI评分知识库检索词: "[\"Google\",\"https://blog.google/\",\"TPU\",\"AI-powered\"]"
AI评分知识库命中: "[{\"id\":\"historical-jan-apr-02\",\"title\":\"二、Google Cloud Next '26：AI Hypercomputer 与第八代 TPU 发布\",\"sourceType\":\"curated_item\",\"time\":\"2026-01_to_2026-04\",\"matchedTerms\":[\"Google\",\"TPU\"],\"rank\":-13.320100201888973},{\"id\":\"runtime-e8b946c3131877fe7add476b\",\"title\":\"Peeling Apart That Supposed $120 Billion Chip Deal Google Inked With Marvell\",\"sourceType\":\"ai_excellent_article\",\"time\":\"2026-08-27\",\"matchedTerms\":[\"Google\",\"TPU\"],\"rank\":-10.829077185808972},{\"id\":\"july-correct-0033\",\"title\":\"Microsoft, Alphabet, Meta Pivot from Buy to Build in AI\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"Google\",\"TPU\"],\"rank\":-6.970349016487022},{\"id\":\"july-correct-0024\",\"title\":\"Salience Labs Wants To Scale Up AI With Silicon Photonics Optical Switch\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"Google\",\"TPU\"],\"rank\":-6.635307933939599},{\"id\":\"historical-jun-025\",\"title\":\"英伟达、谷歌与国产超节点的三种网络选择\",\"sourceType\":\"curated_item\",\"time\":\"2026-06\",\"matchedTerms\":[\"TPU\"],\"rank\":-6.255281506774356}]"
AI摘要: "Google DeepMind研究人员开发WeatherNext AI模型，利用50年历史气象数据，在单个TPU上即可预测气旋路径与强度，取代了依赖巨型超算的传统方法。"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-02T08:11:09.015Z"
采集批次: "2026年9月2日14点42分57秒"
采集批次ID: "20260902-144257-884"
去重键: "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/weathernext-extreme-weather-cyclone-predictions"
---

As extreme weather events impact more communities around the world, the ability to accurately predict a storm's path is critical to keeping people safe. By combining decades of historical atmospheric data with advanced AI, Google researchers have achieved [a massive leap](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/weathernext-2-cyclones/) in cyclone forecasting accuracy — turning what once required building-sized supercomputers into life-saving models that can run on a single TPU. We sat down with Ferran Alet, a research scientist at Google DeepMind who helped develop our [WeatherNext](https://deepmind.google/science/weathernext/) models, to explore how this breakthrough technology is giving people and first responders the vital gift of time before a cyclone strikes.

**What do you do at Google?**

I came to Google DeepMind’s weather team after a background in academia exploring the intersection of AI and physics. Now, I’m using that background to apply AI to traditional physics-based models to better predict future weather. My team works on the WeatherNext forecasting models, which we've built to help meteorologists and scientists access AI-powered extreme weather predictions. These predictions help people get more accurate forecasts and warnings ahead of potentially devastating weather events, like cyclones.

**Before we go further: cyclones, hurricanes, typhoons — which is it?**

They all mean the same thing! Cyclones is the global scientific term for hurricanes and typhoons. In the Atlantic region, they’re called hurricanes, and in the North Pacific, they say typhoons. But all of the terms refer to a large-scale, rotating storm system that forms over a warm ocean. The warmth in the ocean causes warm, moist air to rise, which creates an area of low pressure beneath it and draws in cooler air. The Earth’s rotation causes that cycle of rising moist air and drawing in of cooling air to spin. Combine that with heavy rain and strong winds, and you have a cyclone…or hurricane or typhoon, depending on where you live.

**How have researchers predicted cyclones in the past?**

Historically, forecasting cyclones relied on physics-based models powered by supercomputers the size of shipping containers, or even as big as three-story buildings. Scientists used these machines to simulate the laws of fluid dynamics, predicting future weather based on *current* weather data from global sensors. But the data from those global sensors is often incomplete or noisy, and simulating exact physics requires immense computing power, which slows down the progress we can make in forecasting things like cyclones. The meteorological community historically gained about one day of forecast accuracy per decade, by building better models, launching more satellites and getting more computing power. By applying AI to the data from these physics-based models, we saw that same [decade of progress](https://deepmind.google/blog/weathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones/) in a single generation leap of WeatherNext models.

**How are we using AI to change the way we predict cyclones?**

Physics models are really good at telling us about the past and the present. AI allows us to better predict the future, based on patterns from the past and data from the present. With WeatherNext, we trained AI on 50 years of historical weather data. Instead of having to rely on building-sized supercomputers, researchers can run these faster models on a single TPU. This gives the scientific community a massive leap in accuracy, improving predictions for both where a cyclone will hit and how intense it will be when it makes landfall. Weather agencies are looking to extend their public forecast horizons from five days to seven, giving communities vital extra time to prepare before a cyclone hits.

**What kind of difference can that extra time make for people who need it?**

We saw how crucial that time was last year with [Hurricane Melissa](https://deepmind.google/blog/how-weathernext-helped-the-national-hurricane-center-better-predict-hurricane-melissas-historic-landfall-in-jamaica). Ahead of hurricane season, we partnered with the National Hurricane Center to support their forecasts. What was notable about Melissa was that, while the hurricane initially appeared weak, the AI's high-confidence predictions helped the National Hurricane Center to issue a historic forecast, warning authorities of its rapid intensification from a Category 1 storm into a Category 5. It proved to be one of the most powerful hurricanes in history, which we could see from our models days in advance of it making landfall. The additional day of warning allowed authorities in Jamaica to make life-saving emergency preparations.

**What does technology like WeatherNext mean for the future of extreme weather forecasting?**

We’ve [open-sourced](https://github.com/google-deepmind/weathernext) our WeatherNext 2 and WeatherNext cyclones models, and developed an interactive [Weather Lab](https://deepmind.google.com/science/weatherlab) website to advance science for a global scientific community. This widespread access helps academics and scientists freely experiment with the models and hopefully make new discoveries that help people around the world.