[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-deepmind-weathernext-cyclones-nature-open-source":3,"news-related-22a1a718-0eb6-46e5-8ee8-825400de11d1":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"22a1a718-0eb6-46e5-8ee8-825400de11d1","DeepMind WeatherNext 在 Nature 发论文：用 28 km 粗分辨率做出多一天的飓风预警,代码权重全部开源","DeepMind 与 Google Research 联合发布 Nature 论文《Operational Tropical Cyclone Forecasting with AI》,把 WeatherNext Cyclones 模型同步开源到 GitHub。论文报告显示,这套 0.25°(~28 km) 分辨率的 AI 模型在 2023–2025 年间对飓风的路径、强度、风圈结构都给出平均 1 天以上的提前量,优势相当于过去十年气象学整体进步。模型在 TPU 上 1 分钟跑完 15 天预报,可一次生成 1000 个集合成员,极端事件捕捉能力较传统 50 成员提升 20 倍。","## 一句话结论\n\nGoogle DeepMind 与 Google Research 在 Nature 上发表 WeatherNext Cyclones(WN-C),并把代码与权重全部开源;同一天的工程博客显示,这套 AI 模型用比传统区域模型粗约 100 倍的输入数据,就能对热带气旋的路径、强度、风圈给出平均多 1 天的提前量。\n\n## 技术背景:飓风为什么难预报\n\n飓风(热带气旋)是地球上最危险的天气现象之一。过去 50 年里,全球范围内飓风导致超过 70 万人死亡、1.4 万亿美元经济损失(数据见 DeepMind 官方博客)。飓风预报长期受两类矛盾制约:第一,飓风的**路径**依赖大尺度全球大气环流,这正是全球粗分辨率数值模式擅长的;第二,飓风的**强度**由气旋内核附近高度局地的热力过程驱动,这正是高分辨率区域模式擅长的。\n\n传统路线只能让两种模式并存、靠人工会商拼接。\n\n## WeatherNext Cyclones 的方法\n\nWN-C 是一个**端到端单一 AI 模型**,把全球大气动力学和专家标注的历史气旋数据库联合训练,做到路径、强度、风结构同时输出,论文给出的几个关键设计是:\n\n- **数据**:训练数据来自近 20 TB 的全球大气再分析数据(基于 ECMWF),叠加 IBTrACS 提供的近 5000 个历史气旋轨迹(论文摘要 + DeepMind 博客)。\n- **架构**:采用 Functional Generative Networks(FGN,arXiv:2506.10772)做集合预报,一次推演 1 个 15 天预报只需 1 分钟以内,可在 TPU 上生成 1000 成员的物理一致集合(传统 50 成员)。\n- **输入分辨率**:**只**用 0.25°(约 28 km)输入,比传统区域模式粗约 100 倍;团队还放出了一个 1°(约 111 km)Mini 版本,可在 Colab 单 TPU 上跑。\n- **关键反常识**:DeepMind 团队明确写明,这样粗的输入做出 SOTA 强度预测,**是一个开放研究问题**——他们自己也不完全确定模型从低分辨率里学到了什么信号。\n\n## 实测效果(论文 + Ars Technica 双源印证)\n\nNature 论文摘要给出的核心数字是:2023–2025 年的回看评估中,WN-C 在路径、强度、风圈上拿到平均 1 天以上的提前量,「相当于过去十年运营预报的整体进步」。\n\n更细的数据可见 DeepMind 8 月 6 日博客和 Ars Technica 8 月 8 日的报道:\n\n- **2025 年飓风季实时使用** —— 美国国家飓风中心(NHC)使用 WeatherNext 时,首次在风暴仍处于 1 级时就预测出它将以 5 级强度登陆牙买加,即 5 天前以 80% 置信度命中「牙买加五级飓风」路径(Ars Technica 转述 DeepMind 主作者 Ferran Alet 与 NHC 主任 Mike Brennan 的口述)。\n- **集合规模跃升** —— 去年单风暴 50 个成员,今年扩到 1000 个,用于捕捉「快速增强」这类罕见但关键的极端情景。\n- **业务化部署** —— 论文之外,DeepMind 把训练代码、权重、WN-C 与 WeatherNext 2 一并开源到 `google-deepmind\u002Fweathernext` 仓库,论文之外的预报数据通过 Google Cloud(Vertex AI \u002F BigQuery \u002F Earth Engine)、Weather Lab 可视化页、Open-Meteo 三条渠道放出。\n\n## 与已有 AI 预报的相对位置\n\nWeatherNext 系列并不是新闻。DeepMind 早在 2023 年的 GraphCast 和 2024 年 12 月的 GenCast,都把「AI 跑数值预报」这件事推到与传统 ECMWF 模式同台的位置。本次的关键变化是:WeatherNext Cyclones **首次走完 Nature 同行评审 + 落地到 NHC 业务系统的全程闭环**,且把「低分辨率做出好强度预测」这一反常识结论以论文形式固定下来。\n\n这也意味着「高分辨率等于好强度预测」这个十年来的气象学基本假设,开始被一条 AI 路径撬动。DeepMind 自己在博客和论文里都留了显式 caveat:研究者仍不完全理解低分辨率输入里被学到了什么信号,这个黑箱需要后续研究去拆。\n\n## 一些数字校准\n\n- 模型输入分辨率:**0.25°(约 28 km)**,论文和 README 双方一致;比传统区域模式粗约 100 倍的数字由 DeepMind 博客与 Ars Technica 共同复述。\n- 集合规模:**1000 成员**(2025 飓风季),2024 年及之前为 50 成员,数字来自 DeepMind 8 月 6 日博客。\n- 历史气旋训练数据:**IBTrACS 约 5000 个**(DeepMind 8 月 6 日博客)。\n- 业务化窗口:**WeatherNext 首次业务化进入 NHC 工作流**是 2025 飓风季,5 天前 80% 置信度命中「牙买加五级」由 NHC 主任 Mike Brennan 在 Ars Technica 8 月 8 日报道中口述。\n- 论文发表日期:2026 年 8 月 6 日,DOI 10.1038\u002Fs41586-026-10953-2(Nature)。\n\n## 所以呢\n\n对从业者:这意味着「高分辨率 = 更好」这一默认假设被论文级证据松动——未来做天气 AI 训练,算力预算可以重新分配,「先在 1° 上跑通,再决定要不要叠 0.25°」是更便宜的研究路径。仓库里同时放出 Mini 版本,把研究门槛拉到了单 TPU\u002FColab 级别。\n\n对气象局业务端:NHC 已经在 2025 飓风季的实战里把 1000 成员集合、5 天提前量拿到手,论文和开源意味着其它气象机构可以直接 fork 同一条 AI 路径,而不是从零训练。\n\n对关心通用 AI 进展的读者:这又是一个「基础研究产物反过来改写业务关键流程」的例子——和微软 tokenmaxxing 之后的 valuemaxxing、OpenJDK 禁 LLM 生成代码的政策转向属于同一个时代信号:AI 既是工具,也在变成新的依赖项,它的边界和失效模式需要比过去更早地被审计。\n\n**主参考资料**\n- Nature 论文 *Operational Tropical Cyclone Forecasting with AI*(Alet et al., 2026-08-06):[https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-026-10953-2](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-026-10953-2)\n- DeepMind 8 月 6 日官方博客:WeatherNext AI model achieves breakthrough in forecasting cyclones —— [https:\u002F\u002Fdeepmind.google\u002Fblog\u002Fweathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones](https:\u002F\u002Fdeepmind.google\u002Fblog\u002Fweathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones)\n- Ars Technica 8 月 8 日报道 *DeepMind's hurricane breakthrough has surprised weather scientists*:[https:\u002F\u002Farstechnica.com\u002Fscience\u002F2026\u002F08\u002Fdeepminds-hurricane-model-bought-forecasters-an-extra-day\u002F](https:\u002F\u002Farstechnica.com\u002Fscience\u002F2026\u002F08\u002Fdeepminds-hurricane-model-bought-forecasters-an-extra-day\u002F)\n- GitHub 开源仓库 `google-deepmind\u002Fweathernext`:[https:\u002F\u002Fgithub.com\u002Fgoogle-deepmind\u002Fweathernext](https:\u002F\u002Fgithub.com\u002Fgoogle-deepmind\u002Fweathernext)\n","https:\u002F\u002Fdeepmind.google\u002Fblog\u002Fweathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones\u002F","35ce748f-48b7-4638-88ef-effa57a7e749",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"677acb66-cb2a-4d48-8c02-9dee4acec3c2","en","DeepMind WeatherNext in Nature: an extra day of hurricane warning","DeepMind and Google Research published \"Operational Tropical Cyclone Forecasting with AI\" in Nature and simultaneously open-sourced the WeatherNext Cyclones model on GitHub. According to the paper, this 0.25° (~28 km) resolution AI model adds an average of more than a full day of lead time for hurricane track, intensity, and wind structure forecasts over 2023–2025, equivalent to roughly a decade of operational progress. The model runs a 15-day forecast in under a minute on a TPU and can generate 1,000-member ensembles, expanding extreme-event coverage by 20× over the conventional 50-member baseline.","## One-line takeaway\n\nGoogle DeepMind and Google Research published WeatherNext Cyclones (WN-C) in Nature, open-sourcing the model and weights. According to the same paper, this AI weather model uses inputs roughly 100× coarser than traditional regional models and still adds an average of more than one day of lead time on tropical cyclone track, intensity, and wind-structure forecasts.\n\n## Background: why hurricanes are hard to forecast\n\nTropical cyclones are among the most dangerous and costly weather phenomena. The DeepMind launch blog cites more than 700,000 deaths and $1.4 trillion of economic losses globally over the past 50 years. Hurricane forecasting has long been split between two regimes: tracks depend on the global atmospheric circulation, where coarse global models excel; intensity is driven by highly localized thermodynamic processes around the cyclone core, where higher-resolution regional models excel. The two have to be stitched together by human forecasters.\n\n## The WeatherNext Cyclones approach\n\nWN-C is a single end-to-end AI model that outputs track, intensity, and wind structure jointly. Key design points (per the Nature paper, the DeepMind launch blog, and the open-source repository):\n\n- Training data: ~20 TB of global atmospheric reanalysis (built on ECMWF) plus the IBTrACS historical cyclone database of nearly 5,000 storms.\n- Architecture: built on Functional Generative Networks (FGN, arXiv:2506.10772) for ensemble generation. A single 15-day forecast finishes in under a minute on a TPU, with up to 1,000 physically consistent members (vs. the conventional 50).\n- Input resolution: only 0.25° (~28 km), roughly 100× coarser than traditional regional models. A 1° (~111 km) Mini version is also released, runnable on a single TPU through Colab.\n- Counter-intuitive result: the authors explicitly call out that achieving state-of-the-art intensity forecasts from such coarse inputs is itself an open research question — they are not fully sure what signal the low-resolution inputs encode.\n\n## Empirical performance (Nature + Ars Technica)\n\nThe Nature paper's abstract headline number: across 2023–2025 retrospective evaluation, WN-C delivers an average lead-time advantage of more than a full day for track, intensity, and wind radii, \"comparable to the progress seen over the last decade of operational development.\"\n\nFiner-grained numbers from the DeepMind launch blog (2026-08-06) and Ars Technica (2026-08-08):\n\n- 2025 live deployment — the U.S. National Hurricane Center used WeatherNext operationally and, for the first time, predicted a Category 5 landfall in Jamaica while Hurricane Melissa was still a Category 1. The blog quotes a 5-day, 80% confidence prediction that the system would hit Jamaica as a Category 5.\n- Ensemble scale-up — single-storm ensembles went from 50 members in 2024 to 1,000 in 2025, with the goal of capturing rare rapid-intensification events.\n- Open-sourcing — code, weights, WN-C, and WeatherNext 2 are all released together in the `google-deepmind\u002Fweathernext` GitHub repository, with operational data feeds pushed through Google Cloud (Vertex AI \u002F BigQuery \u002F Earth Engine), the Weather Lab visualization page, and Open-Meteo.\n\n## Where this sits relative to prior AI weather models\n\nWeatherNext itself is not new. DeepMind's GraphCast (2023) and GenCast (December 2024) already pushed AI weather forecasting into ECMWF-class territory. The change here is that WeatherNext Cyclones is the first to complete a full loop: Nature peer review + NHC operational integration in 2025 + public release. It also locks in a counter-intuitive empirical conclusion — high resolution is not a strict prerequisite for state-of-the-art intensity forecasting — which begins to displace a long-standing default assumption.\n\n## Calibration of the key numbers\n\n- Input resolution: 0.25° (~28 km). Confirmed by the Nature paper, DeepMind blog, and GitHub README; the ~100× coarseness figure is restated by both the DeepMind blog and Ars Technica.\n- Ensemble size: 1,000 members in 2025 (50 in prior years), per the DeepMind blog.\n- Historical cyclone training data: ~5,000 storms in IBTrACS, per the DeepMind blog.\n- Operational milestone: first NHC operational use of WeatherNext was in the 2025 Atlantic hurricane season, with the 5-day, 80% Category 5 prediction for Jamaica attributed to NHC director Mike Brennan in Ars Technica's 2026-08-08 report.\n- Paper publication: 2026-08-06, DOI 10.1038\u002Fs41586-026-10953-2 (Nature).\n\n## What this means in practice\n\nFor AI researchers: the \"high resolution = better\" default is now loosened by peer-reviewed evidence. Compute budgets for weather AI training can be reallocated; running a 1° model first and only scaling to 0.25° after proof-of-concept is now a cheaper research path. The Mini release lowers the entry cost to a single TPU on Colab.\n\nFor national weather services: the path used by NHC in 2025 is now open-sourced. Other meteorological agencies can fork the same architecture and training recipe rather than starting from scratch.\n\nFor readers watching general AI progress: this is the latest example of a foundational research result rewriting a mission-critical operational pipeline. It belongs to the same era signal as Microsoft's shift away from tokenmaxxing and Oracle's OpenJDK ban on AI-generated code — AI is simultaneously a tool and a new dependency, and its failure modes need to be audited earlier than its adoption curve.\n\n**Primary references**\n- Nature paper *Operational Tropical Cyclone Forecasting with AI* (Alet et al., 2026-08-06): [https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-026-10953-2](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-026-10953-2)\n- DeepMind launch blog (2026-08-06): WeatherNext AI model achieves breakthrough in forecasting cyclones — [https:\u002F\u002Fdeepmind.google\u002Fblog\u002Fweathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones\u002F](https:\u002F\u002Fdeepmind.google\u002Fblog\u002Fweathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones\u002F)\n- Ars Technica (2026-08-08): *DeepMind's hurricane breakthrough has surprised weather scientists* — [https:\u002F\u002Farstechnica.com\u002Fscience\u002F2026\u002F08\u002Fdeepminds-hurricane-model-bought-forecasters-an-extra-day\u002F](https:\u002F\u002Farstechnica.com\u002Fscience\u002F2026\u002F08\u002Fdeepminds-hurricane-model-bought-forecasters-an-extra-day\u002F)\n- Open-source repository `google-deepmind\u002Fweathernext` on GitHub: [https:\u002F\u002Fgithub.com\u002Fgoogle-deepmind\u002Fweathernext](https:\u002F\u002Fgithub.com\u002Fgoogle-deepmind\u002Fweathernext)\n","deepmind-weathernext-cyclones-nature-open-source","2026-08-10T02:00:00Z","2026-08-10T02:36:21.036308Z","2026-08-10T02:36:21.036319Z",true,"agent",200,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"777afb24-262f-45cc-961f-d5d49ad42883","AgentOPSD 用递归贝叶斯信念破解多轮 Agent 强化学习的信用分配：清华\u002F浙大\u002F美团让 GRPO 学会看哪个 turn 决定胜负","agentopsd-recursive-belief-credit-assignment","2026-08-07T02:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"daf9222a-aebc-4f09-92f8-74b6226fd5f1","把图灵测试变成 RL 损失：MIT 提出 Turing-RL，让用户模拟器更\"像人\"","turing-rl-mit-stanford-discriminative-reward","2026-06-18T02:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"0d8fdf45-4585-47c0-9e78-3652e318b156","Apple Intelligence 中国版落地:通义千问接管语言 AI,百度负责视觉搜索","apple-intelligence-china-qwen-baidu-2026","2026-08-25T12:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"1311adb6-dc19-41a7-a188-6760d9e53672","HF Summer 2026 报告:13 个下载量 Top 25 模型是 2022 年的老面孔","hugging-face-summer-2026-attention-adoption","2026-08-24T08:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"f9cf9f03-6aca-4d29-94d3-5c6acfeaf435","匿名模型 OX Alpha 短暂登顶 OpenRouter 编码榜:研究者推测底座指向智谱 GLM-5.x","ox-alpha-stealth-openrouter-glm-5-zhipu","2026-08-24T03:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"1844afb1-3a1c-4acd-9e4c-f5e2792a2018","下载免费不等于商用免费：HF Summer 2026 隐藏的开源前沿许可证分水岭","frontier-license-shift-hf-summer-2026","2026-08-23T12:30:00+00:00"]