据美国国家公共广播电台（NPR）10月26日报道，飓风预报员正越来越多地借助人工智能进行预测，并称这项技术正在迅速改变整个领域。本月大西洋飓风季首个飓风生成时，一个人工智能模型的路径预测就压过了传统模型。
10月初，飓风“伊萨亚斯”（Isaias）以二级飓风强度在美国佛罗里达州德斯廷附近登陆。常规预报模型最初判断它将继续向西，扑向路易斯安那州东南部。而GoogleDeepMind的人工智能模型从一开始就判断正确：这场飓风将袭击佛罗里达州狭长地带的西部。
“尤其是Google DeepMind，过去几年确实证明了自己是黄金标准模型。”迈阿密资深气象学家、飓风专家迈克尔·劳里说，他曾为美国联邦应急管理署和国家飓风中心提供咨询。“那个模型是我们见过表现最好的计算机模型，遥遥领先。就是这样。”
美国国家飓风中心从去年开始首次将Google DeepMind模型纳入官方预报。迈阿密大学飓风研究员布赖恩·麦克诺尔迪说，到2025年大西洋飓风季结束时，它已经是预测风暴路径和强度最好的模型之一。通常新模型在引入后需要数年打磨。“这是我们不习惯的情况——一个初来乍到的新模型很快就成了最好的模型。”麦克诺尔迪说，“它几乎一登场就排名第一，这是相当了不起的成就。”
传统的基于物理规律的飓风模型，依赖卫星和飞机实时测量的大气数据，如温度、湿度、气压和风切变，再用复杂方程推算风暴可能的路径和强度。人工智能模型的运作方式截然不同：它更依赖历史，通过分析数千场过去的飓风及伴随的大气数据来识别规律。
“AI模型除了当前这张图景之外，其实对大气一无所知。”劳里说，“除此之外，它并不依靠大气运动的主导方程、也就是物理规律来做预测。”劳里指出，设计和开发这些模型的大多数人甚至不是气象学家：“他们是软件工程师，在做神经网络的开发。”所谓“神经网络”，是一组模仿人脑、经过训练能在海量数据中寻找规律的计算机算法。
由此，这些模型看起来像一个“黑箱”。劳里说：“我们不知道它们为什么能预报正确，也不知道它们为什么会预报错误。”对气象学家而言，依赖一个自己并不完全理解的系统会令人不安，尤其是在数百万人可能正仰赖其预报的风暴期间。“这让我们有点不舒服，因为我们想知道你为什么能得出正确答案。”麦克诺尔迪说，“而我觉得我们不知道——至少我不知道。”
不过，并非所有AI模型都表现良好。劳里说，市面上有大量未经检验的天气预报AI模型。目前被证明最准确的两个，尽管各有缺陷，是Google DeepMind和总部位于英国的欧洲中期天气预报中心开发的模型。
在预报“伊萨亚斯”的路径上，Google DeepMind的AI模型击败了传统的物理模型，但在预测风暴强度方面做得并不好。美联社10月10日拍摄的照片显示，在“伊萨亚斯”过后的德斯廷，人们从坍塌的船屋中取回物品。
国家飓风中心的预报员在制作官方预报时会使用多个经过验证的模型，既有人工智能模型也有传统模型。“每场风暴都有不同的预报难题。”国家飓风中心主任迈克尔·布伦南说，“这就是我们采取集合式方法、混合使用最佳模型的原因。”而且，仍须由人来解读这些模型的结果，再决定告知公众什么。飓风活跃期间，国家飓风中心每六小时发布一次更新。模型本身有时会在两次更新之间出现剧烈变化，预报员必须把这些摇摆的预测转化为人们可以据以行动的清晰指引。
“人类预报员提供的一大价值就是连续性。”布伦南说，“我们用自己的专业知识和理解，把飓风预报转化为灾害信息，这才是人们真正需要关注的东西。”
科学家预计相关技术会快速进步。布伦南说，Google DeepMind模型仍需改进。它在追踪“伊萨亚斯”路径上胜过其他模型，但在判断风暴强度方面不如国家飓风中心的混合预报。布伦南团队定期向Google DeepMind提供反馈以帮助改进模型。Google未回应NPR就此事的提问。
布伦南说，预报的“未来一直在变化”，“从来不是某一样东西在推动这些进步”。他指出，不久前预报员还无法在两天之内预测风暴会有多强；这一局面因预测模型的进步以及卫星和飞机采集到更好的数据而改变。“就强度预报而言，误差大约只有‘卡特里娜’飓风时期的一半。”布伦南说。
气象学家劳里认为，AI令人兴奋之处在于改进速度。AI模型的更新只需数月甚至数周，而传统模型最多一年更新一次。“它们在一年之内取得的技巧和飓风预测进步，是我们过去需要10年、甚至十几年才能做到的。”劳里说。
但即便AI来势迅猛，劳里说他并不担心人工智能会取代人类。“对我而言，我没有因为AI而产生某种生存危机。”劳里说，“如果要说的话，我认为它是我们这个行业的福音。”
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
According to the United States National Public Radio (NPR) on 26 October, hurricane forecasters are increasingly relying on artificial intelligence to predict and claim that this technology is rapidly changing the whole area. At the time of the first hurricane in the Atlantic hurricane season this month, the routing prediction of an artificially intelligent model overwhelmed the traditional model.
Early in October, Hurricane Isaias landed near Destin, Florida, United States, with a secondary hurricane intensity. The conventional forecasting model initially determined that it would continue westward towards south-east Louisiana. And Google DeepMind's artificial intelligence model was right from the beginning: the hurricane would hit the west of the narrow edge of Florida.
“In particular, Google DeepMind has proved itself to be a gold standard model over the past few years.” Michael Laurie, a senior meteorologist and hurricane expert in Miami, said he had advised the United States Federal Emergency Management Agency and the National Hurricane Center. "That model is the best computer model we've ever seen, far ahead. That's it."
For the first time since last year, the United States National Hurricane Center has incorporated the Google DeepMind model into official forecasts. Brian McNordi, a hurricane researcher at the University of Miami, said that by 2025, At the end of the wind season, it is already one of the best models for predicting the path and intensity of storms. Usually, new models take years to grind after they're introduced. "This is something that we're not used to -- a new model that's just beginning to be the best one soon." McNoldi said, “It is a remarkable achievement that it ranks first almost as soon as it arrives.”
Traditional models of hurricanes based on physical patterns rely on atmospheric data measured in real time by satellites and aircraft, such as temperature, humidity, air pressure and wind cutting, and use complex equations to extrapolate the path and strength of a storm. Artificial intelligence models work differently: It relies more on history to identify patterns by analysing thousands of past hurricanes and accompanying atmospheric data.
"The AI model has no knowledge of the atmosphere other than the current picture." According to Laurie, “other than that, it does not rely on the dominant equation of atmospheric movements, i.e. the laws of physics”. Laurie points out that most of the design and development of these models are not even meteorologists: “They're software engineers, they're developing a neural network.” The so-called neural network is a set of computer algorithms that mimic the human brain and are trained to find regularity in big data.
So these models look like a "black box." LAURIE said, "We don't know why they forecast right, nor why they forecast wrong." For meteorologists, it would be disturbing to rely on a system that they do not fully understand, especially during a storm on which millions of people may depend. “It makes us a little uncomfortable, because we want to know why you got the right answer.” McNoldi says, “and I think we don't know — at least I don't know.”
However, not all AI models performed well. Laurie states that there are a large number of untested weather forecasting AI models on the market. The two most accurate, albeit flawed, models developed by Google DeepMind and the European Medium-Term Weather Forecast Centre, based in the United Kingdom.
The Google DeepMind AI model defeated traditional physical models on the path to predict Isaias, but did not do well to predict storm intensity. Photo taken by the Associated Press on 10 October shows that, after Isaias, Destin was recovered from the collapsed boathouse.
Forecasters at the National Hurricane Centre use multiple validated models, both artificial and traditional, in the production of official forecasts. "There are different forecasting difficulties for every storm." Michael Brennan, Director of the National Hurricane Center, said, “This is why we have adopted a grouping approach, combining the best models”. Moreover, the results of these models must still be read in person before deciding what to inform the public. The National Hurricane Center was updated every six hours during the hurricane ' s active period by publish. The models themselves sometimes undergo dramatic changes between updates, and forecasters must translate these swing predictions into clear guidelines by which people can act.
“One of the major values offered by human forecasters is continuity.” Brennan said, “It is with our own expertise and understanding that we translate hurricane forecasts into disaster information that people really need attention.”
Scientists expect rapid technological advancement. Brennan stated that the Google DeepMind model still needed improvement. It is better than other models in tracking the Isaias path, but less than a mixed forecast of the national hurricane centre in determining the intensity of the storm. The Brennan team regularly provided feedback to Google DeepMind to help improve the model. Google did not respond to NPR's questions on this matter.
Brennan says that the forecast “the future has been changing” and that “it has never been something that drives these advances”. He pointed out that not long ago the forecaster had been unable to predict how strong the storm would be in two days; the situation had changed as a result of advances in predictive models and better data collected by satellites and aircraft. “In terms of intensity forecasts, the error was only about half that of Hurricane Katrina.” Brennan said.
In the opinion of the meteorologist Laurie, what's exciting about AI is the speed improvement. The update of the AI model takes months or even weeks, while the traditional model is updated at most once a year. “The skills they have gained within a year and the progress they have made in the prediction of hurricanes are what we have been able to do for the past 10 years or more.” Laurie said.
But even though AI is fast, Laurie says he's not worried about artificial intelligence replacing humans. “As far as I'm concerned, I'm not in some kind of existential crisis because of AI.” Laurie said, "If anything, I think it's the gospel of our industry."
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.