外刊 | The high-tech race to improve weather forecasting, Part 4

The high-tech race to improve weather forecasting, Part 4

    这篇文章有点长，我把它分成了考研阅读的长度（600-700words）。

    第一部分介绍了欧洲中期天气预报中心的功能、极端天气与天气预报的重要性。在结尾用1960s至今预测准确率的提升引出第二部分。

    第二部分介绍了天气预测的原理及预测模型计算问题。

    第三部分介绍私人预测机构的优势与功能。

    最后一部分介绍了AI时代气象预测领域的变局。

Sunny with a chance of AI  天晴，可能有AI

"Sunny with a chance of showers"，即“天晴可能下雨”的变体

1.机器学习简化天气预测

Private companies have also been at the forefront of attempts to find new, less computationally onerous ways of predicting the weather. Many are focusing on machine learning, a type of artificial intelligence (AI) that looks for patterns in big piles of data.

Salient, an American startup, uses anAI trained to recognise patterns in historical data to produce forecasts on a seasonal scale, rather than over days or weeks. 

Its customers include Zurich Insurance Group, which hopes to get early warnings of extreme weather its clients might face.

machine learning 机器学习

computationally 计算上

onerous 繁重的

a startup 创业公司

2.AI可以看出研究者忽略的模式

AI can spot patterns that human researchers may have missed. Ray Schmitt, a researcher at the Woods Hole Oceanographic Institution in Massachusetts, is one of Salient’s founders. 

He had theorised about a link between ocean salinity around the east coast of America in spring and rainfall across the Midwest the following summer. AI analysis of weather data seems to confirm the connection, though the precise mechanism remains unclear.

spot 看出

salinity 盐度

3.AI预测不必了解气象原理

That illustrates another intriguing feature ofai-based forecasts. Numerical simulations rely on their programmers having a good understanding of the physical processes that drive the weather. 

But using an AI to spot recurring patterns can help useful forecasts be produced even before the underlying science is fully understood.

4.现有计算模型的不足及AI的优势

Machine learning has already proved its worth with precipitation “nowcasting”—predicting whether it will rain or snow in a given area over the next few hours. 

But predicting them can be tricky for existing numerical models, partly because, by the time they have finished running, the moment has often passed.AI pattern recognition requires less computational grunt, allowing it to make forecasts more quickly.

computational grunt 计算能力
grunt 野猪的呼噜声

5.多方合作AI气象项目

A 2021 collaboration between DeepMind, a part of Google, and the Met Office in Britain usedAI to forecast precipitation based on observations from rain-detecting radar. 

The AI system outperformed existing, numerical forecasting methods nine times out of ten—though it started to stumble when asked to forecast beyond about 90 minutes.

precipitation 降雨量

stumble 出错

6.华为盘古气象大模型

Other big firms with AI expertise are getting involved, too. A paper published in Nature on July 5th described Pangu-Weather, an AI system built by Huawei, a Chinese firm, and trained on 39 years of weather data.

Huawei claims Pangu-Weather can produce week-ahead predictions comparable in accuracy to forecasts from outfits like ECMWF, but thousands of times faster. 

Last year Nvidia, an American chipmaker, claimed that FourCastNet, its AI weather program, could generate, in two seconds, a forecast that can predict hurricanes and heavy rain up to a week in advance.

7.政府官员改变看法

Governmental incumbents are coming around. The ECMWF was surprised by the results of Pangu-Weather, says Florence Rabier, the organisation’s director-general.

“We did see a lot of potential, and they are not exaggerating the claims that it is much cheaper [to run],” she says. The ECMWF is now working with Huawei, as well as with Google and Nvidia.

governmental incumbent 政府现任官员

be coming around 改变态度或想法

director-general 总监

8.AI不能取代数值预测

That does not mean that AI will replace numerical forecasting, though it could help it become more efficient. AI relies crucially on high-quality data onwhich to train models. 

Since many parts of the world lack reliable data from weather stations, old-fashioned numerical simulations must be used retrospectively to fill in the gaps. 

And just as computational approaches face fundamental limits to their utility, so too do AI-based ones. History is a less reliable guide to the future in a world whose weather is being fundamentally altered by climate change.

（1）train the models on high-quality data，这里的on是根据的意思。除此之外，“train A on B”还可以表示瞄准。比如：The security guards trained their flashlights on the suspicious figure lurking in the shadows. 保安们用手电筒照向在阴影中潜伏的可疑人物。

（2）face fundamental limits to，介词to易考完形填空。

（3）“And just as computational approaches face fundamental limits to their utility, so too do AI-based ones.”这是两个并列的主语从句，表示前后两者地位相似。可以理解为，计算方法与AI一样要面临功能的基础限制。computational approaches指的是传统的数值预测（numerical forecast）计算模型。

用“just as A, so too B”造句：

①劳逸结合：“Just as hard work is crucial for achieving success, so too is relaxation.”

②物质与精神生活的同样重要："Just as material comfortplays a significant role in our lives, so too does spiritual well-being."

9.公私合作项目

More public-private collaboration is on the cards. By 2030, the European Commission hopes to have finished “Destination Earth”, a simulation that can handle both short-term weather patterns and longer-term changes in the climate. 

It hopes that users, with the help ofAI, will be able to visualise how animal migration patterns might change as temperatures rise, or what might happen to fish stocks as the oceans warm. Nvidia, whose chips power most of the world’s biggest AI models, has said it will participate.

The firm has also signed up to an even more ambitious plan for a network of “Earth Virtualisation Engines” proposed at a meeting this month in Berlin by a group led by Bjorn Stevens, the director of the Max Planck Institute for Meteorology in Hamburg.

be on the cards 有可能发生

migration patterns 迁徙模式

fish stocks 渔业资源

“power”是动词，可以翻译为驱动

virtualisation engines 虚拟引擎

10.人工智能技术给气象领域带来新变化

Dr Stevens sees all this ferment as part of a shift in how information about the weather is conceived of, produced and used. Turning observationsinto something helpful like a forecast used to require a lot of expert knowledge, he says. 

That made it the domain of a handful of big institutions. But recent technological advances, especiallyAI, have made doing that both easier and cheaper.

“That makes [weather] data valuable,” he says. “And that is transforming everything.”

ferment 骚动

be conceived of 设想

observation 观察数据

domain 领域

a handful of 少量的

used to do 表示过去一直存在的情况，可翻译为“过去总是……”

欢迎指正，感谢阅读。