Future imperfect
外刊原文
Artificial intelligence is helping improve climate models
More accurate predictions will lead to better policy-making
THE DIPLOMATIC ructions at COP29, the United Nations climate conference currently under way in the Azerbaijani capital of Baku, are based largely on computer models. Some model what climate change might look like; others the cost of mitigating it.
No model is perfect. Those modelling climate trends and impacts are forced to exclude many things, either because the underlying scientific processes are not yet understood or because representing them is too computationally costly. This results in significant uncertainty in the results of simulations, which comes with real-world consequences. Delegates’ main fight in Baku, for example, will be over how much money poor countries should be given to help them decarbonise, adapt or recover. The amount needed for adaptation and recovery depends on factors such as sea-level rise and seasonal variation that climate modellers still struggle to predict with much certainty. As negotiations become ever more specific, more accurate projections will be increasingly important.
The models that carry most weight in such discussions are those run as part of the Coupled Model Intercomparison Project (CMIP), an initiative which co-ordinates over 100 models produced by roughly 50 teams of climate scientists from around the world. All of them attempt to tackle the problem in the same way: splitting up the world and its atmosphere into a grid of cells, before using equations representing physical processes to estimate what the conditions in each cell might be and how they might change over time.
When CMIP started in 1995, most models used cells that were hundreds of kilometres wide—meaning they could make useful predictions about what might happen to a continent, but not necessarily to individual countries. Halving the size of cells requires roughly ten times more computing power; today’s models, thousands of times more powerful, can simulate cells of around 50km per side.
Clever computational tricks can make them more detailed still. They have also grown better at representing the elaborate interactions at play between the atmosphere, oceans and land—such as how heat flows through ocean eddies or how soil moisture changes alongside temperature. But many of the most complex systems remain elusive.
Clouds, for example, pose a serious problem, both because they are too small to be captured in 50km cells and because even small changes in their behaviour can lead to big differences in projected levels of warming.
Better data will help. But a more immediate way to improve the climate models is to use artificial intelligence (AI). Model-makers in this field have begun asserting boldly that they will soon be able to overcome some of the resolution and data problems faced by conventional climate models and get results more quickly, too.
Engineers from Google have been among the most bullish. NeuralGCM, the company’s leading AI weather and climate model, has been trained on 40 years of weather data and has already proved itself to be as good at forecasting the weather as the models for and by which these data were originally compiled. In a paper published in Nature in July, Google claimed its model will soon be able to make projections over longer timescales faster, and using less power, than existing climate models. With additional training, the researchers also reckon NeuralGCM will be able to offer more certainty in important areas like shifts in monsoons and tropical cyclones.
This optimism, say the researchers, comes from the unique abilities of machine-learning tools. Where existing models sidestep intractable physics problems by using approximation, NeuralGCM’s creators claim it can be guided by spotting patterns in historical data and observations. These claims sound impressive, but are yet to be evaluated. In a preprint posted online in October, a team of modellers from the Lawrence Livermore National Laboratory in California noted that NeuralGCM will remain limited until it incorporates more of the physics at play on land.
Others are more sceptical that AI methods used in short-term weather forecasting can be successfully applied to the climate. “Weather and climate are both based on physics,” says Gavin Schmidt, a climate scientist who runs NASA’s Goddard Institute for Space Studies, but pose different modelling challenges. For one thing, the available data are rarely of the same quality. For weather forecasting, huge swathes of excellent data are generated every day and, therefore, able to continuously validate the previous day’s predictions. Climate models do not enjoy the same luxury. In addition, they face the challenge of simulating conditions more extreme than any previously observed, and over centuries rather than days.
AI can nonetheless help improve climate models by addressing another major source of uncertainty: human behaviour. Until now, this has been overcome by codifying different social and political choices into sets of fixed scenarios which can each then be modelled. This method makes evaluations possible, but is inflexible and often vague. With the help of AI, existing tools known as emulators can customise conventional models to suit their end users’ needs. Such emulators are now used by cities planning infrastructure projects, by insurers assessing risk and by agricultural firms estimating changes in crop yields.
Unlike models such as Google’s NeuralGCM, which is trained on the same weather data as today’s top climate models, emulators are typically trained on the outputs of full-scale climate models. This allows them to piggyback on improvements to the models themselves—both the new physics they are able to model and the ways in which they extrapolate beyond historical data. One such emulator, developed by the Commonwealth Scientific Industrial Research Organisation in Australia in 2023, for example, was capable of adjusting predictions linked to future emissions levels one million times faster than the model it was trained on.
Reducing the uncertainties in climate models and, perhaps more important, making them more widely available, will hone their usefulness for those tasked with the complex challenge of dealing with climate change. And that will, hopefully, mean a better response.
参考译文
并不完美的未来
人工智能助力气候模型升级
精准预测为决策护航
当下，在阿塞拜疆首都巴库召开的 COP29 气候大会上，各国外交官们正就一系列基于计算机模型的预测展开激辩。这些模型或是描绘气候变化的可能走向，或是评估减排措施的成本效益。
然而，世上没有完美的模型。目前的气候模型由于科学认知的局限性和计算资源的制约，不得不舍弃诸多变量。这种简化inevitably带来预测结果的不确定性，进而影响现实决策。以巴库会议为例，与会各方最大的分歧在于应向发展中国家提供多少气候援助资金。而这笔资金的具体数额又取决于海平面上升和季节变化等难以精准预测的因素。随着气候谈判日趋具体，提高预测准确度的需求愈发迫切。
在众多气候模型中，最具权威性的当属耦合模型比较计划（CMIP）。该计划汇集了全球50余个研究团队开发的百余个模型。这些模型采用相同的方法论：将地球及其大气层划分为网格状单元，再用物理方程计算每个单元的状态及其演变。
自1995年 CMIP 启动以来，模型精度突飞猛进。早期模型的网格单元动辄数百公里，只能对大陆尺度的变化做出粗略预测。如今，得益于计算能力的千倍提升，模型可以将单元缩小至50公里见方。
通过巧妙的计算技术，模型的精度还可进一步提高。它们在模拟大气、海洋、陆地间的复杂互动方面也取得长足进步，比如海洋涡流的热传递、土壤湿度随温度的变化等。但仍有许多复杂系统难以准确描述。
以云层为例，就给模型制作者出了一道难题。不仅因为50公里的网格难以捕捉云的微观变化，更因为云层行为的细微差异可能导致温度预测产生巨大偏差。
更丰富的数据固然有助于提高模型精度，但人工智能（AI）的引入则开辟了一条更直接的进化路径。模型研究者们信心满满地表示，AI 技术不仅能突破传统模型在分辨率和数据处理上的瓶颈，还能大幅提升计算效率。
谷歌工程团队对此尤为乐观。他们开发的 AI 气象气候模型 NeuralGCM 已在40年气象数据的训练中，展现出与传统模型相媲美的预测能力。今年7月发表在《自然》杂志上的论文显示，该模型有望以更低的能耗实现更长期的气候预测。研究人员还表示，通过深入训练，NeuralGCM 有望在季风变迁、热带气旋等关键领域提供更可靠的预测。
研究团队的信心源自机器学习的独特优势。传统模型往往需要通过简化来规避复杂的物理问题，而 NeuralGCM 则可以通过挖掘历史数据中的规律来寻找突破。不过，这些雄心勃勃的主张仍待验证。加州劳伦斯利弗莫尔国家实验室的研究人员在10月的预印本中指出，在完善陆地物理过程的模拟之前，NeuralGCM 的潜力仍将受限。
对于短期天气预报中的 AI 方法能否成功迁移到气候预测领域，学界也存在分歧。NASA 戈达德空间研究所所长 Gavin Schmidt 指出，虽然天气和气候都遵循物理定律，但它们在建模上各有挑战。天气预报可以依靠每日海量的高质量数据不断自我验证，而气候模型则需要模拟跨越数世纪、超出历史记录的极端情况，难度显著提高。
尽管如此，AI 在应对另一个重要变量——人类行为方面展现出独特优势。传统方法是将社会政治选择编码为固定场景进行建模，虽然可行但缺乏灵活性。借助 AI，现有的模拟器可以根据用户需求灵活调整模型。这些工具已在城市规划、保险评估、农业生产等领域发挥重要作用。
与直接在气象数据上训练的 NeuralGCM 不同，这类模拟器通常基于成熟气候模型的输出结果进行训练，可以充分借鉴现有模型在物理模拟和数据外推方面的优势。澳大利亚联邦科学工业研究组织在2023年开发的模拟器就实现了比原模型快100万倍的预测调整能力。
随着气候模型的不确定性逐步降低，其应用范围不断扩大，人类应对气候变化的能力必将显著增强。在这场关乎全人类命运的挑战中，科技创新正在为我们开辟新的希望之路。
感谢大家阅读，Enjoy!
来源：20241116-The Economist 《经济学人》最新一期
原文标题：Artificial intelligence is helping improve climate models
* 参考译文为AI辅助翻译。