The latest report from the World Meteorological Organization: AI is revolutionizing weather forecasting, making it faster, cheaper and more accessible

“We are still very far from meeting global climate goals.”

“2023 was the warmest year on record… The first eight months of 2024 were also the warmest on record.”

said Celeste Saulo, Secretary-General of the World Meteorological Organization (WMO).

There is an 86% chance that at least one of the next five years will surpass 2023 and become the hottest year on record. This conclusion comes from the recently released WMO report "United in Science 2024".

At the same time, the report points out that in the next five years, there is an 80% chance that the global average near-surface temperature will temporarily exceed the pre-industrial level by 1.5°C in at least one year.

In addition, according to the WMO's previously released "Global Climate Status 2023" report, the pain and chaos caused by heat waves, floods, droughts, wildfires and rapidly intensifying tropical cyclones have disrupted the daily lives of millions of people and caused billions of dollars in economic losses.

In order to cope with the climate crisis, it is particularly important to create a fast, economical and accurate weather warning system. Artificial intelligence (AI) is bringing disruptive changes to the field of weather warning with its complex algorithms and powerful computing power.

As Saulo points out, “AI has revolutionized the science of weather forecasting by making it ‘faster, cheaper and more accessible’”.

AI is revolutionizing weather forecasting

In this report, a research team from the European Centre for Medium-Range Weather Forecasts (ECMWF) and its collaborators provide a comprehensive account of the latest applications of AI in extreme weather warning, as well as existing shortcomings and future prospects.

Specifically, in the field of weather forecasting, AI models have broken through the numerical weather forecast (NWP) model based on physical models and have surpassed physical models in predicting certain weather variables and extreme or dangerous events (such as tropical cyclones). For example, research by scholars such as Keisler and Pathak has demonstrated the significant advantages of AI models in cyclone forecasting.

Figure | The latest progress in artificial intelligence/integrated forecasting systems (AIFS) has improved cyclone detection capabilities. (Source: The report)

By using the Fourier forecast neural network, an emerging global data-driven weather forecast AI model, it has completely achieved accurate prediction of high-resolution, fast-time-scale variables such as surface wind speed, precipitation, and atmospheric water vapor. It can generate a week's forecast in less than 2 seconds, which is several orders of magnitude faster than IFS.

Studies have also shown that AI prediction models have broken through traditional physical models. By improving data quality, integrating different data sources, and processing prediction outputs at downscale, they have significantly reduced the computational cost of creating data required to support weather forecasts and the threshold for running high-quality prediction models.

These capabilities were previously limited to large global forecast centers due to computational burdens, but now they are available to institutions without sufficient resources. The entry threshold for running high-level forecast models has been significantly lowered, and the lower cost has also enabled smaller public and private entities to enter the field of weather forecasting with the help of AI, greatly changing the traditional landscape of the weather forecasting industry.

Other studies have pointed out that the integration of large language models (such as ChatGPT) can effectively promote the interpretation and communication of complex meteorological information, support the decision-making of Early Warnings for All, Sustainable Development Goals (SDGs), the Paris Agreement and other disaster reduction frameworks, and enhance disaster prevention, response and adaptation capabilities.

In theory, services such as ChatGPT, which is built on large language models (LLMs), are good at expanding the scope of climate information to any interested individual anywhere on the planet, providing local climate services to everyone. Although this idea has not been implemented so far due to the lack of detailed local information about future weather and climate change and its impacts, if this obstacle can be overcome, it will help people better cope with global climate challenges.

In recent years, with the rapid development of AI technology, many technology giants and research institutions such as Google, NVIDIA, and Huawei have made major breakthroughs in the field of weather forecasting and developed a series of impressive AI weather forecast products.

These products not only improve the accuracy and speed of weather forecasts, but also demonstrate unprecedented potential in key areas such as extreme weather forecasting.

Among them, the Pangu-Weather model developed by Huawei Cloud has become one of the world's most innovative achievements. The model was published in Nature magazine in July 2023. It was trained using 39 years of global reanalysis weather data. Its prediction accuracy is comparable to that of the world's top numerical weather forecast system IFS (European Centre for Medium-Range Weather Forecasts), but the prediction speed is more than 10,000 times faster than IFS at the same spatial resolution. This breakthrough shows the huge advantages of AI forecasting models in efficiency and cost.

Figure | Evolution of two-meter temperature error during 10-day forecasts in the Southern Hemisphere for different AI modeling systems (Huawei's Pangu-Weather, NVIDIA's FourCastNet, AIFS, and Google DeepMind's GraphCast) in 2022. (Source: European Center for Medium-Range Weather Forecasts, 2024)

Another study published in Nature introduced the NowcastNet model developed by a team led by Professor Michael Jordan of the University of California, Berkeley and Professor Wang Jianmin of Tsinghua University. NowcastNet combines physical laws with deep learning technology to enable real-time precipitation forecasts, significantly improving the accuracy of short-term forecasts (nowcasting). This method of combining physics and AI demonstrates the great potential of AI in capturing the dynamic process of rapid and subtle weather changes.

Figure ｜ NowcastNet design architecture. (Source: Nature)

In November of the same year, Google DeepMind launched another groundbreaking AI weather forecasting model, GraphCast. This model can predict hundreds of weather variables for the next 10 days within one minute at a global 0.25° resolution. Compared with traditional weather forecasting methods, GraphCast not only significantly improves forecasting efficiency, but also performs well in forecasting extreme weather events.

In addition, the AI ​​model developed by the Google Research team in early 2024 has also made significant progress in the field of global flood warning. The model beat the existing most advanced global flood warning system, trained with data from 5,680 measuring instruments, and can accurately predict daily runoff in unmeasured basins within a 7-day forecast period, providing a more powerful tool for addressing flood risks caused by climate change.

Figure ｜ LSTM-based river forecast model architecture. Two LSTMs are applied in sequence, one receiving historical weather data and the other receiving predicted weather data. The model output is the probability distribution parameters of the flow at each forecast time step.

Another important player in the field of AI weather forecasting is IBM, which developed the GRAF (Global High-Resolution Atmospheric Forecasting System) system, which uses AI and machine learning to update global weather every hour. GRAF is the world's first high-resolution weather forecasting system, capable of providing forecasts with a resolution of up to 3 kilometers, providing higher accuracy in predicting extreme weather events.

Microsoft's AI for Earth project is also using machine learning and big data to improve climate prediction and weather forecasts. Although it does not directly develop a dedicated AI weather forecast model, the research supported by the project has played an important role in improving the accuracy of weather forecasts and helping the world better cope with the challenges brought by climate change.

Overall, these cutting-edge AI weather forecast models have greatly improved the speed, accuracy and coverage of weather forecasts through the training of large-scale historical data and the application of advanced deep learning algorithms. AI technology is transforming traditional weather forecasts in a more accurate and faster way, and is playing an increasingly important and critical role in addressing the global challenges of extreme weather events and climate change.

Problems and Prospects

However, data quality and availability remain major issues . In particular, the AI ​​model training phase has a great demand for high-quality and consistent data.
The report points out that due to economic, political and geographical differences between countries, data availability is uneven. In addition, the gap in spatial and temporal data, the lack of data on small-scale weather phenomena, and the lack of high-resolution global reanalysis data all have a negative impact on the training effect of AI models. At the same time, the lack of complex variables that can affect the accuracy of weather forecasts, such as oceans, land, cryosphere and carbon cycle, in current AI models is also a major challenge that needs to be faced urgently.

On the other hand, transparency and fairness issues may also limit the application of AI in weather forecasting. Although efforts are being made to improve the interpretability of models and integrate physical constraints, increasing transparency to ensure public trust remains a major issue that needs to be addressed. In addition, due to limitations in data, computing power, and user skills, AI technology is difficult to popularize, which will further exacerbate the global digital divide and inequality.

Looking ahead , the focus of the next stage of AI weather forecasting models will include data assimilation and the development of basic models. After large-scale and diversified data training, AI models can adapt to more specific applications. In addition, expanding existing AI models to cover the entire earth system to enhance climate prediction capabilities, leveraging the potential of commercial satellites and crowdsourcing data (such as the Internet of Things) and combining low-cost data storage platforms and standardized tools to promote the popularization of AI weather forecasting technology worldwide will also be important directions for future exploration.

It is worth noting that AI will be increasingly used to support decision-making and help the global community cope with the risks of climate change and extreme weather. Strong global governance and frameworks are essential to ensure that AI technology is beneficial to all humanity and widely accessible.

To this end, it is necessary to enhance the transparency and traceability of AI models, promote trust building and develop standards for responsible use. It is also necessary to pay attention to the solution of systemic bias and fair access issues in the development of ethical AI, especially in some vulnerable communities.

In addition, training and capacity building are crucial to narrowing the digital divide and ensuring the fair and effective application of AI tools.

Author: Ruan Wenyun