Generate 10-day weather forecast in 1 minute, AI beats the world's best forecasting system

When it comes to weather forecasting, artificial intelligence (AI) is upending traditional methods, promising more accurate predictions at a faster pace and at a lower cost.

Google DeepMind has launched a machine learning-based weather forecasting model, GraphCast, which can predict hundreds of weather variables for the next 10 days within one minute at a global resolution of 0.25°, significantly outperforming traditional weather forecasting methods . In addition, the model also performs well in predicting extreme events.

The relevant research paper, titled “Learning skillful medium-range global weather forecasting”, has been published in the authoritative scientific journal Science. In addition, the relevant open source code has also been released on Github.

The findings suggest that future forecasts of both everyday weather and extreme events such as hurricanes, heat waves and cold snaps could become more accurate.

Hurricane Lee in the North Atlantic in September is an example of a successful prediction. Rémi Lam, co-first author and co-corresponding author of the paper, said, "GraphCast was able to correctly predict that Lee would land in Nova Scotia nine days before the hurricane, while traditional methods could only predict six days. This gave people three more days to prepare for its arrival ."

In response, Matthew Chantry, machine learning coordinator at the European Centre for Medium-Range Weather Forecasts (ECMWF), said that progress in AI systems in meteorology is "much faster and more impressive than we expected even two years ago."

“GraphCast consistently outperforms other machine learning models such as Nvidia’s FourCastNet, and in many ways is more accurate than our own prediction system.”

Weather forecast for the next 10 days in 1 minute

Weather has a wide-ranging and profound impact on human beings, involving many aspects such as life, health, and economy.

Weather forecasting is one of the oldest and most challenging tasks in science. Medium-term predictions play a vital role in supporting critical decision-making across sectors ranging from renewable energy to event logistics, yet they are difficult to achieve accurately and effectively.

Typically, weather forecasting relies on numerical weather prediction (NWP), which starts with precisely defined physical equations that are then converted into computer algorithms that run on supercomputers. While this traditional approach has been successful in science and engineering, designing equations and algorithms is time-consuming, and making accurate predictions requires deep expertise and expensive computing resources.

According to the paper, GraphCast is a weather forecasting system based on machine learning and graph neural networks (GNNs), which may be 1,000 times cheaper than traditional methods in terms of energy consumption.

GraphCast makes predictions at a high resolution of 0.25 degrees longitude/latitude (28 km x 28 km at the equator), with over a million grid points covering the entire Earth’s surface. At each grid point, the model predicts five Earth surface variables (including temperature, wind speed and direction, and mean sea level pressure) and six atmospheric variables (including specific humidity, wind speed and direction, and temperature) at each of 37 altitudes.

Although GraphCast training is computationally intensive, the resulting prediction model is very efficient. Using GraphCast to make a 10-day prediction takes less than a minute on a Google TPU v4 machine. In contrast, using traditional methods (such as HRES) to make a 10-day prediction may require hours of computation on a supercomputer.

To evaluate GraphCast’s forecasting skills, the researchers compared GraphCast with HRES, the most accurate medium-range weather forecast model currently available, and found that GraphCast significantly outperformed HRES in 90% of the 1,380 validation targets.

Additionally, the model is better at predicting extreme events, such as tropical cyclone tracks, atmospheric rivers (narrow regions of the atmosphere responsible for poleward water vapor transport), and extreme temperature anomalies.

In addition to weather forecasting, GraphCast can also open up new directions in other geographic spatiotemporal forecasting, including climate and ecology, energy, agriculture, human and biological activities, and other complex dynamical systems.

Previously, some researchers have expressed concerns about AI’s ability to accurately forecast extreme weather, in part because there are relatively few such events in the past to draw on. However, GraphCast reduced cyclone forecast track errors by about 10-15 miles with a lead time of 2-4 days, improved water vapor forecasts associated with atmospheric rivers by 10%-25%, and provided more accurate forecasts of extreme heat and cold 5-10 days in advance.

"It's generally thought that using AI might not be very good at predicting rare anomalies. But it seems to do a pretty good job of that," said Peter Battaglia, director of research at Google DeepMind and one of the co-authors of the study. " It also suggests that the model is capturing something more fundamental about how the weather evolves over time, rather than just looking for more superficial patterns in the data ."

But this does not mean that AI can replace all traditional forecasting methods. There are other challenges that need to be overcome before AI models such as GraphCast can be reliably used for operational forecasting.

For example, an important limitation of this approach is how to deal with uncertainty. The research focus is mainly on deterministic predictions.

GraphCast’s mean squared error (MSE) training objective encourages spatially blurring its predictions in the presence of uncertainty, which may not be ideal in some applications, especially in the context of knowing the tail or joint probabilities of events.

And, due to limitations in training data and engineering design, global AI models cannot yet generate as many parameters or as granular forecasts as traditional models. This makes AI models less useful for predicting smaller-scale phenomena such as thunderstorms and flash floods, or for forecasting larger weather systems that can produce huge variations in precipitation over a small area.

In addition, meteorologists do not yet particularly trust AI models because the inner workings of these models are not as transparent as traditional models. Jacob Radford, a data visualization researcher at the Cooperative Institute for Research in the Atmosphere at Colorado State University, said in an email: "A key role of forecasters is to explain and communicate information to partners, a task made more challenging by the lack of tools to determine why AI models make such predictions. These models are still in their infancy and still need to build trust in the research and forecaster communities before they can be considered for use."

Although the study has many limitations, the researchers are confident that it marks an important turning point in weather forecasting and opens up a new path for humanity .

Moreover, they say, the approach should not be seen as a replacement for traditional weather forecasting methods , which have been developed over decades, have been rigorously tested in many real-world settings, and offer many capabilities that have yet to be explored.

“Rather, our work should be interpreted as evidence that AI weather forecasting can meet the challenges of real-world forecasting problems and has the potential to complement and improve upon current state-of-the-art methods.”

Some progress in AI weather forecasting

In the past two years, major technology companies including Google, Microsoft and Nvidia have made a number of advances in AI weather modelling, all of which have published academic papers claiming that their AI models perform at least as well as the European model. These claims have been confirmed by ECMWF scientists.

In July this year, two research papers on "AI weather forecasting" published in Nature also mentioned two AI-based weather forecasting methods.

The Pangu-Weather model developed by Huawei Cloud uses 39 years of global reanalysis weather data as training data. Its prediction accuracy is comparable to that of the world's best numerical weather forecast system IFS, and is more than 10,000 times faster than the IFS system at the same spatial resolution.

In addition, the model NowcastNet proposed by a joint research team led by Michael Jordan, a leading figure in the field of machine learning and professor at the University of California, Berkeley, and Wang Jianmin, a professor at Tsinghua University, can combine physical laws and deep learning to make real-time precipitation forecasts.

Last month, the UK's Met Office announced a partnership with the Alan Turing Institute, an AI research centre, to develop its own neural network for weather forecasting and incorporate it into existing supercomputer infrastructure.

Simon Vosper, director of science at the UK Met Office, said climate change needs to be factored into forecasts. “ If AI-based systems are only ‘trained’ on previous weather conditions, it is compelling to question whether these systems will be able to capture new extreme weather conditions .”

“Our goal is to leverage the best that AI has to offer while leveraging traditional computer models based on atmospheric physics,” Vosper said. “We believe this fusion of technologies will provide the most powerful and detailed weather forecasts in this era of great change.”

It is foreseeable that the use of AI in weather forecasting will benefit people's daily lives, but AI will never stop there.

As Google DeepMind mentioned in a blog post: "Our research goes beyond predicting the weather to understanding broader climate patterns. By developing new tools and accelerating research, we hope AI can help the international community tackle the biggest environmental challenges we face."

Reference Links:

https://www.science.org/doi/10.1126/science.adi2336

https://github.com/google-deepmind/graphcast

https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/

https://www.ft.com/content/ca5d655f-d684-4dec-8daa-1c58b0674be1

https://www.washingtonpost.com/weather/2023/11/14/weather-forecasting-artificial-intelligence-google/

Author: Yan Yimi Editor: Academic Jun