In today's increasingly severe climate change, accurate weather and climate forecasting has become more important than ever.

Recently, Google researchers have developed a new type of weather and climate forecasting model called NeuralGCM, driven by artificial intelligence technology. This latest achievement was published in the journal Nature.

The model ingeniously combines the advantages of artificial intelligence technology and traditional physical models. In terms of accuracy for forecasts from 1 to 15 days, its performance is comparable to that of the European Centre for Medium-Range Weather Forecasts (ECMWF), and its computational speed is very fast, capable of generating a 22.8-day atmospheric simulation in just 30 seconds.

Advertisement

Integration of Traditional Models and AI Models

Since the 1950s, weather forecasting has mainly relied on mathematical models based on physical equations.These models operate on supercomputers, using complex calculations to predict future weather conditions.

Over time, these models have become increasingly detailed, requiring more and more computing power.

However, even the most advanced supercomputers struggle to fully simulate the complexity of the Earth's atmospheric system.

It is noteworthy that the NeuralGCM model, while capable of simulating large-scale physical processes, can also use artificial intelligence technology to handle small-scale phenomena such as cloud formation and regional microclimates.

Artificial intelligence is primarily used to handle small-scale phenomena that traditional models find difficult to simulate accurately, and to correct errors that accumulate at the small scale.Stephan Hoyer, an artificial intelligence researcher at Google Research and co-author of the paper, said: "This is not a confrontation between physical (models) and artificial intelligence (models), but a combination of the two."

A dual breakthrough in performance and accuracy

The research shows that the accuracy of NeuralGCM in weather forecasting for 1-15 days is on par with the ECMWF model.

ECMWF is a partner organization in this study, and its model is considered one of the most accurate weather forecasting models in the world.

More importantly, NeuralGCM has performed well in long-term climate prediction. Its 40-year climate prediction simulation results are consistent with the global warming trend shown in the ECMWF data.The NeuralGCM model is trained on 40 years of historical weather data provided by ECMWF. This training method based on historical data enables the model to learn the complex patterns and regularities of weather systems, thereby performing well in future predictions.

At the same time, NeuralGCM also demonstrates excellent advantages in simulating seasonal cycle phenomena, including basic atmospheric dynamics, annual cycles of global precipitable water and total kinetic energy, as well as the unique seasonal behaviors of Hadley circulation and monsoon circulation.

Aaron Hill, an assistant professor of meteorology at the University of Oklahoma, pointed out that simulating global climate change or long-term climate trends is a highly computationally intensive task, which is often difficult to achieve in traditional models due to high computational costs. Therefore, the real prospect of this technology lies in simulating large-scale climate events.

Compared with traditional models, NeuralGCM has a significant advantage in computational efficiency. Researchers say that the model can process 70,000 days of simulation in 24 hours using a single tensor processing unit chip.

Another outstanding advantage of NeuralGCM is its small code volume. After training, machine learning-based models such as Google's GraphCast require less than 5,500 lines of code, while the model of the National Oceanic and Atmospheric Administration in the United States requires more than 377,000 lines of code.This simplification not only improves operational efficiency but also makes the model easier to maintain and update.

Potential and challenges coexist.

Although NeuralGCM has performed well in short-term weather forecasting, Hoyer said that its ultimate goal is to be used for long-term modeling, especially for predicting extreme weather risks. This capability is crucial for addressing the challenges brought by climate change.

In addition to scientific research, NeuralGCM may also attract the interest of a wider range of groups, such as commodity trading, agricultural planning departments, and insurance companies, because these fields attach great importance to more accurate weather forecasts, and NeuralGCM may play an important role.

Despite the great potential of NeuralGCM, it will take time to completely replace traditional models.Hill pointed out that people have not yet established sufficient trust in these artificial intelligence-based prediction systems.

He said: "People perform well at work partly because they understand the strengths and weaknesses of the models currently in use, and know when certain models perform well and when other models may have biases."

Another challenge is the "black box" nature of artificial intelligence models. Unlike traditional physical models, the internal workings of artificial intelligence models are often difficult to explain or replicate, which is a problem for climate research that requires rigorous scientific verification.

In addition, climate scientists have also pointed out that if the model is trained only based on historical data, it may encounter difficulties in predicting unprecedented phenomena caused by climate change.

Hill believes that although many artificial intelligence skeptics in the field of weather forecasting have been convinced by recent developments, the rapid pace of development makes it difficult for the academic community to keep up.He said: "I feel like every two months, Google, Nvidia, or Huawei releases a new model."

In fact, this is indeed the case. In July 2023, Huawei released the Pangu meteorological model, and its research results were published in Nature; in November 2023, Google DeepMind announced the machine learning-based weather forecasting model GraphCast, with the results published in Science; in March 2024, Nvidia launched the "Earth 2" digital twin platform for estimating climate change on Earth.

It is worth mentioning that NeuralGCM will be released in an open-source format. This decision not only benefits the research of climate scientists but also provides the broader technology community with the opportunity to participate in and improve this model.

It is foreseeable that with the rapid development of artificial intelligence technology in the field of weather and climate forecasting, both academia and industry will continue to actively explore its application potential.