A commentary of “Artificial intelligence models bring new breakthroughs in global accurate weather forecasting”: Top 10 Scientific Advances of 2023, China

Skillful weather forecasts have the potential to save lives, support emergency management, and mitigate economic and social losses, which capture the public's attention. The present framework of numerical weather prediction (NWP) can trace its origins back to the 1950s, by solving partial differential equations (PDEs) that describe atmospheric motion, to infer future atmospheric states. Such forecasts typically require several hours on a supercomputer with hundreds of nodes for the upcoming days. Among the various numerical models used by operational centers, the Integrated Forecasting System (IFS) from the European Centre for Medium-Range Weather Forecasts (ECMWF) stands out for its superior skills in medium-range weather forecasts. The quiet revolution of NWP has also been evaluated by the World Meteorological Organization as one of the most significant scientific, technological and social advances in the twentieth century [1].

In recent years, the advancement of artificial intelligence (AI) has opened up possibilities for developing numerical models using an AI-based method. In contrast to the conventional approach, such AI meteorological models are trained with massive amounts of data instead of relying on prescribed physical laws. Pangu-Weather [2] proposed by Huawei stands out as one of the most prominent models in this regard (Fig. 1). Numerous reforecast evaluations indicate that its accuracy is comparable to or even better than IFS, and its computational cost is much smaller.

Fig. 1.

The 3D neural network of Pangu-Weather[2].

Using a data-driven modeling paradigm but achieving better skills positions Pangu-Weather as one of the “Top 10 Scientific Advances of 2023, China”. Specifically, Pangu-Weather employs a customized Transformer-style structure and is trained by ERA5 reanalysis data with a resolution of 0.25°, providing forecasts of the global variables in three dimensions, including geopotential, wind, pressure, temperature, and humidity. Notably, Pangu-Weather outperforms IFS by extending the skillful lead time to approximately 0.6 days. Simultaneously, the accuracy of typhoon track in reforecasts has also been improved by approximate 25% compared with IFS. Moreover, it can generate global weather forecasts for several days within a few seconds, resulting in a power consumption reduction of over 10,000 times compared with traditional approaches. Moreover, together with GraphCast [3] by Google DeepMind and FourCastNet [4] by Nvidia, AI meteorological models have garnered recognition as one of the “Top 10 Breakthroughs of Science in 2023” and have become popular tools in atmospheric research.

The fundamental question about the success of AI meteorological models lies in their approaches to achieving skillful forecasts without predefined physical laws, sparking unprecedented discussions in the scientific community. The crucial issue prompts various explorations, such as whether AI models uncover effective solutions for solving the PDEs describing the atmospheric evolutions, or if they have learned concrete physical mechanisms from the extensive datasets. If so, how can these captured physical mechanisms be explicitly or mathematically characterized? The answers to these questions may carry broader implications beyond the realm of AI models alone. Despite the challenge of lacking clear scientific explanations, AI models contribute value in applications due to their high skill and low energy consumption. Particularly, AI models can provide supplementary information for weather forecasts, especially for high-impact weather and climate events that conventional numerical models struggle to simulate accurately. Additionally, the rapid processing speed of AI models enables swift and timely forecasts. Presently, many global meteorological operational forecast agencies, including the China Meteorological Administration and the ECMWF, have initiated the real-time assessment of forecast capabilities exhibited by various AI meteorological models, with a commitment to putting these models into operations.

In addition to Pangu-Weather, a series of independently developed AI atmospheric and oceanic models have emerged in China, such as FuXi [5], FengWu [6], NowCast [7], AI-GOMS [8], and XiHe [9]. Many universities and institutes including Fudan University, Tsinghua University, National University of Defense Technology, Shanghai AI Lab, and the National Meteorology Center are collaborating to create AI models with advanced skills. Some of these models incorporate pre-existing physical laws to a certain extent, possessing notable physical interpretability and exhibiting higher forecast skills compared with conventional numerical models. Many AI meteorological models have also been developed by foreign institutes and companies, including GraphCast from Google, FourCastNet from Nvidia, and AIFS [10] from ECMWF, featuring distinct network structures. All the aforementioned models stand at the forefront of the global trend in AI geoscience, exhibiting growing influence and effectiveness of AI in meteorology.

In addition to constructing or enhancing more skillful AI models, it is crucial to expand their applications with the advantages in AI approaches. Here are several aspects that AI models can be further investigated.

First, the hybrid construction of AI models incorporating both physical laws and data is a promising avenue. The PDEs can be utilized to describe the dynamic processes while the ambiguous processes can be constructed with AI approaches. The hybrid constructed model like NeuralGCM [11] can not only enhance model interpretability, but also contributes to improved forecast skills.

Second, AI models can find their applications in computationally intensive fields like ensemble forecasts, data assimilation and targeted observations. The rapid computation capability of AI approaches can be harnessed to expedite crucial processes like parameterization schemes, radiative transfer, and more. This application of AI approaches has the potential to significantly reduce the time required to obtain results from hours to minutes.

Third, the incorporation of multiple layers into AI models is essential for developing a fully-coupled AI Earth system. Presently, many AI models focus on single-layer variables (e.g., only atmosphere or ocean) due to the short timescales involved in forecasting. However, for extended timescales, particularly in the subseasonal to seasonal range, it is crucial to introduce variables related to land and oceanic processes. The development of an AI Earth system, encompassing these multiple layers, could prove to be an effective solution to current challenges in subseasonal to seasonal forecasts.

In summary, AI models have the potential to significantly advance geoscience research and, to some extent, bring about a paradigm shift, rather than serving as substitutes for numerical models. Embracing a more open-minded approach is essential for tightly integrating the strengths of dynamics-driven and data-driven methodologies. However, it is crucial to conduct thorough investigations and explorations to maintain a certain degree of interpretability and credibility in the application of these AI models.

Declaration of competing interest

The authors declare that they have no conflicts of interest in this work.

Biographies

Mu Mu(BRID: 08901.00.70383) is a professor in the Department of Atmospheric and Oceanic Science & Institute of Atmospheric Sciences at Fudan University in China. He obtained his Ph.D. in mathematics from Fudan University in 1985 and was elected as a member of the Chinese Academy of Sciences in 2007. Prof. Mu's research interests include predictability of weather and climate, ensemble forecast and targeted observation in atmosphere and ocean, as well as nonlinear stability and instability problems in geophysical fluid dynamics.

Bo Qin(BRID: 09976.00.90871) is a post-doctor in the Department of Atmospheric and Oceanic Science & Institute of Atmospheric Sciences at Fudan University in China. He received the B.S. degree and the Ph.D. degree in School of Software Engineering at Tongji University in China, in 2017 and 2023 respectively. Dr. Qin is mainly engaged in building intelligent simulating/forecasting models with a certain degree of physical interpretability for multiple weather/climate phenomena by artificial intelligence.

Guokun Dai(BRID: 06636.00.67182) is an associated research scientist in the Department of Atmospheric and Oceanic Science & Institute of Atmospheric Sciences at Fudan University in China. He obtained his Ph.D. in meteorology from the Institute of Atmospheric Physics, Chinese Academy of Sciences, and conducted postdoctoral research at Fudan University. Dr. Dai's research interests encompass the predictability of extreme Eurasian events, the linkage between the Arctic and Eurasia and the application of machine learning in geosciences.