Breakthrough progress and scientific value of artificial intelligence in weather forecasting

<sec><p indent="0mm">The field of weather forecasting is undergoing a profound paradigm shift, driven by rapid advancements in artificial intelligence (AI) that challenge the dominance of traditional numerical weather prediction (NWP). This review provides a comprehensive synthesis of the breakthrough progress of data-driven weather forecasting in meteorology, critically examining its technical foundations, scientific implications, and future potential. The motivation for this transformation stems from the escalating societal demand for more accurate, timely, and computationally efficient forecasts to mitigate the impacts of severe weather extremes, a need amplified by a changing global climate. While NWP models rely on the explicit numerical integration of atmospheric governing equations, data-driven AI models learn the complex, non-linear dynamics of atmospheric evolution directly from vast repositories of historical reanalysis and observational data. This fundamental change in methodology has yielded orders-of-magnitude improvements in computational speed, enabling rapid-update cycles, high-resolution outputs, and large-ensemble generation at a fraction of the traditional cost. </sec><sec> Significant achievements have been documented across all forecast timescales. In the nowcasting <sc>(0–2 h)</sc> and short-range <sc>(2–24 h)</sc> domains, deep learning architectures have delivered unprecedented performance gains. Models such as Google’s MetNet series and the CMA-Tsinghua “Fenglei” (NowcastNet) have established new benchmarks. MetNet-3 provides precipitation forecasts at <sc>1 km</sc> spatial resolution with 2-minute temporal updates for up to 24 h. Concurrently, Fenglei, the first operational deep-learning nowcasting model at National Meteorological Center, extends the skillful prediction of extreme precipitation from the conventional <sc>1–2 h</sc> limit to 3 hours by uniquely embedding physical constraints within a multi-scale generative framework. These models substantially outperform traditional extrapolation and NWP-based systems in capturing the genesis, evolution, and decay of rapidly developing severe weather events. </sec><sec> For medium-range <sc>(3–15 d)</sc> global forecasting, a cohort of landmark AI models—including Huawei’s Pangu-Weather, Google DeepMind’s GraphCast, Shanghai AI Lab’s FengWu, and the CMA-Tsinghua “Fengqing”—has demonstrated the capacity to match or exceed the skill of the world’s leading NWP systems, such as the ECMWF Integrated Forecasting System (IFS). Pangu-Weather and GraphCast, for instance, deliver 10-day global forecasts in minutes, with superior accuracy in key metrics like RMSE and ACC for critical variables and demonstrating remarkable skill in tropical cyclone track prediction. The operational deployment of ECMWF’s AIFS and CMA’s Fengqing signals the formal entry of AI into the global operational forecasting landscape. </sec><sec> This technological leap forwards raises pivotal scientific questions. A central challenge is the optimal integration of physical principles with data-driven architectures. Promising pathways include the design of physics-informed network structures, the incorporation of physical conservation laws (e.g., energy conservation) into loss functions to enhance long-term stability, and the development of hybrid models that couple neural networks with traditional physics cores. The inherent “black box” nature of deep learning also necessitates advances in interpretability, for which model designs like GraphCast’s graph neural network and the analysis of attention mechanisms in Transformers offer promising avenues. Furthermore, while AI models excel at predicting certain extreme phenomena, accurately forecasting their intensity and lifecycle, especially under non-stationary climate conditions, remains a formidable challenge. Concurrently, AI-based generative and diffusion models are revolutionizing uncertainty quantification, offering computationally inexpensive, high-skill probabilistic ensembles that challenge traditional methods, although their physical consistency requires rigorous verification. </sec><sec> The scientific value of these breakthroughs extends far beyond improved forecast skill. AI models serve as new instruments for atmospheric science, capable of revealing previously obscured statistical-physical relationships, enabling novel counterfactual experiments for climate attribution studies, and facilitating an integrated Earth system science approach by breaking down disciplinary barriers. The demanding nature of weather prediction, with its complex, multi-scale, and physically constrained dynamics, simultaneously acts as a powerful catalyst for fundamental AI research. To fully realize this potential, future efforts must focus on establishing a robust theoretical foundation for physics-AI integration, developing comprehensive methods for uncertainty quantification, and fostering a new generation of interdisciplinary talent. This convergence of AI and atmospheric science is poised not only to revolutionize weather and climate prediction but also to advance the fundamental understanding of the Earth system, thereby strengthening global resilience to environmental hazards. </sec>