Abstract: Recent progresses in applying artificial intelligence (AI) to meteorological monitoring, forecasting, post-processing, and environmental services at Chinese Academy of Meteorological Sciences are reviewed. In data sensing and preprocessing, AI techniques are employed to learn radar and satellite features to improve sea-clutter identification, spatiotemporally consistent reconstruction of missing observations, and intelligent retrieval of key variables such as negative oxygen ions and land-surface temperature, improving the quality, continuity, and operational usability of datasets. For nowcasting/short-term forecasting, extrapolation integrating generative adversarial networks and physical constraints, and deep learning-numerical weather prediction coupled frameworks are being explored. For post-processing, Dice loss and multi-model ensemble correction are used to alleviate extreme-precipitation class imbalance and dust-forecast uncertainty, while deep error correction also improves El Niño-Southern Oscillation and regional precipitation forecasts. For end-to-end global AI forecasting, a modular assimilation-forecast cycle model (Fengyuan) has been developed to build an observation-to-forecast pipeline, thereby reducing dependence on numerical analysis fields. Cross-disciplinary progress is also highlighted, including long-term terrestrial water storage reconstruction, atmospheric pollution attribution, and first machine-learning-driven global aerosol-meteorology model, which together support water-resource and environmental management.
Building on this progress, AI application to meteorology is moving from a post-processing aid to a driver of operational transformation, yet significant bottlenecks persist. Future progress should advance along 4 strategic directions. The first, physical constraints must be more deeply embedded to improve interpretability and physical credibility alongside predictive accuracy, including tighter coupling of differentiable physical processes, conservation constraints, and deep neural networks. The second, the assimilation-forecast chain should move toward fuller observation-driven end-to-end framework, with expanded direct use of heterogeneous observations, flow-dependent error modeling, and reduced reliance on reanalysis fields. The third, deterministic mean prediction should be supplemented by probabilistic generative forecasting, integrating diffusion models, adversarial learning, and imbalance-aware strategies to better capture low-probability, high-impact extremes and improve the reliability of early warnings. The fourth, AI meteorology should expand toward coupled multi-sphere applications, integrating atmosphere, land, hydrology, chemistry, ocean, cryosphere, and human-activity drivers to support digital-twin-enabled earth-system services. Next-stage development should be anchored in high-quality observations, physics-aware intelligent modeling, and rigorous validation across multiple scenarios. Improving Fengyuan in terms of resolution, lead time, stability and interpretability can more effectively support disaster risk reduction, carbon-neutral transition planning, energy scheduling, ecological governance, and high-quality development of weather services. Meanwhile, open benchmarking protocols, reproducible workflows, and cross-regional transfer evaluation frameworks are essential to ensure that Fengyuan delivers more robust and reliable performance under diverse climatic and operational scenarios.