Chen et al. (2026) Environment-conditioned deep learning for severe convection nowcasting

Identification

• Journal: npj Climate and Atmospheric Science
• Year: 2026
• Date: 2026-10-02
• Authors: Lei Chen, Zijian Zhu, Xiaoran Zhuang, Tianyuan Qi, Yuxuan Feng, Xiaohui Zhong, Hao Li
• DOI: 10.1038/s41612-026-01558-7

Research Groups

• Fudan University: Artificial Intelligence Innovation and Incubation Institute
• Shanghai Innovation Institute
• Jiangsu Meteorological Observatory
• Jiangsu Key Laboratory of Severe Storm Disaster Risk/Key Laboratory of Transportation Meteorology of CMA, Nanjing, China
• FuXi Intelligent Computing Technology Co. Ltd, Shanghai, China

Short Summary

This paper presents an environment-conditioned deep learning system called FuXi-Nowcast for severe convection nowcasting. The model combines high-resolution observations with three-dimensional atmospheric forecasts to predict composite reflectivity, precipitation, wind gusts, and surface variables up to 12 h ahead.

Objective

• Develop a deep learning-based nowcasting framework that can better represent convective initiation and maintain intense convective signals.
• Improve the accuracy of severe convection nowcasting by conditioning high-resolution forecasts on three-dimensional atmospheric fields.

Study Configuration

• Spatial Scale: East China, spanning 29.01∘N to 36.68∘N and from 114.67∘E to 122.34∘E, on a 768×768 domain with a spatial resolution of 0.01∘.
• Temporal Scale: April–July 2024 evaluations over East China.

Methodology and Data

• Models used: FuXi-Nowcast, CMA-MESO, Persistence, PySTEPS
• Data sources: ERA5 reanalysis, FuXi-2.0 forecasts, radar composite reflectivity, observations from ground-based weather stations, HRCLDAS, and a static land-sea mask.

Main Results

• FuXi-Nowcast outperforms operational numerical weather prediction, persistence, and extrapolation baselines for composite reflectivity and precipitation.
• The model shows its clearest advantage for convective initiation and maintenance of intense echoes over longer lead times.
• Case studies demonstrate that FuXi-Nowcast can predict convective initiation and subsequent convective evolution more realistically than the baselines.

Contributions

• This study presents a deep learning strategy for forecasting high-impact events whose precursors are weak or non-existent in the target observation stream.
• The model conditions high-resolution forecasts on three-dimensional atmospheric fields, providing pre-convective environmental information unavailable to observation extrapolation models.
• FuXi-Nowcast mitigates two major limitations of radar-only nowcasting methods: they cannot infer new convective development when precursors are absent from the target reflectivity field, and they often smooth or dissipate intense convective signals during autoregressive prediction.

Funding

• This research was supported by the National Natural Science Foundation of China (Grant No. 91837204) and the Shanghai Municipal Science and Technology Commission (Grant No. 19QH1410300).

Citation

@article{Chen2026Environmentconditioned,
  author = {Chen, Lei and Zhu, Zijian and Zhuang, Xiaoran and Qi, Tianyuan and Feng, Yuxuan and Zhong, Xiaohui and Li, Hao},
  title = {Environment-conditioned deep learning for severe convection nowcasting},
  journal = {npj Climate and Atmospheric Science},
  year = {2026},
  doi = {10.1038/s41612-026-01558-7},
  url = {https://doi.org/10.1038/s41612-026-01558-7}
}