Generative machine learning for skilful 3D radar nowcasting

Abstract

Timely, reliable, and robust radar nowcasting is an essential tool for extreme precipitation predictions and weather-dependent decision-making, yet existing methods still face two limitations: effective utilization of 3D radar data and robust prediction with occlusions or missing observations. We propose EchoCast-3D, a generative AI-based 3D ensemble probabilistic nowcasting model. Based on a Mask Diffusion Transformer backbone and trained using 3D radar echo data, EchoCast-3D delivers spatiotemporally consistent 3D forecasts, and generates reliable and complete predictions even when observations contain missing areas, a situation common in operational practice. In multiple real-world rainstorm case studies, EchoCast-3D precisely predicts the 3D evolution of severe convective systems and precipitation processes. Quantitative verification indicates that compared to existing powerful 2D nowcasting systems, EchoCast-3D achieves remarkable improvements of 34.5% in Continuous Ranked Probability Score, 14.5% in Mean Absolute Error, and 17.6% in Critical Success Index at echo intensity exceeding 40 dBZ. Even with 15% data missing, EchoCast-3D still can deliver stable and reasonable predictions, reaching the current state-of-the-art. Our research demonstrates practical application value in extreme weather preparation, and provides accurate, robust radar nowcasting in operations. We anticipate this work will serve as a foundation for new insights in nowcasting research.