Reproduce “Nowcast3D: Reliable precipitation nowcasting via gray-box learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Huaguan Chen, Wei Han, Haofei Sun, Ning Lin, Xingtao Song, Yunfan Yang, Jie Tian, Yang Liu and 4 more

Nowcast3D is a gray-box learning framework for 3D precipitation nowcasting that leverages volumetric radar reflectivity by combining physics-inspired neural operators with a conditional diffusion model. This framework provides reliable, physically consistent, and probabilistically quantified forecasts for extreme precipitation, demonstrating superior performance in quantitative metrics and a 57% preference rate in blind meteorologist evaluations compared to existing 2D methods.

Key insights

  • Directly modeling the 3D nature of precipitation systems with volumetric radar data, rather than 2D composites, significantly enhances nowcasting accuracy and the ability to capture complex, rapidly evolving extreme events.
  • Integrating physically constrained neural operators allows a deep learning model to infer underlying physical processes (such as 3D wind fields, diffusion, and microphysical tendencies) without explicit labeled training data for these variables, improving interpretability and physical consistency.
  • Conditional diffusion models can effectively refine deterministic physical forecasts into reliable probabilistic ensembles, enabling robust uncertainty quantification that is critical for operational decision-making in severe weather scenarios.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.