Abstract
Extreme precipitation nowcasting demands high spatiotemporal fidelity andextended lead times, yet existing approaches remain limited. Numerical WeatherPrediction (NWP) and its deep-learning emulations are too slow and coarse forrapidly evolving convection, while extrapolation and purely data-driven modelssuffer from error accumulation and excessive smoothing. Hybrid 2D radar-basedmethods discard crucial vertical information, preventing accuratereconstruction of height-dependent dynamics. We introduce a gray-box, fullythree-dimensional nowcasting framework that directly processes volumetric radarreflectivity and couples physically constrained neural operators withdatadriven learning. The model learns vertically varying 3D advection fieldsunder a conservative advection operator, parameterizes spatially varyingdiffusion, and introduces a Brownian-motion--inspired stochastic term torepresent unresolved motions. A residual branch captures small-scale convectiveinitiation and microphysical variability, while a diffusion-based stochasticmodule estimates uncertainty. The framework achieves more accurate forecasts upto three-hour lead time across precipitation regimes and ranked first in 57\%of cases in a blind evaluation by 160 meteorologists. By restoring full 3Ddynamics with physical consistency, it offers a scalable and robust pathway forskillful and reliable nowcasting of extreme precipitation.