Fig. 1. NowcastNet for extreme-precipitation nowcasting.

a, Architecture of NowcastNet, a physics-conditional deep generative model. The nowcast encoder learns contextual representations. The nowcast decoder conditions on the physics-informed evolutions ${\mathbf{x}}_{1:T}^{″}$ and transforms draws from a latent Gaussian vector z into mesoscale and convective-scale predictions ${\hat{\mathbf{x}}}_{1:T}$. b, Evolution network, a neural implementation of the advection schemes informed by the 2D continuity equation, which imposes compliance with the precipitation physics and outputs mesoscale predictions ${\mathbf{x}}_{1:T}^{″}$. c, Evolution operator, a neural operator that iteratively advects x₀ by motion fields v_1:T to get ${\mathbf{x}}_{1:T}^{\prime }$ and adds by intensity residuals s_1:T to get ${\mathbf{x}}_{1:T}^{″}$. Precipitation data obtained from the MRMS²⁶ dataset and maps produced with cartopy and Natural Earth.