Nowcast3D: Reliable precipitation nowcasting via gray-box learning

TL;DR

Nowcast3D introduces a gray-box 3D nowcasting framework that explicitly models three physical processes—advection, anisotropic diffusion, and a residual microphysical source—using volumetric radar reflectivity as input. A physics-informed predictor yields a deterministic, dynamically consistent forecast, which is then refined by a conditional diffusion model that produces ensembles and quantifies uncertainty. Evaluated on large-scale Chinese radar data, Nowcast3D outperforms extrapolation, hybrid, and purely data-driven baselines in location accuracy, structural realism, and perceptual quality, while also validating the inferred wind fields against profiler measurements. The approach demonstrates robust generalization to higher-resolution urban domains and offers interpretable physical fields and probabilistic forecasts suitable for risk-aware decision-making and potential integration with NWP.

Abstract

Extreme-precipitation nowcasting requires high spatial and temporal resolution together with extended lead times, yet current approaches remain constrained. Numerical weather prediction systems and their deep-learning emulators operate at relatively coarse space-time resolution and struggle to capture rapidly evolving convective systems. Radar extrapolation methods, which advect recent fields using estimated motion, have difficulty capturing the complex evolution of precipitation. Purely data-driven models often produce overly smoothed reflectivity fields and underestimate intensity. Hybrid 2D radar-based methods discard crucial vertical information, preventing accurate reconstruction of height-dependent dynamics. We introduce Nowcast3D, a gray-box, fully three-dimensional nowcasting framework that operates directly on volumetric radar reflectivity and couples physically constrained neural operators with data-driven learning. The model learns three fields that govern reflectivity evolution: a three-dimensional flow field for advective transport, a spatially varying diffusion field for local dispersive spreading, and a residual source term for unresolved microphysical effects. These learned operators advance the forecast in time under explicit physical constraints, while a conditional diffusion model, conditioned on both the observations and the physics-based forecast, generates ensembles of future radar volumes that quantify forecast uncertainty. In a blind evaluation by 160 meteorologists, Nowcast3D is preferred in 57% of post-hoc and 51% of prior assessments. By explicitly embedding three-dimensional dynamics and uncertainty into a single framework, Nowcast3D offers a scalable and robust approach for reliable nowcasting of extreme precipitation.

Figures (9)

• Figure 1: (a) Physical decomposition of reflectivity evolution. Reflectivity changes are partitioned into three contributions: advection by the 3D wind field, local diffusion representing small-scale spreading, and microphysical tendencies that alter reflectivity in situ. (b) Overall model schematic. Phy-Pred Network first infers latent physical fields from 3D radar; a physics-based predictor then evolves the three processes above. The physics forecast conditions a conditional diffusion model, and multiple independent noise samples produce an ensemble of probabilistic forecasts. (c) Deterministic physics backbone. The deterministic predictor inserts explicit advection, diffusion and microphysical operators into each temporal update. (d) The diffusion model captures physics-scale structure and stochastic residuals; varying noise trajectories under fixed conditioning produce calibrated probabilistic forecasts.
• Figure 1: Overview of the Nowcast3D framework. The pipeline consists of three main components: data processing (green), a physics-based prediction network (blue) and a diffusion-based generative refinement network (red).
• Figure 2: Nowcast3D accurately forecasts the evolution of a severe mesoscale convective system.a, Radar reflectivity of the convective system over North China at the forecast initialization time. The black box indicates the forecast domain, and red dots mark the locations of wind profiler radar stations. b, Quantitative forecast skill comparison using the CSI at 30 dBZ and 40 dBZ reflectivity thresholds over the 3 hour forecast horizon. c, Analysis of spatial scale fidelity using PSD. Forecast spectra are compared to the observed spectrum (black dashed line). d, Perceptual similarity assessment using the LPIPS metric. Lower values indicate better performance. e, Vertical wind profile validation at the Haidian station. The predicted zonal (east–west) and meridional (north–south) wind components are compared against wind profiler observations at 60 minute. f, Comparison of the predicted horizontal wind field with wind profiler observations in Beijing at an altitude of 2,500 m, 60 minutes into the forecast. g, Visual comparison of forecast reflectivity fields from Nowcast3D and baseline models against the observed radar sequence.
• Figure 2: Radar network coverage and low-level imputation examples.a, Spatial distribution of the operational weather radar network over China and surrounding regions, with coloured disks indicating approximate horizontal coverage at 500, 1500, 3000, 4500, 6500 and 10,000 m above sea level. The dashed rectangle marks the analysis domain used in this study. b, Schematic of the three-dimensional radar sampling geometry, illustrating how the beam widens with range and leads to sparse coverage near the surface despite dense sampling aloft. c, Example of low-level reflectivity in the Huanan area at heights of 500 m and 1,000 m before (left) and after (right) imputation. d, Same as c but for a convective event near Maoming, Guangdong, highlighting the recovery of coherent low-level structures beneath well-observed upper-level echoes.
• Figure 3: Generalization of Nowcast3D to high resolution, fine scale forecasting.a, Radar reflectivity over Maoming, Guangdong Province, at the forecast initialization time. The analysis was conducted at 0.01° resolution. b, Quantitative forecast skill comparison using the CSI at 30 dBZ and 40 dBZ reflectivity thresholds over the 3 hour forecast horizon. c, Analysis of spatial scale fidelity using PSD. Forecast spectra are compared to the observed spectrum (black dashed line). d, Perceptual similarity assessment using the LPIPS metric. Lower values indicate better performance. e, Probabilistic forecast skill assessed using the CRPS, shown as both average-pooled and max-pooled values. Lower scores indicate better-calibrated ensemble forecasts. f, Vertical profiles of forecast error and skill. MAE and LPIPS are plotted as a function of height, overlaid with the mean observed reflectivity and data coverage profiles. g, Visual comparison of forecast reflectivity fields from Nowcast3D and baseline models against the observed radar sequence.