StormDiT: A generative AI model bridges the 2-6 hour 'gray zone' in precipitation nowcasting

TL;DR

StormDiT is presented, a unified generative model that treats weather evolution as a holistic spatiotemporal problem, learning the coupled physics of the gray zone without human-imposed structural priors, and establishes the first robust baseline for 3-hour forecasting.

Abstract

Accurate short-term warnings for extreme precipitation are critical for global disaster mitigation but are hindered by a persistent predictability barrier at the 2-6 hour horizon -- the "nowcasting gray zone." In this window, traditional observation-based extrapolation fails due to error accumulation, while numerical weather prediction is computationally too slow to resolve storm-scale dynamics. Recent generative AI approaches attempt to bridge this gap by decomposing precipitation into separate deterministic advection and stochastic diffusion components. However, this decomposition can sever fundamental causal links between entangled atmospheric processes, such as the dynamic initiation of convection triggered by boundary advection. Here we present StormDiT, a unified generative model that treats weather evolution as a holistic spatiotemporal problem, learning the coupled physics of the gray zone without human-imposed structural priors. Trained on a massive dataset of 7,720 precipitation events from China, our model achieves a breakthrough in long-horizon stability. On a heavy-rainfall test set, it maintains skillful prediction for strong convection ($\ge$ 35 dBZ) with a Critical Success Index (CSI) near 0.2 across the full 6-hour forecast at 6-minute resolution. Crucially, the model exhibits superior probabilistic calibration, accurately quantifying operational risks. On the public SEVIR benchmark, our unified paradigm more than doubles the state-of-the-art 1-hour performance for heavy rain and establishes the first robust baseline for 3-hour forecasting. Furthermore, interpretability analysis reveals that the model attends to non-local physical precursors, such as outflow boundaries, explicitly validating its emergent understanding of convective organization.

Figures (9)

• Figure 1: The StormDiT Frameworka, The landscape of deep learning-based radar nowcasting. The visualization maps representative models by publication year and forecast horizon. StormDiT targets the 2--6 hour Gray Zone, a critical operational window where traditional methods struggle to balance high-resolution detail with long-term consistency. Bubble size scales with model parameter count. b, Workflow. The system decouples large-scale dynamics from pixel-level redundancy by operating within a compressed latent manifold. A Causal VAE encoder projects high-dimensional radar sequences into tokens. The DiT backbone then processes these states as a unified sequence, conditioning on historical context (blue tokens) to predict the evolution of target states initialized from a Gaussian prior (noise-textured tokens). c, Rectified Flow. A schematic visualization of transport trajectories from the noise distribution ($\pi_0$) to the physical data distribution ($\pi_1$). Conventional deterministic regression inherently collapses towards the conditional mean, resulting in blurry predictions (blue halo). While standard diffusion models traverse curved, stochastic paths (red dashed line), StormDiT employs Rectified Flow to learn straight-line Optimal Transport trajectories (green solid lines), ensuring both sampling efficiency and structural sharpness. d, DiT block. The core DiT block ($N=28$) utilizes Adaptive Layer Normalization (adaLN) to enforce continuous-time dynamics. Time embeddings are injected into every block to regress dimension-wise scale and shift parameters, effectively modulating feature statistics to direct the latent transport trajectory.
• Figure 2: Statistical performance of precipitation nowcasting on the 2025 China dataset. a, Temporal evolution of the Critical Success Index (CSI) for varying reflectivity thresholds (5–45 dBZ) over the 6-hour forecast horizon. The solid lines and shaded regions represent the mean and 95% confidence intervals, respectively, derived from the test set of $n=2,624$ samples collected in 2025. The slow decay rate illustrates the model's stability in the forecasting "gray zone" (2–6 h). b, Aggregated mean CSI scores for discrete intensity thresholds. Error bars indicate the standard deviation. The model demonstrates sustained predictive skill for high-intensity events ($\ge$45 dBZ), avoiding the collapse in performance often seen in long-lead-time forecasts.
• Figure 3: Forecasting the genesis and organization of a mesoscale convective system. a–c, Quantitative verification metrics (PSD, PDF, and CSI) confirming the spectral fidelity and forecast skill for this specific case. d, Spatio-temporal evolution of a squall line event initialized at 18:30 UTC on 13 March 2025. The sequence captures the rapid consolidation of scattered cells into a highly organized bow-echo system. e, details at T+3h and T+6h. The red highlights indicate the reconstruction of sharp reflectivity gradients at the squall line's leading edge, contrasting with the smoothing effects typical of deterministic baselines.
• Figure 4: Capturing the rotational dynamics and dissipation of Typhoon "Co-May". a–c, Spectral and probabilistic metrics demonstrating the preservation of physical realism during the storm's decay. d, Forecast sequence initialized at 09:30 UTC on 31 July 2025, during the typhoon's weakening phase. The model captures the cyclonic rotation while simultaneously resolving the fragmentation and physical dissipation of the outer spiral rainbands. e, Detailed views of the rainband structure at T+3h and T+6h, showing the model's capability to maintain high-frequency texture and avoid blurring artifacts throughout the forecast period.
• Figure 5: Qualitative comparison on the 1-hour SEVIR forecast task. A visual comparison example of precipitation forecasts from different models for a convective event from the SEVIR dataset.