🌩️ StormInsight: Hierarchical Environmental Forcing and Vertical Coupling for Convective Systems Evolution

Official repository and dataset for the paper "StormInsight: Hierarchical Environmental Forcing and Vertical Coupling for Convective Systems Evolution" (ICML 2026).

🔗 Dataset and Code Access: Hugging Face Repository

---

📖 Overview

Weather nowcasting forms the first line of defense against rapidly evolving weather hazards. However, existing systems predominantly extrapolate 2D radar reflectivity, which consistently struggles under rapid intensification regimes. Convection evolution is fundamentally a three-dimensional, environment-driven process.

🏗️ Model Architecture

StormInsight rethinks nowcasting as an environment-conditioned, vertically structured dynamical modeling task. It consists of two tightly coupled components:

• Storm Evolution Encoder: Explicitly disentangles the convective system state from vertical thermodynamic coupling and large-scale environmental forcing.
  • Convective State Encoder: Extracts instantaneous physical states from multi-modal observations.
  • Vertical Interaction Encoder: Models directional and temporally decoupled cross-layer feedbacks (e.g., latent heat release, moisture convergence) via an expert-driven Mixture-of-Experts (MoE) component.
  • Atmospheric Environment Encoder: Captures synoptic-scale constraints using Multi-mesh Message Passing.
• Convective System Decoder: Predicts future radar echoes by adaptively aggregating cross-layer interactions conditioned on evolving environmental conditions. It leverages a Conditional Flow Matching framework, using Global and Local Adapters to enforce synoptic constraints and inject vertical feedbacks.

---

🗄️ The StormBench Dataset

To support comprehensive evaluation, we built StormBench, a new large-scale benchmark that unifies multi-source observations and atmospheric reanalysis data across multiple regions.

Spatial & Temporal Coverage

• USA: Covers 2017-2019, utilizing SEVIR and GHCNh, with a 5-minute sampling interval.
• France: Covers 2016-2018 across Northwestern and Southeastern regions, utilizing MeteoNet, with a 5-minute cadence for radar/station and hourly for satellite.

Multi-Modal Variables Included

• Satellite: VIS, IR channels (e.g., IR069, IR107, IR039, WV062) to capture cloud-top morphology.
• Radar: Vertically Integrated Liquid (VIL) and Radar Reflectivity (CR) for mapping convective intensity.
• Surface Weather Stations: Temperature, Mean Sea-Level Pressure, Wind Speed/Direction, Relative Humidity, and Accumulated Precipitation.
• ERA5 Reanalysis Data: Pressure-level fields (e.g., divergence, geopotential, specific humidity) and Single-level fields (e.g., CAPE, boundary layer height).