AirCast-SR: A Foundation Model for Kilometer-Scale Atmospheric Super-Resolution via Latent Consistency Diffusion

2026-05-20

AI Summary

This study addresses the challenge of efficiently generating kilometer-scale high-resolution weather forecasts—a capability lacking in traditional numerical weather prediction and critical for applications in energy, agriculture, and disaster management. The authors propose the first foundational atmospheric super-resolution model capable of zero-shot global generalization without fine-tuning, advancing AI-based forecasting resolution from approximately 28 km to 1 km with hourly outputs over a 67-hour lead time across eight surface variables. Built upon a latent-consistency diffusion architecture with a 3D U-Net backbone, the model is trained on contiguous U.S. data using GraphCast as input and NOAA AORC as target. Experiments demonstrate near-zero bias across all variables and lead times, while multi-season case studies and ground observations from India and Germany confirm its exceptional global generalization and fidelity in preserving fine-scale structures at wavelengths of 10–100 km.

Abstract

Operational weather prediction at kilometer scales remains computationally prohibitive for traditional numerical weather prediction (NWP) models, limiting forecast access for applications in energy, agriculture, and disaster management that require fine-grained spatiotemporal detail. Here we introduce AirCast-SR, a foundation model for atmospheric super-resolution that downscales global AI weather forecasts from 0.25 degree (~28 km) to 1 km horizontal resolution at hourly temporal resolution, producing 67-hour forecasts of eight coupled surface variables simultaneously. EarthMind-SR employs a three-dimensional U-Net conditioned within a Latent Consistency Model (LCM) diffusion framework, trained on patch-based samples over the contiguous United States (CONUS) using GraphCast forecasts as input and NOAA's Analysis of Record for Calibration (AORC) as the target. The model achieves near-zero bias across all variables and lead times, and its radial power spectral density analysis demonstrates preservation of fine-scale atmospheric structure at wavelengths of 10 km to 100 km where coarser models lose spectral power. We validate EarthMind-SR across three CONUS case studies spanning winter, summer, and spring seasons, and demonstrate zero-shot global transferability over India and Germany using independent surface station observations without any retraining or fine-tuning. As an open-weights foundation model, EarthMind-SR establishes a new paradigm for kilometer-scale AI weather prediction and provides a platform for regional fine-tuning, distillation, and downstream applications in climate services and hazard forecasting.

Problem

• kilometer-scale weather prediction
• computational prohibitive
• fine-grained spatiotemporal detail
• atmospheric super-resolution
• high-resolution forecasting

Innovation

• Latent Consistency Model
• Atmospheric Super-Resolution
• Foundation Model
• Zero-shot Transfer
• Kilometer-scale Forecasting