APELIOTES: A DIFFUSION-BASED MODELING FRAMEWORK FOR KM-SCALE MULTI-LEVEL ATMOSPHERIC FIELDS

Research Approach

Apeliotes addresses the computational bottleneck of traditional numerical weather prediction (NWP) by replacing expensive dynamical downscaling with a two-stage machine learning pipeline. First, it uses the pre-trained Aurora foundation model to generate coarse-resolution global weather forecasts. Second, it employs a regionally trained 'Corrective Diffusion' (CorrDiff) model to map these coarse inputs to 4-km resolution regional outputs. This approach allows the model to capture complex, non-linear atmospheric dynamics and generate high-resolution fields—such as vertical wind profiles and wind power density—that are not explicitly provided by global models.

Key Findings

The authors demonstrate that Apeliotes achieves competitive performance compared to established baselines while operating at a fraction of the computational cost. The model shows high predictive accuracy, with correlations of 0.91 for 10-meter wind speed and 0.99 for 2-meter temperature. Notably, the model predicts vertical wind profiles with less than 3% error, providing critical data for applications like wind energy assessment and boundary-layer characterization that require high vertical resolution near the surface.

High-resolution weather data are essential for local decision-making, urban planning, and renewable energy management, yet generating these data via traditional physics-based models is often too slow for real-time applications. By enabling rapid, kilometer-scale forecasting, Apeliotes bridges the gap between global-scale climate models and the specific, high-resolution needs of regional stakeholders. Its ability to generate physically consistent multi-level fields directly from coarse inputs offers a scalable, data-driven alternative for environmental monitoring and infrastructure safety.