Super-Resolving Coarse-Resolution Weather Forecasts With Flow Matching

1. Problem Statement

Machine learning (ML) weather forecasting models have recently surpassed traditional Numerical Weather Prediction (NWP) systems in skill and efficiency. However, training and operating these models at high spatial resolutions (e.g., 0.25°) remains computationally prohibitive due to massive data throughput and memory requirements. Conversely, models trained at coarse resolutions (e.g., 1.5°) are computationally efficient and capture large-scale atmospheric dynamics well, but they fail to resolve subgrid-scale phenomena critical for regional impacts (e.g., heavy precipitation, tropical cyclones, and complex terrain interactions).

The core challenge is to achieve high-resolution (HR) forecast quality without the computational cost of end-to-end high-resolution training. Existing downscaling methods often focus on static climate states or lack consistency with forecast trajectories.

2. Methodology

The authors propose a modular framework that decouples the forecasting process from the spatial resolution. The approach treats super-resolution (SR) as a stochastic inverse problem, applied as a post-processing step to coarse-resolution forecast trajectories.

A. Problem Formulation

• Inputs: A coarse-resolution (LR) forecast trajectory $\hat{x}^{LR}_{t+k\delta}$ generated by a base model (e.g., ArchesWeatherGen).
• Goal: Reconstruct a high-resolution (HR) trajectory $\hat{x}^{HR}_{t+k\delta}$ consistent with the LR input.
• Residual Decomposition: The HR state is modeled as the sum of an interpolated LR state and a residual:
  $$\hat{x}^{HR}_t = \uparrow x^{LR}_t + \hat{r}_t$$
  Where $\uparrow x^{LR}_t$ is the LR state interpolated to HR resolution (preserving large-scale structure), and $\hat{r}_t$ is the unresolved high-frequency residual predicted by a neural network.
• Stochastic Nature: Since a single LR state can correspond to multiple valid HR realizations, the model learns the conditional probability distribution $p(r_t | \uparrow x^{LR}_t)$ rather than a deterministic mapping.

B. Model Architecture & Training

• Architecture: A 3D Swin U-Net Transformer (adapted from ArchesWeather) is used as the backbone. It processes the concatenated input of the interpolated LR state and the noisy residual.
• Generative Method: The model is trained using Flow Matching. This technique learns a velocity field that transports samples from a standard Gaussian prior to the data distribution along linear interpolation paths.
• Training Data: The model is trained exclusively on ERA5 reanalysis data (1979–2018) at 0.25° resolution, which is regridded to 1.5° to create LR-HR pairs.
• Zero-Shot Application: Crucially, the model is applied to forecast outputs (generated by ArchesWeatherGen) without retraining or fine-tuning. This works because the base forecasting model is trained to sample from the ERA5 distribution, making the inverse problem well-posed even for forecast trajectories.

C. Workflow

1. Forecasting: The base model generates a low-resolution (1.5°) ensemble trajectory.
2. Super-Resolution: For each time step in the trajectory, the SR model generates a stochastic residual conditioned on the interpolated LR state.
3. Reconstruction: The residual is added to the interpolated state to produce the final 0.25° HR forecast.

3. Key Contributions

1. Modular Decoupling: Demonstrates a framework where forecasting and downscaling are independent, allowing for the reuse of efficient coarse-resolution models while adding high-resolution detail.
2. Stochastic Inverse Modeling: Formulates super-resolution as a generative task using Flow Matching, enabling the reconstruction of physically consistent small-scale variability rather than just sharpening edges.
3. Zero-Shot Transfer: Proves that a model trained solely on reanalysis data can effectively super-resolve ML-based forecast trajectories without domain adaptation.
4. Residual Formulation: Explicitly preserves large-scale dynamics by predicting only the high-frequency residuals, ensuring the HR output remains consistent with the LR forecast's large-scale structure.

4. Results

The framework was evaluated against operational baselines (IFS ENS) and other ML models (GenCast) at 0.25° resolution.

• Design Consistency:
  • Re-coarsening: When super-resolved fields were re-gridded back to 1.5°, they maintained spatial correlations >0.99 with the original LR forecasts.
  • Variance: Activity ratios remained near 1.0, indicating no spurious variance injection or excessive smoothing.
• Forecast Skill:
  • Probabilistic Skill: The super-resolved ArchesWeatherGen forecasts achieved competitive CRPS (Continuous Ranked Probability Score) and Energy Scores compared to GenCast and IFS ENS, particularly at longer lead times (7–10 days).
  • Variable Performance: Significant improvements were observed for variables with rich small-scale variability (e.g., wind, humidity). Geopotential height, being smoother, showed less gain over simple bicubic interpolation.
• Physical Realism:
  • Spectral Analysis: Power spectra showed that the SR model recovers fine-scale energy (wavelengths < 1.5°) significantly better than bicubic interpolation, closely matching ERA5 statistics.
  • Case Studies:
    • Hurricane Teddy: The model successfully reconstructed fine-scale features like cyclonic moisture spirals and dry intrusions, preserving the large-scale storm track while adding realistic subgrid structure.
    • Storm Alex (Alps): The model captured orographic channeling and sharp moisture fronts in complex terrain, demonstrating utility for regional extreme weather.
• Efficiency:
  • Training cost for the SR module was modest (7 A100-days) compared to the base forecasting model (23 A100-days).
  • Inference is fast (~8 minutes for a 10-day global forecast on a single GPU).

5. Significance and Implications

• Computational Efficiency: This approach offers a pathway to kilometer-scale (or 0.25°) global forecasting at a fraction of the cost of training native high-resolution models. It leverages the "large-scale predictability" of coarse models while adding necessary local detail.
• Uncertainty Quantification: By generating stochastic ensembles, the method provides a more complete depiction of uncertainty at fine scales, which is critical for risk assessment in extreme weather events.
• Scalability: The modular nature allows the framework to be applied to any coarse-resolution forecasting system (ML or NWP) that produces physically plausible large-scale flows.
• Future Directions: The authors suggest extending this to sub-daily temporal resolution and applying it to climate subseasonal-to-seasonal (S2S) forecasting, where large-scale dynamics are well-resolved but local variability is often under-represented.

In conclusion, the paper establishes learned generative super-resolution as a viable, physically grounded, and computationally efficient strategy for bridging the gap between coarse-resolution global forecasts and the high-resolution needs of regional weather impact prediction.