High-Resolution Weather Forecasting May Be Data-Bound, Not Model-Bound

Introduction

Machine-learning weather forecasting is moving toward finer global grids, but higher resolution creates a basic data problem. Models operating at 0.1 degrees need extensive high-resolution training data, while decades of global reanalysis are generally available at a coarser 0.25-degree resolution. The gap makes it difficult to build and scale high-resolution forecasting systems, even when model architectures continue to improve.

The study behind BaguanHR proposes a different solution. Rather than taking a 0.25-degree forecasting model and fine-tuning it with a small collection of 0.1-degree examples, the framework focuses on transferring the data itself.

Key points

• Data transfer instead of model transfer. The authors argue that coarse-resolution forecasting discards information in a way that cannot be fully reversed. A model trained at 0.25 degrees may therefore carry a fundamental limitation when adapted directly to a finer grid.
• Super-resolution as the transfer mechanism. The paper compares forecasting with super-resolution and reports that super-resolution has lower conditional entropy and lower input amplification. In practical terms, this makes it a more stable route for converting information between resolutions.
• Variable-wise processing. Atmospheric variables do not share identical spatial patterns or scales. BaguanHR applies super-resolution by variable, generating large amounts of synthetic 0.1-degree data from ERA5 before combining those samples with real high-resolution data.
• A measurable scaling law. The reported results follow a power-law pattern: doubling the dataset lowers RMSE by 4.6% for 72-hour forecasting and by 4.9% for 120-hour forecasting.

Why it matters

Training on the combined synthetic-and-real dataset allows BaguanHR to outperform both other ML-based approaches and IFS-HRES across more than 85% of lead times within 72 hours. The result does not imply that super-resolution can perfectly reconstruct every small-scale atmospheric event. Its value is that it turns a long record of coarse-grid information into a much larger training resource for high-resolution models.

That shift could be important for the next generation of weather foundation models. Collecting or producing decades of globally consistent 0.1-degree reanalysis is difficult. A variable-wise super-resolution pipeline offers a comparatively direct way to unlock existing coarse-resolution archives and make them useful for fine-grid training.

The work also reframes the question of scaling. Once architectures become capable enough, adding parameters may not be the main path to better forecasts. The limiting factor may instead be the amount, resolution, and diversity of usable weather data. At the same time, synthetic detail must be checked for physical consistency and evaluated against independent high-resolution observations. BaguanHR is therefore best viewed as a data-expansion strategy, not a replacement for genuine high-resolution measurements and reanalysis.

Source: Hugging Face Daily Papers