Generalization of OMG-HD beyond CONUS-trained setting

Determine whether OMG-HD, which is trained using Real-Time Mesoscale Analysis (RTMA) labels limited to surface variables over the contiguous United States (CONUS), can generalize to produce accurate forecasts in other geographic regions, particularly in areas at different latitudes and with distinct surface conditions such as oceanic regions.

Background

OMG-HD is trained using RTMA as supervisory labels. RTMA focuses on surface variables and is restricted to the CONUS domain, which inherently constrains the diversity of training targets and regional conditions the model experiences during training.

Because training is limited to CONUS, the authors explicitly flag uncertainty about whether the learned representations and forecasting skill of OMG-HD will transfer to other regions, especially those at different latitude ranges or with substantially different surface characteristics, such as oceans. Establishing this generalization property is essential for expanding OMG-HD from a regional to a broader or global operational context.

References

This constraint restricts its broader application, as RTMA restricts the training process to focus only on the CONUS region, making it uncertain whether the resulting model can generalize to different regions, particularly those located in different latitude ranges or with differing surface conditions, such as those over the ocean.

— OMG-HD: A High-Resolution AI Weather Model for End-to-End Forecasts from Observations  (2412.18239 - Zhao et al., 2024) in Section 5 (Discussion)

Could the algorithm be applied in Europe, where GRAF also has a 4-km grid? While we do not provide objective verification statistics for Europe in the results presented, Fig.~\ref{fig:europe} shows an example of a transfer-learning \citep{pan2010} forecast in Europe using the US-based training.

— High-resolution Calibrated Probabilistic Hourly Precipitation from a Deterministic Forecast  (2608.12685 - Hamill, 13 Aug 2026) in Section 6, Results, paragraph accompanying Figure Europe

Weight reuse is demonstrated across properties and likelihoods, while regional generalization remains untested.

— From Image Morphology to Multiphysics Earth Posteriors  (2609.31229 - Yu et al., 25 Sep 2026) in Section “Structural learning and physical inference draw on different knowledge”