A research weather model that zooms in on Scandinavia while keeping the rest of the globe blurry just beat an operational forecasting system at its own game.
The model uses what's called a stretched grid: 2.5 km resolution over a Nordic region of interest, 31 km everywhere else, updated every 6 hours. It generates ensembles, meaning multiple probable weather outcomes rather than one deterministic guess, for 87 variables. Researchers trained it with a loss function based on the Continuous Ranked Probability Score, measured in both grid-point and spectral space. That second part matters: without the spectral component, the model produced fields that looked spatially incoherent, an artifact standard mean-squared-error training doesn't fix either. Tested against Norway's MetCoOp Ensemble Prediction System using real surface station data, it beat MEPS by 13% on 2-meter temperature and 10% on mean sea-level pressure.
The catch is in the details nobody puts in a press release: wind speed and precipitation improvements were marginal, and when the model was thrown at Storm Dave, it correctly placed the storm system but underestimated peak wind speeds. That's the unglamorous reality of AI weather models right now. They're excellent at smoothing out the variables with predictable statistical structure and shakier on the violent, localized extremes that actually make a forecast useful during a storm.
This fits a pattern seen across data-driven forecasting efforts: strong on averages, weaker on tails. Cheaper, faster forecasts are real progress. Trusting one to call the peak gust in the next big storm is not there yet.