Original title: Bridging short- and medium-range weather forecasting with machine learning
Timothy A. Smith, Mariah Pope, Sergey Frolov, Brett Basarab +3 · August 2026
Weather agencies usually run separate forecast systems for short-range and medium-range weather. Researchers wanted to combine both ranges into a single, clearer system. They created Nested-EAGLE, a global machine learning model that zooms in with extra detail over the Contiguous United States. It works by blending detailed regional data directly into its global training. The model achieved much lower error for conditions near the ground across the Contiguous United States than NOAA's existing systems, while staying competitive across the rest of the world. Because of how it trained, it predicted total precipitation amounts less skillfully than NOAA's regional system. However, even with blurred peak values, it predicted storm locations most accurately at longer lead times.
You are packing for a camping trip. You check two separate forecasts: one for the next two days and another for the rest of the week, but they give conflicting advice because they use different data. Now suppose one forecast combines both, and for your exact campsite it zooms in to show local conditions using the same detailed data. In this paper, the model merges short and medium forecasts into one, with a high-resolution inset over the US, giving more accurate temperatures and winds for your area, though exact rainfall amounts are still hard to predict.
Weather forecasting currently uses separate models for short-term (next few days) and medium-term (week ahead) predictions. This machine-learning model merges both into one system, giving more accurate near-surface forecasts over the US and better storm location at longer leads. The public gets more consistent global weather information, though heavy rain predictions still need improvement.
The model's skill gains are limited to the US; global performance is only competitive, not superior. Precipitation forecasts are less accurate than the high-resolution rapid refresh model because deterministic training smooths out extremes. Storm location is good, but rain amounts are blurred. The method relies on high-resolution regional analysis data, which may not be available everywhere. Future work is needed to extend gains globally and improve precipitation representation.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Bridging short- and medium-range weather forecasting with machine learning — Timothy A. Smith, Mariah Pope, Sergey Frolov et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.