AI-guided regional model gives 3 km, 72 h forecasts in about 21 s with lower error

Why it matters

High-resolution regional forecasts support disaster mitigation and decisions in sectors like transport and energy, but numerical regional models need a lot of computing. This approach produces hourly regional fields from global initial conditions in seconds on one GPU. It may make frequent, multi-region high-resolution forecasting cheaper.

What they did

The authors built AERO-ODE from two separately trained networks. One refines pressure-level tendencies from NeuralGCM’s dynamical core using terrain and time encodings. The other predicts surface pressure, wind and 2 m temperature. They trained on 2021 to 2023 hourly data with HRRR 3 km analyses as targets. They tested on daily 00 UTC starts in 2024 over an eastern, flat region and a western, mountainous region. They compared with WRF-ARW, YingLong-WRF, NeuralGCM, IFS and PanGu-Weather, and with HadISD station observations.

Key findings

• At 48 h, the median RMSE reduction across 29 outputs is 29–32% versus WRF-ARW and 28–33% versus YingLong-WRF.
• All RMSE reductions relative to WRF-ARW are significant in both regions (p < 0.05). Median anomaly correlation is 0.04–0.05 higher than WRF-ARW.
• AERO-ODE generally has lower errors than interpolated NeuralGCM, IFS and PanGu-Weather fields, especially for low-level and surface variables in the mountainous region.
• Extreme-event CSI and ETS improve for most variables. Lower-level humidity is a weaker case, with lower ETS than the regional baselines over much of the forecast.
• A 72 h hourly forecast over one region takes about 21 s on average.

Limitations

• The model does not impose hard physical constraints, so the authors make no claim that it conserves mass or energy. Precipitation is not forecast.
• At the earliest lead times, AERO-ODE has slightly higher RMSE than a regional baseline for several variables. The reference analysis comes from the same system as WRF-ARW, which may favor it.
• Surface weights trained in the western region raise RMSE when moved to the eastern one, so new regions need retraining or fine-tuning. Only two US regions were tested.

Glossary

• WRF-ARW: A widely used numerical regional weather model; here it stands for the HRRR operational forecasts.
• HRRR: A 3 km, hourly updated US analysis and forecast system used here as the reference truth.
• NeuralGCM: A global model that mixes equations of atmospheric motion with neural networks; its dynamical core feeds AERO-ODE.
• ETS: Equitable threat score, which rates how well forecasts hit extreme events after removing chance hits.