Explicit uncertainty quantification for GNN-based nowcasting

Develop an explicit uncertainty quantification methodology for the Graph Neural Network-based nowcasting system constructed within the Anemoi framework for Switzerland (combining surface observations, radar and satellite inputs, and ICON-CH1 NWP states at 1 km resolution and 10-minute intervals), so that the generated short-term forecasts of 2-meter temperature, 2-meter dew point, 10-meter wind components, and precipitation include quantified uncertainty in a form suitable for operational use.

Background

The paper introduces a high-resolution nowcasting approach for Switzerland using a Graph Neural Network within the Anemoi framework, integrating surface station observations, radar-derived precipitation, satellite channels, and past and future ICON-CH1 NWP states. The model produces 10-minute forecasts across a 12-hour horizon over a 1 km grid, and demonstrates performance advantages over ICON and, beyond the first hour, over INCA.

While the approach implicitly accounts for some uncertainty through its design and verification, the authors explicitly identify the lack of formal uncertainty quantification as an unresolved aspect. In operational meteorology, quantified uncertainty (e.g., probabilistic forecasts or calibrated confidence intervals) is essential for decision-making, risk communication, and ensemble-based workflows; therefore, establishing an explicit framework for uncertainty quantification is a critical next step for this GNN-based nowcasting system.