Seasonal forecasting using the GenCast probabilistic machine learning model

Area of Science:

• Earth and Climate Science
• Artificial Intelligence in Meteorology
• Atmospheric Physics

Background:

• Machine-learnt weather prediction (MLWP) models rival conventional numerical weather prediction (NWP) in medium-range forecasts.
• Extending MLWP performance to longer timescales, crucial for seasonal forecasting, remains an area of active research.
• Interactions with slower Earth system components are vital for seasonal climate prediction.

Purpose of the Study:

• To evaluate the efficacy of the GenCast MLWP model for seasonal forecasting using prescribed and observed sea surface temperatures (SST).
• To compare GenCast's seasonal forecasting performance against the European Centre for Medium-Range Weather Forecasts (ECMWF) seasonal forecasting system, SEAS5.
• To assess the potential of MLWP models for comprehensive seasonal forecasting, including coupled Earth system components.

Main Methods:

• GenCast, a probabilistic MLWP model, was adapted for seasonal forecasting in two configurations: GenCast-Persisted (climatological SST anomalies) and GenCast-Forced (observed SSTs).
• Forecasts were generated and compared with SEAS5, focusing on precipitation, 2-meter temperature, mean sea level pressure (MSLP), and the North Atlantic Oscillation (NAO) index.
• Reliability diagrams were used to assess forecast skill relative to climatology and SEAS5.

Main Results:

• GenCast-Persisted accurately captured precipitation patterns associated with El Niño and La Niña events, with GenCast-Forced correcting some erroneous patterns.
• Ensemble-based uncertainty in precipitation forecasts from GenCast compared favorably with SEAS5.
• While SEAS5 excelled in tropical 2-meter temperature and MSLP, GenCast-Persisted showed higher skill in higher latitudes and mountainous regions for MSLP, correlating better with the NAO index.

Conclusions:

• GenCast-Forced demonstrated forecast reliability comparable to SEAS5, indicating the potential of MLWP for seasonal forecasting.
• GenCast-Persisted showed limited skill relative to climatology, highlighting the importance of accurate sea surface temperature forcing for seasonal predictions.
• These findings suggest that MLWP models, when coupled with other Earth system components, could significantly advance seasonal forecasting capabilities.