Sub-seasonal forecasts: a hybrid flow matching model draws 90 scenarios from four physical members

Guided by major climate modes (MJO, NAO, ENSO), the model improves the probabilistic score of weekly rain and temperature forecasts by 22 to 34%, without calibrating the extremes.

After a few weeks, physical models remain efficient but suffer from systematic biases and weakened teleconnections, and very large assemblies are expensive. Purely data-driven models, on the other hand, see too little realization of the slow modes that make this deadline predictable. Abdullah A. Fahad and Manmeet Singh propose a compromise: start from a small set of dynamic forecasts and derive 90 stochastic members of weekly precipitation and temperature at 2 m, on a one-degree global grid, thanks to a U-Net flow matching that takes into account the time frame.

A structured preconception rather than white noise

The members are initialized by perturbations conditioned by the Madden-Julian oscillation, the North Atlantic Oscillation, El Niño, and the maturity week, so that the network refines a large-scale covariance structure instead of inventing it. On the 2021-2025 forecasts that are held aside, the CRPS improves by 22% for rain and 25.6% for temperature at equal initialization, and by 32.2% and 34% in the production configuration (90 members versus four). Ablation tests show that this a priori weighs more than the rest.

The authors remain cautious: out of 120 extreme events observed, the gain is real but the whole remains too little dispersed, and the score on rainfall threshold exceedances becomes almost neutral again after the first week. The cases of the British heatwave in 2022, the atmospheric rivers in California in 2023 and the floods in Pakistan in 2022 illustrate both the scope and the limits of the approach.