Dynamical-Model-Informed Stochastic Flow Matching for Global Probabilistic Subseasonal Forecasting of Means and Extremes

Abstract

Physics-based models provide skillful subseasonal forecasts but suffer from systematic biases, weakened teleconnections, and parameterized physics, while very large ensembles remain computationally expensive. Purely data-driven models also see too few realizations of the low-frequency modes that provide subseasonal predictability. We introduce methodname, a hybrid framework that converts modelshort subseasonal ensemble guidance into 90-member stochastic forecasts of weekly precipitation and 2 m temperature on a global 1-degree grid using a lead-aware flow-matching U-Net. Its stochastic prior initializes members with EOF perturbations conditioned on the Madden-Julian Oscillation, North Atlantic Oscillation, El Nino-Southern Oscillation, and lead week, allowing the learned flow to refine large-scale covariance rather than create it from white noise. On held-out 2021-2025 forecasts, empirical CRPS improves by 22.0% for precipitation and 25.6% for temperature in a same-initialization four-versus-four comparison, and by 32.2% and 34.0% in the 90-versus-four production comparison. Ablations across five training seeds show distinct gains from the structured prior and a learned error-variance pathway, with the prior contributing more. Verification over 120 observed-extreme cases shows positive CRPS skill, but the ensemble remains underdispersive, four-member upper quantiles are poorly resolved, and precipitation exceedance BSS is near neutral after week 1. Results support improved probabilistic skill and tail representation, but not calibrated operational rare-event probabilities. Case studies of the 2022 UK heatwave, 2023 California atmospheric rivers, and 2022 Pakistan floods illustrate the reach and limits of dynamically informed refinement.