Marchuk-S2S: Efficient Generative Modeling of Dense Subseasonal Weather Trajectories

Subseasonal weather prediction requires models that remain useful under long autoregressive rollouts while representing uncertainty over many plausible trajectories. We introduce Marchuk-S2S, a compact 276M-parameter generative temporal model that produces 50-member global weather ensembles at six-hour resolution for up to 48 days. It parameterizes a conditional flow-matching process in the latent space of a frozen weather autoencoder and uses one forecasting network across medium-range and subseasonal horizons, combining variable-horizon training, calendar-phase conditioning, and CRPS fine-tuning. Aggregated to daily resolution, its ensemble mean remains competitive with FuXi-S2S and ECMWF S2S reforecasts on temporal anomaly correlation and Madden--Julian Oscillation diagnostics, and spectral and propagation diagnostics give preliminary evidence of coherent 48-day rollouts. A 50-member, 48-day forecast takes 9.7 minutes on one NVIDIA H100, showing that dense stochastic trajectory generation can be practical at subseasonal scale.