Skip to main content
Back to timeline
arXivSource publication:

Running weather backward: Rewind-AR concentrates the precursors of a pinned extreme, but up to 22% of its members go unstable within 66 hours

Synopsis

Given an observed extreme-weather end state, can the varied precursor paths that lead to it be sampled, as scenario planning needs? The video-diffusion emulator cBottle-video previously produced 66-hour final-conditioned ensembles, but they were underdispersed, geostrophically inconsistent, and reproduced the conditioning frame only approximately. Rewind-AR instead uses a FourCastNet 3 backbone as a backward-in-time probabilistic autoregressive model; for the 2021 Pacific Northwest heatwave and Hurricane Ian, its leading antecedent 500 hPa geopotential-height modes explain far more across-member variance than cBottle-video's, at the cost of up to 22% of members becoming numerically unstable within 66 hours.

Source-provided article image: Run it Back: Autoregression vs. Video Diffusion for Final-Conditioned Weather Trajectories
arXiv

Interpretation

Final-conditioned weather trajectory ensembles can now come from an autoregressive model trained explicitly for reverse-time rollout, rather than from a general-purpose masked-conditioning video diffusion model. The earlier cBottle-video was a pre-trained, off-the-shelf video diffusion climate emulator not fine-tuned for this task; its ensembles were underdispersed and geostrophically inconsistent, and it reproduced conditioned frames only approximately, which hindered autoregressive rollout. For Hurricane Ian, cBottle-video's leading eight free-end principal components explain 9%, 8%, and 14% of in-sample variance (7%, 6%, and 12% cross-validated) in free-end storm latitude, longitude, and central MSLP; the same regressions on Rewind-AR's 867 stability-screened members give 71%, 87%, and 31% in-sample and 70%, 87%, and 29% cross-validated.

The model starts from the pinned final state and rolls backward in six-hour steps, drawing from stochastic noise fields at each step, so one end state yields a set of differing precursor paths. Rewind-AR configures a FourCastNet 3 backbone as a probabilistic autoregressive model of the reverse-time conditional distribution, with 273.5M parameters counting each complex weight as two real parameters, trained on 2009–2018 ERA5 reanalysis with 2019 for validation on a 25 km regular latitude–longitude grid. Each model generates 1,000 members per case; Rewind-AR is initialized from the ERA5 state at the pinned frame, analyzes the first 66 hours (11 steps), differs across members only in random draws, and is run with batch size four and 250 seeds per 1,000-member ensemble.

Perspective

The demonstrated scope is three case ensembles constrained to an observed ERA5 end state within a 66-hour window: the 2021 Pacific Northwest heatwave, Hurricane Ian, and Superstorm Sandy, the last falling inside both models' training periods and so not testing generalization to an unseen event. The intended reader is a researcher using AI weather and climate emulators for extreme-event scenario analysis; all Rewind-AR statistics use only members surviving numerical stability screening.

Numerical stability screening removes 13.3% of Ian members and 22.2% of Pacific Northwest members by 66 hours, and none of the Sandy members, so survivors may not be a random sample of the ensemble. A closer link between geopotential-height leading modes and antecedent storm position does not by itself imply higher physical plausibility, and for Sandy some trajectories start at implausible inland locations. The two models also differ in resolution, training data, architecture, and size, so this is not a fully standardized comparison of autoregression and video diffusion.