Rescene Adds Band-Limited Stochastic Wrapper to Frozen Vision-Transformer for Drift-Free 100-Year Climate Emulation
Rescene stabilizes a frozen neural weather operator for 100-year climate runs via a lightweight deterministic climatology blend and band-limited stochastic perturbations. The 0.4 M-parameter wrapper restores observed variability and blocking statistics while preserving the original model's small-scale energy cascade. Evidence comes from ERA5-based multi-decadal integrations showing no drift and calibrated ensembles.
This architecture directly addresses the stability barrier that has limited prior ML weather models to short lead times. By demonstrating that band-limited forcing plus climatological relaxation suffices for century-scale runs, Rescene lowers the computational cost of ensemble climate projections and opens a route to rapid experimentation with different frozen operators. Future work could test whether the same wrapper transfers to higher-resolution or physics-informed backbones now emerging from operational centers.
Rescene team: The same wrapper will keep spread-skill ratio above 0.75 when applied to a 0.25-degree frozen operator by end of 2027.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2608.09971)
- [2]Supporting Source(https://www.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era5)