Rescene: A New AI Climate Emulator from Frozen Weather Models
A team of researchers has introduced Rescene, a machine learning framework that converts a static neural weather forecasting model into a reliable climate emulator. This innovation is detailed in a paper available on arXiv (2608.09971) and tackles the issues of instability found in existing ML weather models beyond their training limits. Rescene consists of a 0.4 million-parameter wrapper surrounding a fixed 1.5-degree, 6-hourly vision-transformer operator, which was trained using ERA5 reanalysis data. It features a deterministic 'slow clock' (0.33 million parameters) and a generative head (0.06 million parameters) that introduces stochastic variations, ensuring stability and preserving seasonal cycles. The findings highlight the effectiveness of this method for economical long-term climate forecasts.
Key facts
- Rescene is a 0.4 M-parameter wrapper around a frozen 1.5 degree, 6-hourly vision-transformer operator.
- The wrapper includes a deterministic 'slow clock' (0.33 M parameters) and a generative head (0.06 M parameters).
- It uses ERA5 reanalysis data for training.
- The method adds spectrally shaped stochastic perturbation at every step.
- It aims to recover stability and seasonal cycle from frozen ML weather models.
- The paper is available on arXiv with ID 2608.09971.
- The work addresses the problem of ML weather models blowing up or drifting when integrated beyond training horizon.
- Rescene blends the forecast toward a lead-aware day-of-year climatology.
Entities
Institutions
- European Centre for Medium-Range Weather Forecasts (ECMWF)
- arXiv