By predicting only the slow-changing components of the climate, this model – based on the JEPA architecture – matches the CEPMMT reference set and outperforms it on several criteria in weeks 5 and 6.
Between two weeks and two months of maturity, the "infraseasonal to seasonal" horizon (S2S) is crucial for agriculture, energy or water management, but is considered very difficult to predict. Weather models based on excellent AI up to two weeks and then deteriorate, in part because they seek to reproduce fine details that are no longer predictable at these deadlines.
Chenyu Dong and Gianmarco Mengaldo propose to target only slow components, which remain predictable. For this purpose, they recapture the JEPA architecture from computer vision, which predicts in a latent space by removing random details, and adapt it with elements borrowed from the best weather models of AI. Result: S2S-JEPA achieves performance comparable to the physical set of the European Centre for Medium-Term Weather Forecasts (ECMMT), the domain reference, and surpasses it on several indicators at weeks 5 and 6.
Read the full article on the original source
Go to the source →