Two machine learning weather models just passed a much harder test than they were designed for.
ArchesWeather and ArchesWeatherGen were originally built to forecast weather up to 10 days out. Researchers adapted them for multi-decadal climate simulation by adding boundary conditions — monthly mean sea surface temperature and sea ice cover — then ran them through the AI Model Intercomparison Project (AIMIP) Phase 1 protocol, a standardized benchmark modeled on the long-established Atmospheric Model Intercomparison Project (AMIP). The results: both models produced stable long-term simulations, maintained a consistent annual cycle, and captured interannual variability and the tails of climate distributions. They also faithfully reproduced ERA5 climatology and large-scale circulation patterns.
The significance here is the distance between the original task and the new one. Weather forecasting is a bounded prediction problem; climate simulation is an unbounded stability problem where small errors compound over years. The fact that models trained for the former can be coaxed into performing the latter — without retraining from scratch — suggests that the underlying learned representations of atmospheric dynamics may generalize further than the research community assumed. It also adds pressure on traditional numerical climate models, which are expensive to run and slow to iterate.
ArchesWeatherGen's probabilistic design, which uses flow-matching to generate ensemble forecasts, gave researchers a way to quantify uncertainty across long runs — something purely deterministic models can't offer. Whether these models hold up under more demanding AIMIP phases, or whether the boundary-condition crutch limits their applicability, remains an open question.