Evaluating Skill and Stability of ArchesWeather and ArchesWeatherGen under Multi-Decadal Climate Simulations

Researchers demonstrate that ArchesWeather and ArchesWeatherGen, machine learning models originally designed for weather forecasting, can be successfully adapted for multi-decadal climate simulations by conditioning on sea surface temperature and sea ice data. The models produce stable long-term climate outputs that faithfully reproduce observational climatology and large-scale atmospheric patterns, suggesting ML-based weather models may have untapped potential for climate modeling applications.

This research addresses a significant gap in machine learning applications to climate science by repurposing weather forecasting models for extended climate simulations. The key innovation involves adapting ArchesWeather and ArchesWeatherGen—originally trained for 10-day weather predictions—into forced atmospheric models by incorporating monthly boundary conditions of sea surface temperature and sea ice cover. This approach follows the standardized AIMIP Phase 1 protocol, enabling direct comparison with traditional numerical climate models and establishing benchmarking consistency across emerging ML-based climate tools.

The work demonstrates that despite their weather-focused origins, these models maintain stability over decades while capturing critical climate features including annual cycles, interannual variability, and distribution tails. Performance metrics show faithful reproduction of ERA5 climatology and large-scale circulation patterns, suggesting the models successfully learned generalizable atmospheric physics during weather training rather than memorizing short-term patterns. This finding challenges assumptions that weather and climate modeling require fundamentally different architectures.

For the AI research community, this validates ML models as viable alternatives to computationally expensive numerical simulations for certain climate applications. However, the work remains primarily academic with limited near-term commercial implications. The research establishes important validation frameworks for future ML climate models and provides evidence that forced configurations can mitigate some limitations of purely unforced approaches. Future development should focus on assessing predictive skill for climate-relevant phenomena like monsoons and extreme events, and exploring whether these models can eventually reduce computational barriers to climate projections.

• →ML weather models can be effectively adapted for stable multi-decadal climate simulations using boundary condition conditioning on SST and sea ice
• →ArchesWeather and ArchesWeatherGen faithfully reproduce ERA5 climatology and large-scale atmospheric circulation patterns without traditional physics-based modeling
• →Forced model configurations demonstrate superior stability compared to unforced setups, suggesting boundary constraints are critical for climate applications
• →AIMIP Phase 1 protocol provides standardized evaluation framework enabling direct comparison between ML-based and numerical climate models
• →Research validates that atmospheric physics learned during weather training generalizes to extended climate timescales beyond original forecasting scope