Weather models as climate models: a successful experiment
AI weather models are already confidently forecasting the atmosphere several days ahead. But climate is the statistics of weather over decades, so transferring weather algorithms to the climate domain has long been considered risky. If a model is trained only on short periods, what happens after a year, ten, or thirty years of simulation? Won't it start "drifting" into an unrealistic state? The authors of a recent preprint on arXiv decided to test this in practice, using two AI models as test subjects — ArchesWeather and its probabilistic "sister" ArchesWeatherGen.
Both models were developed for ten-day forecasts, and both had to be seriously adapted for completely different time scales. What came of this is described below.
Two approaches to forecasting
Interestingly, the models are designed differently. ArchesWeather is a deterministic system: for each input, it has one output. This approach is fast but does not provide insight into the range of possible atmospheric states. ArchesWeatherGen is a probabilistic model based on flow-matching: it generates a distribution of possible weather scenarios rather than a single answer. At the same time, it uses ArchesWeather's forecasts as a basis for its generation, allowing an ensemble to be built. The ensemble, in turn, provides an uncertainty estimate — important in both weather and climate.
How the models were adapted for climate tasks
It would be naive to expect a weather model to become a climate model on its own. The ocean is critical for climate: it accumulates heat and drives slow processes. Therefore, the researchers equipped both models — ArchesWeather and ArchesWeatherGen — with additional input channels carrying boundary conditions. The model now receives monthly mean sea surface temperature (SST) and sea ice concentration (SIC) and regularly "adjusts" atmospheric dynamics to them.
The runs were performed according to the standardized protocol of the first phase of AIMIP — an analogue of the well-known AMIP project, but for machine learning models. This protocol allows different AI models to be compared with each other and with classical climate models. The authors also conducted ablation experiments: they disabled individual adaptation components to understand what exactly provides stability. Additionally, forced configurations, where the ocean is prescribed externally, were compared with unforced ones, where the model operates without external coupling.
What the multi-year simulations showed
Fears of rapid degradation did not materialize. In the forced configuration, both ArchesWeather and ArchesWeatherGen demonstrated stable multi-year runs: the annual cycle does not break down, and climate "crashes" that kill long simulations were avoided. The models successfully reproduce:
- the basic climatology of the ERA5 reanalysis;
- large-scale atmospheric circulation;
- year-to-year variability;
- rare, extreme states — the tails of distributions.
In other words, average values and the spread of values look plausible not only overall, but also in extreme cases. On a number of parameters, AI simulations are on par with classical numerical climate models. It was separately noted that forced configurations perform noticeably better than unforced ones: without ocean "hints," errors accumulate faster.
Conclusions
The work confirms that machine learning models originally created for weather forecasting can serve as the basis for climate experiments. To do this, it is enough to add external oceanic conditions and carefully follow the comparison protocol. This is a relatively cheap and fast way to obtain long-term simulations, although a full replacement of classical models is still far off — broader validation is needed, especially on regional details and extremes. But the first steps look promising.
