Over the last two years, large-scale AI weather models have reached or exceeded the accuracy of numerical weather prediction for short- and medium-range forecasts. Those predictions focus on some atmospheric variables, known to have a fast evolution, and lack land-atmosphere coupling.With the advent of foundation models, e.g., Aurora, it has been shown that AI models can be fine-tuned for diverse tasks. But land-surface hydrology has not yet been explored.In this work, we show that Aurora is stable when rolled out for at least two years. Our results also illustrate that seasonal fluctuations and long-term trends are captured by Aurora. Finally, a proof-of-concept experiment indicates that Aurora can be extended to new physics with reduced training costs by training separate decoders.
