Unlike many ML applications, where out-of-distribution behaviour is an undesirable edge case, in climate modelling it is the problem itself.
By training ML models on historical data and tasking them with rolling out a future, unknown, and changed climate, we guarantee that the model is faced with significant distribution shift. As a result, understanding how these models will perform in such a scenario is vital. But how can we do this when we don't have a test set for the future climate?
In this work, we take inspiration from the field of emergent constraints to use seasonal shift as a proxy to investigate the robustness of ML models to distribution shift. We find a systematic degradation of performance with increasing shift, highlighting that the current ML toolbox is not sufficient for the task of climate emulation.
Taking inspiration from a more standard numerical approach, we explore how learning regime-specific experts can improve generalisation whilst preserving the non-linear expressiveness that makes ML attractive for this task.
We hope this work motivates a stronger focus on robustness in future model designs and presents a potential approach to achieving more robust models.
Check out the preprint on arXiv (https://arxiv.org/abs/2605.22248)