No Epoch Like the Present: Robust Climate Emulation Requires Out-of-Distribution Generalisation
A new research paper highlights the critical need for out-of-distribution (OOD) generalization in climate emulation models. Current machine learning models, while performing well on present-day data, are prone to failure when faced with the inevitable shifts caused by climate change. The study proposes using seasonal variations as a proxy for these long-term shifts and introduces a new evaluation framework to test emulator robustness, finding significant degradation in state-of-the-art models. The paper suggests that compositional generalization, by decomposing physical systems, offers a path toward more reliable ML-driven climate emulators. AI
IMPACT Highlights the limitations of current ML models in predicting future climate scenarios, emphasizing the need for OOD generalization to ensure reliability.