Origins and control of long-term drift attractors in autoregressive ML Earth system models

Characterize the mechanisms that generate long-term drift attractors in autoregressive machine learning Earth System Models using the Spherical Fourier Neural Operator architecture, specifically in the coupled atmosphere–ocean Ola configuration, and develop effective strategies to control or suppress these drifts to enable stable simulations beyond six months.

Background

The authors report substantial drifts at higher latitudes in the Ola coupled model that lead to dynamical instabilities for rollout periods longer than six months, even though similar ML models have demonstrated century-long stability when sufficiently tuned and trained. They emphasize that understanding and mitigating these long-term drift attractors is necessary to advance coupled ML models toward climate-scale simulation capabilities.

References

Understanding the origins of such long-term drift attractors in autoregressive ML, and how to control them, is an open area deserving of systematic empirical testing.

— Coupled Ocean-Atmosphere Dynamics in a Machine Learning Earth System Model  (2406.08632 - Wang et al., 2024) in Section 'Discussion, Limitations, and Future Work'

However, how to model the full temporal evolution of such latent variables while controlling error accumulation remains unresolved.

— Learning Prognostic Variables for AI Convective Parameterizations via Symbolic Distillation  (2609.24882 - Schönfeld et al., 21 Sep 2026) in Section 1, Introduction