Earth on a chip: AI-based autoregressive models of the Earth system

Recent advances in machine learning offer transformative potential for Earth system modeling, yet purely data-driven climate and weather emulators face critical challenges of instability and unphysical behavior when integrated over long timescales. This talk presents a unified framework for AI-based autoregressive Earth-system models that are stable, physically consistent, and computationally efficient. It begins with a theoretical eigenanalysis of neural autoregressive models, establishing how inference-time stability is linked to the spectral properties of the model’s Jacobian and how architectural constraints can suppress unstable dynamical modes. Building on this foundation, the root cause of long-term instability - spectral bias in deep networks under-representing high-wavenumber, fine-scale dynamics - is identified, and remedies via higher-order integration schemes and spectral regularization are demonstrated. The talk culminates with the LUCIE 3D climate emulator, a lightweight data-driven model trained on reanalysis data that accurately captures broad climatological statistics, forced climate responses, and variability, while remaining stable over multidecadal simulations. Together, these developments chart a rigorous pathway from mathematical theory to trustworthy AI Earth-system emulation, capable of supporting long-term climate studies and estimation of extremes

Speaker: Ashesh Chattopadhyay, UC Santa Cruz