LUCIE: A Lightweight Uncoupled ClImate Emulator with long-term stability and physical consistency for O(1000)-member ensembles

We present LUCIE, a data-driven atmospheric emulator that remains stable during autoregressive inference for a thousand of years with minimal drifting climatology. LUCIE was trained using 9.5 years of coarse-resolution ERA5 data, incorporating 5 prognostic variables, 2 forcing variables, and one diagnostic variable (6-hourly total precipitation), all on a single A100 GPU over a two-hour period. LUCIE autoregressively predicts the prognostic variables and outputs the diagnostic variables similar to AllenAI’s ACE climate emulator. Unlike all the other state-of-the-art AI weather models, LUCIE is neither unstable nor does it produce hallucinations that result in unphysical drift of the emulated climate. The low computational requirements of LUCIE allow for rapid experimentation including the development of novel loss functions to reduce spectral bias and improve tails of the distributions. Furthermore, LUCIE does not impose true sea-surface temperature (SST) from a coupled numerical model to enforce the annual cycle in temperature. We demonstrate the long-term climatology obtained from LUCIE as well as subseasonal-to-seasonal scale prediction skills on the prognostic variables. LUCIE is capable of 6000 years of simulation per day on a single GPU, allowing for O(100)-ensemble members for quantifying model uncertainty for climate and ensemble weather prediction.