LUCIE-3D: A Lightweight Uncoupled ClImate Emulator

We present LUCIE-3D, a data-driven atmospheric emulator that remains stable during autoregressive inference for 100 years with little climate drift. This work builds upon the existing 2-dimensional version of LUCIE (https://arxiv.org/abs/2405.16297), which has demonstrated success with just a few key atmospheric variables. LUCIE-3D was trained on 29 years of coarse-resolution ERA5 data with 4 prognostic variables on 8 vertical levels. Owing to the cheap computational cost of inference, ensembles are executed for 100 years to compute an uncertainty-quantified climatology for the prognostic. Unlike other state-of-the-art machine learning weather models, LUCIE-3D is neither unstable nor does it produce hallucinations that result in unphysical drift of the emulated climate.

The climatology of LUCIE-3D matches ERA5 well, with LUCIE-3D being able to reproduce the general circulation of the atmosphere (Fig. 1) with minimal biases. LUCIE-3D also has a realistic representation of tropical convective processes, including the MJO and Kelvin Waves. The precipitation global bias is 0.4 mm/day and comparable to higher-resolution, modern climate models.

Figure caption: LUCIE-3D zonal climatology, calculated using a 10-year free run, compared to ERA5.