In this talk, we will present the LUCIE framework, a family of lightweight, stable, and physically consistent climate emulators designed for long-term autoregressive simulation and rapid scientific experimentation. Beginning with LUCIE-2D, a fully data-driven global climate emulator trained on as little as two years of 6-hourly ERA5 data, we demonstrate that careful architectural choices—such as a hard-constrained first-order integrator, spectral regularization to preserve fine-scale variability, and data-efficient optimization—enable stable 100-year simulations with large ensembles and minimal climatological bias. LUCIE-2D reproduces ERA5 mean fields and variability for temperature, winds, precipitation, and humidity, and supports credible estimation of extreme event statistics and return periods at negligible computational cost, training in just 2.4 hours on a single A100 GPU.
Building on this foundation, we introduce LUCIE-3D, an extension that captures the vertical structure of the atmosphere and responds to external forcings. Using a Spherical Fourier Neural Operator backbone trained on 30 years of ERA5 across eight σ-levels, LUCIE-3D incorporates CO2 as a prognostic forcing variable and optionally accepts prescribed SSTs to emulate coupled ocean–atmosphere interactions. The model reproduces climatological means, modes of variability, and climate-change signals—such as surface warming and stratospheric cooling—while credibly simulating dynamical phenomena including equatorial Kelvin waves, the MJO, and annular modes. Despite its increased dimensionality, LUCIE-3D remains computationally accessible, training in under five hours on four GPUs.
Together, LUCIE-2D and LUCIE-3D provide a versatile platform for fast, controlled, and physically grounded exploration of climate dynamics. Their stability, efficiency, and transparency enable new scientific workflows, from large-ensemble studies of extremes and uncertainty, to ablation experiments probing the role of forcings and boundary conditions, to long-horizon emulation relevant for paleoclimate analysis and future climate projections.
References:
Guan, Haiwen, et al. "LUCIE: A lightweight uncoupled climate emulator with long-term stability and physical consistency." Journal of Advances in Modeling Earth Systems 17.11 (2025): e2025MS005152.
Guan, Haiwen, et al. "LUCIE-3D: A three-dimensional climate emulator for forced responses." arXiv preprint arXiv:2509.02061 (2025).
Copyright. All Rights Reserved.