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

Haiwen Guan, Troy Arcomano, Ashesh Chattopadhyay, Romit Maulik

TL;DR

LUCIE is a lightweight, data-driven climate emulator that achieves 100-year simulation stability and physical consistency while training on as little as two years of data. By combining a Spherical Fourier Neural Operator (SFNO) with a hard-constrained Euler integrator and a novel spectral regularization strategy, it runs efficiently on a single GPU, enabling massive ensemble simulations at a fraction of the cost of traditional models.

What Problem Does This Paper Address?

While recent AI-driven weather models (like GraphCast or Pangu-Weather) match or beat traditional numerical weather prediction models for short-term forecasts, they struggle with long-term climate emulation. When integrated over months or years, these models frequently suffer from "spectral bias"—becoming numerically unstable, drifting unphysically near the poles, or becoming excessively diffusive (blurring out fine-scale atmospheric features).

Furthermore, these state-of-the-art AI models are massive, computationally expensive to train, and require decades of high-resolution training data. The authors introduce LUCIE (Lightweight Uncoupled ClImate Emulator) to bridge this gap, providing a low-complexity, physically consistent model that can be trained quickly and used as a "sandbox" for large-ensemble climate experiments.

Methods

LUCIE builds upon the Spherical Fourier Neural Operator (SFNO) architecture to naturally handle the Earth's spherical geometry. To overcome the stability and data-scarcity challenges of long-term emulation, the authors introduce three key innovations:

  1. Euler Integration as a Hard Constraint: Instead of directly predicting the atmospheric state at the next time step, LUCIE predicts the tendency (the rate of change) of the prognostic variables. It then uses a built-in, first-order Euler integration scheme to compute the next state, which significantly suppresses autoregressive error growth.
  2. Spectral Regularization in the Tendency: The loss function includes a novel penalty on the Fourier spectrum of the prognostic variables' tendencies, specifically over the extra-tropics. This prevents the model from losing power at high wavenumbers (fine spatial scales) over long integrations.
  3. Aggregated Gradient Method (AGGM): To enable training on highly limited datasets without causing the model to blow up during inference, AGGM accumulates gradients across the entire training batch over an epoch rather than resetting them after every batch.
Figure 1: A schematic of LUCIE show the training stage of the model (left panel), in- ference stage for an individual ensemble member (top right), and global annual mean temperature for T0.95 for the the first 10 years of inference and ERA5 from 2010-2019 for reference (bottom right).

Datasets

LUCIE is trained on the ERA5 reanalysis dataset, coarse-grained and regridded to a T30 Gaussian grid (roughly 3.75° horizontal resolution) across specific vertical $\sigma$-levels.

LUCIE is highly data-efficient, utilizing a minimal state vector (~2GB of data) split into three categories:

LUCIE Variable Breakdown

For evaluation, the authors compare LUCIE against a 40-year free run of SPEEDY, a simplified, physics-based atmospheric general circulation model operating at the same horizontal resolution.

Key Insights and Ideas

  • Data Scarcity is Not a Barrier: Traditional AI climate emulators rely on decades or centuries of simulation data. LUCIE proves that with the right optimization strategy (AGGM), a data-driven model can learn a stable, long-term climate from just 2 to 10 years of 6-hourly data.
  • Targeting Tendencies Matters: Penalizing the Fourier spectrum of the rate of change (tendency) rather than the full atmospheric state is highly effective at maintaining physical consistency and preventing the typical spectral decay seen in pure ML models.
  • Hybrid Inductive Biases Work: Embedding simple numerical integration directly into the neural network architecture acts as a powerful constraint that keeps long-term simulations bounded.

Key Results

  • Computational Efficiency: LUCIE trains in just 2.4 hours on a single A100 GPU, making it accessible for rapid prototyping and sandbox experimentation.
  • Century-Long Stability: The model successfully completed 100-year autoregressive simulations across 100 ensemble members without numerical blow-ups or unphysical drift.
  • Outperforming Physics Baselines: LUCIE reduced root-mean-squared biases in long-term climatology by at least a factor of two compared to the physics-based SPEEDY model.
  • Capturing Variability and Extremes: Unlike SPEEDY, LUCIE accurately captured the diurnal cycle of near-surface temperature. It also successfully emulated out-of-distribution extreme weather events ("gray swans") and estimated their return periods, despite these events being absent from its short training dataset.

Limitations and Future Work

LUCIE is an uncoupled emulator, meaning it lacks an explicit ocean or sea-ice component. Consequently, it fails to capture massive year-to-year climate variations driven by sea-surface temperatures, such as the El Niño-Southern Oscillation (ENSO). It also exhibits a dry bias over regions like the Maritime Continent, which the authors attribute to its limited vertical resolution.

Additionally, because it generates a stable, stationary climate, it cannot currently simulate non-stationary climate change. The authors note that future iterations will focus on expanding to a full 3D atmosphere, adding vertical levels, incorporating atmospheric chemistry (like greenhouse gases), and coupling the system with a slab ocean emulator.

Summary updated

Abstract

We present a lightweight, easy-to-train, low-resolution, fully data-driven climate emulator, LUCIE, that can be trained on as low as $2$ years of $6$-hourly ERA5 data. Unlike most state-of-the-art AI weather models, LUCIE remains stable and physically consistent for $100$ years of autoregressive simulation with $100$ ensemble members. Long-term mean climatology from LUCIE's simulation of temperature, wind, precipitation, and humidity matches that of ERA5 data, along with the variability. We further demonstrate how well extreme weather events and their return periods can be estimated from a large ensemble of long-term simulations. We further discuss an improved training strategy with a hard-constrained first-order integrator to suppress autoregressive error growth, a novel spectral regularization strategy to better capture fine-scale dynamics, and finally an optimization algorithm that enables data-limited (as low as $2$ years of $6$-hourly data) training of the emulator without losing stability and physical consistency. Finally, we provide a scaling experiment to compare the long-term bias of LUCIE with respect to the number of training samples. Importantly, LUCIE is an easy to use model that can be trained in just $2.4$h on a single A-100 GPU, allowing for multiple experiments that can explore important scientific questions that could be answered with large ensembles of long-term simulations, e.g., the impact of different variables on the simulation, dynamic response to external forcing, and estimation of extreme weather events, amongst others.
BibTeX
@article{guan2025lucie,
  author = {Haiwen Guan and Troy Arcomano and Ashesh Chattopadhyay and Romit Maulik},
  title = {{LUCIE: A Lightweight Uncoupled ClImate Emulator with long-term stability and physical consistency for O(1000)-member ensembles}},
  journal = {Journal of Advances in Modeling Earth Systems},
  year = {2025},
  doi = {10.1029/2025ms005152},
}
Figure