Simplifying Earth System Projections: Mimicking ESM Results with a Diffusion Model

Ensemble simulations using Earth System Models (ESMs) have historically been used to gain insights into future climate scenarios. However, they present notable disadvantages, particularly their long computing times and the high technical threshold required for accessibility. The recent rise of data-driven approaches offers a promising alternative, making long-term climate projections more efficient, accessible to policymakers and regional planners, and scalable for specific regions.

During the Winter School “Data-Driven Modeling and Predictions of the Earth System,” we compared the results of a simple diffusion model with the ensemble results from the CESMv.2.1.5 Large Ensemble from model year 2015 to 2090. The diffusion model, trained on CESM data, uses only CO₂ concentration and the month of the year as context channels to predict spatially-resolved, monthly averaged air temperature, precipitation, and atmospheric pressure on a global scale. The project aimed to demonstrate how effectively the diffusion model simulates global and regional variability and long-term trends in these atmospheric variables compared to the ESM. Particular attention was given to its representation of the El Niño–Southern Oscillation (ENSO) region. Additionally, a bias correction was applied to the diffusion model results against the ESM to evaluate distortions in trends and variability.

The study concluded that even a simple diffusion model has significant potential for predicting meteorological parameters based solely on projected greenhouse gas emissions and the time of year. However, its performance weakened near the poles in reproducing ESM results, highlighting the importance of incorporating additional geographic variables (e.g., grid cell size) during training. Despite these limitations, combining the strengths of coupled ESMs with diffusion models can leverage the physical accuracy of ESM outputs and the computational efficiency and adaptability of diffusion models, offering a more comprehensive understanding of Earth system dynamics.