CSE Community Seminar | November 13, 2026

Abstract

Earth System Models (ESMs) are the primary tools scientists use to study how Earth’s climate responds to anthropogenic greenhouse gas emissions. However, the simulations are computationally expensive, and the projections suffer from biases relative to observations, especially for weather extremes. ESM emulators offer a cheaper alternative to rapidly explore different future scenarios, but they inherit the biases of the ESMs on which they are trained. We show that a generative ESM emulator, trained on ESM simulations, can be fine-tuned with ERA5 reanalysis data to overcome biases in extreme events. We validate the approach by repeating the emulator’s fine-tuning on one historical run from a high-resolution ESM ensemble and then assessing its performance in simulating end-of-century extremes from the same ESM ensemble. This work suggests that ESM emulators are not only useful for reducing the computational cost of climate projections, but they can also be fine-tuned to improve the accuracy of those projections.