Skill Superiority of ML-Based Climate Emulators over Pattern-Scaling Emulators

Determine whether machine-learning-based climate emulators, such as ClimateBench v1.0 employing Gaussian process regression and neural networks, are more skillful than traditional pattern-scaling emulators that regress regional temperature on global mean temperature or cumulative emissions.

Background

ML-based climate emulators have been proposed to interpolate projections from expensive simulations and estimate climate impacts of anthropogenic emissions. ClimateBench exemplifies such emulators trained with ML techniques to produce annual projections to 2100.

Traditional pattern-scaling emulators, which relate regional climate variables to global mean temperature or cumulative emissions, are well-established. The paper explicitly notes that it remains to be shown whether ML-based emulators offer superior skill relative to these simpler, widely used baselines.