Deep Learning Climate Model Emulation with PEAR - Adapting a Transformer-Based Weather Forecasting Model onthe HEALPix Grid to the Climate Domain

Abstract

Climate model emulation aims to approximate the forcing–response behaviour of computationally expensive Earth system models at a fraction of their original cost. This thesis adapts PEAR, a HEALPix-based Volumetric Swin Transformer originally developed for medium-range weather prediction on the sphere, to the task of climate model emulation. Instead of predicting future atmospheric states from initial conditions, the adapted model learns to map greenhouse-gas and aerosol emission forcings to global fields of near-surface air temperature and precipitation. The model is evaluated on ClimateSet, a machine-learning-ready dataset containing processed CMIP6 outputs from multiple Earth system models, and benchmarked against UNet and ClimaX baselines from the ClimateSet benchmarks. Across 15 climate models and 5 random seeds, PEAR achieves competitive performance despite its small size, with 4.3M trainable parameters compared with 18.3M for UNet and 106.6M for ClimaX. Although the larger ClimaX model obtains the strongest overall results, PEAR performs favourably compared with UNet, outperforming it on 13 of the 15 climate models in overall RMSE on normalized labels. The strongest results are found for precipitation, where PEAR achieves the lowest RMSE on a majority of the evaluated models. These results suggest that PEAR provides a promising parameter-efficient architecture for climate model emulation, with further potential for improvement through architectural, temporal, and capacity-related refinements.