Nadiya Shore

Topic

Making extremes local: modelling extreme climate events with machine learning downscaling techniques

School of Earth and Ocean Sciences

Date & location

• Monday, August 31, 2026
• 12:00 P.M.
• Clearihue Building, Room B017

Examining Committee

Supervisory Committee

• Dr. Adam Monahan, School of Earth and Ocean Sciences, University of Victoria (Supervisor)
• Dr. Alex Cannon, School of Earth and Ocean Sciences, UVic (Member)
• Dr. Julie Bessac, Computational Science, National Laboratory of the Rockies (Outside Member)

External Examiner

• Dr. Julie Carreau, Department of Mathematical and Industrial Engineering, Polytechnique Montréal

Chair of Oral Examination

• Prof. Malcolm Gaston, School of Public Administration, UVic

Abstract

Extreme climate events are one of the immediate and high-impact consequences of anthropogenic climate change, and are projected to increase in frequency and severity into the 21^st century. Global Earth System Models simulate current climates and project future ones, but are too low-resolution to be useful to local adaptation planning, where communities will directly experience extreme climate events. Downscaling, turning low-resolution (LR) climate fields into high-resolution (HR) fields, is a critical tool that can bridge this gap and properly inform decision makers. Other physics-based downscaling models are too computationally expensive to be implemented on large scales, making new tools, including Machine Learning (ML) models, attractive. This research employs an ML conditional Wasserstein Generative Adversarial Network (cWGAN) model to downscale LR ERA5 reanalysis fields into stochastically generated HR ensembles to assess how well the model can capture near-tail climate extremes in temperature and precipitation. The univariate cWGAN models produce well-calibrated ensembles of downscaled fields that capture characteristic spatial structures with realistic textures. The near-tail percentiles for both temperature and precipitation are skilfully reproduced in the models, as are the complex spatial dependencies across topographic features. While some high-topography biases in the upper temperature extremes were improved by adding snow cover information to the model, the lower tail maintained a larger stubborn warm bias that was not improved by snow cover information or when trained on seasonal-only data. The bivariate cWGAN model downscaling both temperature and precipitation simultaneously produced generally similar extremes as the univariate models, although the visual quality of the precipitation fields was reduced. However, the bivariate model captured dependencies in temperature and precipitation that were not as well-captured in the univariate models. Compound near-tail extremes for the upper tails of both temperature and precipitation were well captured over land but overestimated by the cWGAN model over the ocean. The spatial dependencies of an extreme over space were found to be influenced by the quality of the small-scale features in the generated fields. The cWGAN model produced highly realistic percentiles of near-tail temperature and precipitation extremes and shows promise as an affordable and effective downscaling product.