My Mila internship on climate downscaling

Paper: arXiv

This paper comes out of my internship with Prof. David Rolnick’s group at Mila. We evaluate deep learning–based climate downscaling or super-resolution for transferability and try to check if models trained on diverse climate datasets can generalize to new regions, variables and even new climate products. We pretrain on multiple climate products and compare CNNs, Fourier Neural Operators (FNOs) and a CNN–ViT hybrid. I presented this work at the ICML 2024 Machine Learning for Earth System Modeling workshop.