[ESAC Student Seminar] Björn Lütjens

“Machine Learning for Earth System Modeling (EGU practice talk)”

Abstract: Machine learning has been proposed for many tasks across Earth system modeling. With the advent of these new techniques it has become too easy to forget about the quality of existing domain-informed baselines. In this talk, I will first share a cautionary tale in deep learning-based climate emulators, i.e., fast approximations of climate models. I will share how resurfacing a linear regression-based model from the 1980s can outperform existing deep learning models on a popular benchmark and highlight the importance of explicitly addressing noise or internal variability in the climate system. Second, I will share a work-in-progress benchmark dataset for downscaling daily high-resolution (100m) maps of surface meltwater in Greenland and compare domain-informed baselines with deep generative vision-based models.

ESAC Student Seminar Series

A forum for students and postdocs to share recent research, hone presentation skills, and build community among peers, sponsored by the EAPS Student Advisory Committee. Open to current EAPS graduate and undergraduate students and postdocs. Typically hosted on Thursdays during the semester, including pizza lunch.

Contact: esac.officers@gmail.com