AstroAI Lunch Talks - April 29, 2024 - Björn Lütjens
29 Apr 2024 - Joshua Wing
The video can be found here: https://www.youtube.com/watch?v=-3vKe1nO2SE
Speaker: Björn Lütjens, Department of Earth, Atmospheric, and Planetary Sciences, MIT
Title: A Cautionary Tale about Deep Learning-based Climate Emulators
Abstract: Climate models are computationally very expensive for exploring the impacts of climate policies. For example, simulating the impacts of a single policy emission scenario can take 20.000 CPU hours or the equivalent of USD 300K in cloud compute. Compellingly, deep learning models can now forecast the weather in seconds rather than hours in comparison to conventional weather models, and being proposed to achieve similar reductions by approximating climate models. Climate approximations or emulators, however, have already been developed since the 1990s and I will present how we created a simple linear regression model that outperforms a novel 100M-parameter transformer-based deep learning model on the most common climate emulation benchmark. I will use this result to discuss how internal climate variability influences model performance and highlight the need for domain-informed baselines in applied AI projects.
Watch the talk below!