Atmosphere Ocean Science Colloquium

---

Machine Learning for Earth System Modeling

Speaker: Björn Lütjens, Massachusetts Institute of Technology

Location: 60 Fifth Avenue Open Space

Date: Wednesday, April 2, 2025, 3:30 p.m.

Synopsis:

Analysing climate risks with a full-complexity Earth system model (ESM) can take multiple weeks and require computing resources worth hundreds of thousands of USD, for each assumption on future greenhouse gas and aerosol emissions. As a result, scientists and industry practitioners that require custom ESM outputs use 'climate emulators', which are approximate functions that are multiple orders of magnitude faster, but can break down for negative carbon emissions, extreme statistics, and other nonlinear relationships. Recently, large deep learning models have been proposed to learn more accurate emulators. But, I will show how a linear regression model can outperform these novel techniques, including a transformer-based foundation model, on the common benchmark dataset for this task. I will propose an update to the benchmark dataset and share more general recommendations on how to avoid overconfidence in deep learning techniques. One of the recommendations is to augment physics with ML, only when necessary, and I will further illustrate this with a generative vision model that we enhanced with physics-consistency blocks. I will conclude with a roadmap towards establishing hybrid physics-ML emulators as a novel tool in the climate modeling pipeline.