Machine Learning for Trustworthy Climate Emulators
- Date
- Friday 10 February 2023, 8am - 5pm
Friday 10th February 2023
2-3pm
Hybrid / Room 11.87
Speaker: Björn Lütjens, PhD candidate (MIT)
Decision-makers in industry are grappling with climate change and ask for local climate risk analyses. Local analyses, however, are largely inaccessible, because running Earth system models at the local scale suffers from the curse of dimensionality and climate datasets can reach the size of petabytes. This talk will present how our work in physics-informed machine learning can alleviate this issue by developing the displayed trustworthy climate emulators. Emulators are copies, surrogates, or reduced-order models that are orders of magnitudes faster than climate models, but ML-based emulators must exploit physical knowledge to be trusted. First, I will explain how multiscale physics are inherent in climate data and can be exploited via multiscale neural operators. Then, I will illustrate how diffusion-based models can exploit the probabilistic nature of downscaling Earth system models, in the case of Greenland surface meltwater maps.
Bio:
Björn Lütjens is a PhD Candidate at MIT and his research is tackling climate change with machine learning, little-by-little, together with Prof. Dava Newman, Cait Crawford, and Chris Hill. His research has won grants by NSF, Climatechange.ai, ESA, Portugal Space, NASA, IBM, MIT MSCS, MIT Pkg, MIT Legatum and compute grants by Microsoft and NVIDIA. He has worked with IBM Future of Climate and BRT (John Deere), advised two teams of senior researchers at the NASA SETI Frontier Development Lab, and co-founded the ForestBench Consortium for creating a public forest carbon dataset. He has earned his M.Sc. from MIT in robust deep reinforcement learning and a B.Sc. from TUM in engineering science, which positions him at the intersection of machine learning and computational science. He also windsurfs poorly, jams, jokes, and loves meeting new people.