Computing and Sustainability Seminar: Patrick Beukema

"EarthSystem: A foundation for solving global crises"

Earth AI is surging in popularity. In addition to myriad academic labs, some of the largest industrial AI labs became involved this year. We even have access to embeddings spanning the whole world at high resolution. But what problems are we solving? Do we need larger foundation models or smaller ones? How would we even know, when everyone uses a different benchmark and evaluation strategy? Should we bridge the gap between weather, climate, and earth-observation data? There are many questions and a lot of noise. In this talk, I will try to pierce through that noise by anchoring on why we are building these models in the first place. I will present a new open source platform at Ai2 that was built in collaboration with many partners who need this technology to solve urgent and complex problems like food security, wildfires, and deforestation. I will show how these models can be deployed to provide cost effective and continual, real-time, global coverage. I'll explore whether a universal foundation modeling approach can ever work for planetary-scale problems, and propose ways academia, the non-profit sector, and frontier labs can collaborate more effectively. By grounding our discussion in concrete use cases rather than abstract capabilities, we can better align technological development with societal needs.

Patrick Beukema's current work focuses on developing AI systems that address global humanitarian and environmental challenges. Prior to Ai2, he developed real-time brain-decoding models for patient populations at an Applied Physics Laboratory (Johns Hopkins) spinout. His earlier work spanned neuroscience, biomedical research, and applied machine learning, with support from NIH, Meta, and Ai2. He holds a Ph.D. in neuroscience from Carnegie Mellon and the University of Pittsburgh, an M.S. in Logic from Carnegie Mellon, and a B.A. in from McGill University.