Keynote Talk: Roots and Routes to Robust Climate Projections: Balancing Physical Constraints with Deep Learning

Uncertainties in Earth’s future climate are driven largely by the complex, microscopic interactions between aerosols and clouds, which are computationally expensive to simulate. In this talk, I will explore the "roots" of this challenge by demonstrating a hybrid framework that anchors detailed physical process models with "top-down" observations of historical warming and the Earth's energy budget. This combination allows us to rule out implausible climate futures and significantly narrow projection uncertainty. I will then discuss the "routes" offered by machine learning emulation to accelerate this science, examining the critical trade-offs between traditional pattern scaling and frontier deep learning. Drawing on recent benchmarking work, I will show how the chaotic noise of natural climate variability can confound complex deep learning models, often making simpler, linear methods more robust for predicting temperature. Ultimately, I argue that while AI offers powerful new routes for discovery, we must rigorously benchmark these tools against physical roots to distinguish the climate signal from the noise.