SFA IV: Climate modelling

Thunderstorms organize themselves: rain evaporating beneath a storm cloud creates pools of cold air that spread out and trigger new storms nearby. Through this chain reaction, individual storm cells just a few kilometers wide can grow into storm clusters spanning hundreds of kilometers – systems that drive tropical circulation, cause extreme rainfall and flooding, and can even seed hurricanes. Standard climate models are too coarse to capture the crucial small-scale interactions, and simulating everything in full detail is computationally unaffordable. This project therefore develops simplified physical models and generative AI that learn the small-scale storm interactions from ultra-fine simulations and represent them in coarser models. It is multi-scale because it connects kilometer-sized cold-air pools, over hours, with global circulation patterns evolving over months and years, and interdisciplinary because it combines atmospheric physics, computational science, and machine learning.

Some parts of the climate system – like the Gulf Stream system, the Greenland Ice Sheet, or Arctic sea ice – can pass "tipping points": once pushed too far, they change irreversibly. Because such tipping unfolds over centuries, it has not been seen as an immediate threat. This project asks a new question: what happens to fast, small-scale weather fluctuations while such a slow, large-scale tipping process is underway? Using theory and climate models, the team will trace how the tipping of huge climate components translates into changed day-to-day weather variability and extremes that affect people, agriculture, and infrastructure. The project is multi-scale because it links processes that take centuries and span oceans with weather events lasting hours, and interdisciplinary because it connects climate physics, mathematics of instabilities, and climate-impact research – with possible parallels even to turbulence in astrophysics.

Global climate models divide the atmosphere into grid boxes tens of kilometers wide – but crucial processes like cloud formation and turbulence happen on much smaller scales and must be represented approximately. These approximations ("parameterizations") are a major source of uncertainty in climate projections. This project replaces them with generative AI – the same class of models behind modern image generators – trained on very-high-resolution simulations and built directly into a climate model that is fully "differentiable", meaning the physics and the AI can be trained together as one system while respecting conservation laws. It is multi-scale because it teaches coarse global models what unresolved small-scale weather is doing, and interdisciplinary because the approach applies wherever fluids are simulated on grids – from Earth's atmosphere to gas flows in galaxies, connecting climate science, machine learning, and astrophysics.