Deep Learning Emulators for Accessible Climate Projections

Abstract

Climate change has shifted from a purely scientific topic to a deeply politicized issue. To combat climate change we need to create mutual understanding on the links between policies, global warming, and city-scale impacts. Climate models have been incredibly helpful in generating this causal understanding, but running them requires supercomputers and is only accessible to the minority of researchers.

This thesis explores how emulating climate models with deep learning can make them more accessible and, at the same time, raise novel challenges in deep learning on physical, long-term time-series, and high-dimensional data. This dissertation shows that deep learning can decrease runtime in dynamical models, increase accuracy in local climate projections, and generate visualizations of climate impacts. Specifically, this thesis contributes a hybrid model, called multiscale neural operator, that corrects fast low-resolution simulations by learning a hard-to-model parametrization term. This achieves to cut runtime complexity from quadratic to quasilinear which can result in a 1000x faster model on selected equations in multiscale dynamics. This thesis also contributes satellite imagery of the future that visualizes climate data using physically-consistent deep generative vision models.

The thesis contributions are framed in an envisioned online tool that rapidly emulates the city-scale impacts of various climate policies. In the future, such an emulator could accelerate local climate risk analyses, attribution of extreme events, and the understanding of causal links between between impacts and policies.