Exploring Local Climate Impacts via physics-informed machine learning models.
Björn Lütjens, Matthew Kearney
Thursday, March 3, 2022
How would a 10% reduction in deforestation impact the flood risk at MIT? It is hard to say, as exploring the local impacts of global climate policies requires high-resolution climate projections, that are complex to generate; if naively done, generation with a high-resolution climate model would take multiple days on the world’s largest supercomputers. This talk presents two fundamental methods towards overcoming the computational complexity of high-resolution climate projections. First, Björn will present Matryoshka Neural Operator: a novel machine learning-based partial differential equations (PDE) solver that reduces the complexity of solving select PDEs from quadratic, O(N^2), to quasilinear, O(N logN), by exploiting multiscale relationships. Then, Matt will present initial results on creating an ensemble of high-resolution from low-resolution climate projections via a novel probabilistic deep learning method, called diffusion-based models.