Climate models often operate on resolution which is too coarse to capture relevant atmospheric processes. We will investigate and develop novel machine learning methodologies accounting for the multiscale and multiresolution structures. The goal is to construct scalable techniques directly applicable to large datasets arising from satellites, as well as in the context of physical emulators to quantify uncertainty in model simulations of aerosol-cloud effects on climate. The new approaches to scalable and expressive multiresolution Gaussian Processes, coupled with increasingly accurate observations from satellites, promise improving the understanding of complex variable interactions and further refinements to physical models.
Dino Sejdinovic (supervisor), Athanasios Nenes (co-supervisor), Javier Gonzalez Hernandes (Amazon, non-academic advisor), Nial Robinson (Met Office, non-academic advisor)
4 months at Amazon Cambridge, 2.5 months at Met Office; 2.5 months at EPFL in Lausanne
Enrolment in Doctoral degree
PhD in Statistics at University of Oxford