Generative multi-fidelity modeling and downscaling via autoregressive Gaussian processes

Oregon Convention Center 

Computer models are often run at different fidelities or resolutions due to tradeoffs between computational cost and accuracy. For example, global circulation models can simulate climate on a global scale, but they are too expensive to be run at a fine spatial resolution. Hence, regional climate models (RCMs) forced by GCM output are used to simulate fine-scale climate behavior in regions of interest. We propose a highly scalable generative approach for learning high-fidelity or high-resolution spatial distributions conditional on low-fidelity fields from training data consisting of both high and low-fidelity output. Our method learns the relevant high-dimensional conditional distribution from a small number of training samples via autoregressive Gaussian processes with suitably chosen regularization-inducing priors. We demonstrate our method on simulated examples and for emulating the RCM distribution corresponding to GCM forcing using past data, which is then applied to future GCM forecasts

Gaussian Processes

Spatial Fields

Downscaling

Bayesian Transport Map

Climate-model emulation

Generative Modelling 

Section on Statistics and the Environment