Scalable Composite Transformations for Generative Climate Model Emulation

Johannes Brachem, University of Göttingen

Co-authors: Paul F. V. Wiemann, Ohio State University; Matthias Katzfuss, University of Wisconsin–Madison

Abstract: We address the challenge of modeling high-dimensional, non-Gaussian spatial fields based on limited – but more than one – observed samples, a common problem in climate science and geostatistics. Inspired by copula modeling, we combine a Bayesian transport map (Katzfuss & Schäfer, 2023) with flexible marginal models to construct a highly expressive joint model.

Using only 20 observations, our method outperforms previous models fitted to 80 observations, as measured by predictive log scores on holdout data.