Climate Model Intercomparison with Multivariate Information Theoretic Measures


Earth system models, or climate models, are fundamental tools to understand climate change. The Coupled Model Intercomparison Project (CMIP) provides outputs from many coupled atmosphere-ocean general circulation models for a number of different climate forcing scenarios. CMIP is essential to assess model performance and representativity during the historical period and quantifications of the causes of the spread in future projections. In some aspects, model accuracy has improved significantly over the different CMIP phases, but biases and uncertainties in their projections still remain, and notable differences between models exist. In this work we adopt a novel information theory (IT) perspective for evaluating and comparing climate models. Information content is model independent, and allows us to compare climate model simulations and study their differences in information units. IT measures, such as entropy, total correlation, divergences and mutual information, allow us to potentially encapsulate some interactions and phenomena that each climate model may exhibit. We introduce the rotation-based iterative Gaussianization (RBIG) method to address the inherent problem of high-dimensionality in probability density function estimation and in turn IT measures estimation. The RBIG method is a generative model that is robust to noise and dimensionality and is computationally efficient. We will show intercomparison scenarios between CMIP5 model simulations at a monthly resolution and for key models and variables. We will show assessment of the relative information content, and divergences among models, across time and space. These results provide a better unbiased evaluation of models in addition to being a valid comparative measure of shared information.

Dec 5, 2019 12:00 AM
San Francisco, California, USA