Bayesian Structure Learning for Climate Model Evaluation

A Bayesian structure learning approach is employed to compare and contrast interactions between the major climate teleconnections over the recent past as revealed in reanalyses and climate model simulations from leading Meteorological Centers. In a previous study, the authors demonstrated a general framework using homogeneous Dynamic Bayesian Network models constructed from reanalyzed time series of empirical climate indices to compare probabilistic graphical models. Reversible jump Markov Chain Monte Carlo is used to provide uncertainty quantification for selecting the respective network structures. The incorporation of confidence measures in structural features provided by the Bayesian approach is key to yielding informative measures of the differences between products if network-based approaches are to be used for model evaluation, particularly as point estimates alone may understate the relevant uncertainties. Here we compare models fitted from the NCEP/NCAR and JRA-55 reanalyses and Coupled Model Intercomparison Project version 5 (CMIP5) historical simulations in terms of associations for which there is high posterior confidence. Examination of differences in the posterior probabilities assigned to edges of the directed acyclic graph provides a quantitative summary of departures in the CMIP5 models from reanalyses. In general terms the climate model simulations are in better agreement with reanalyses where tropical processes dominate, and autocorrelation time scales are long. Seasonal effects are shown to be important when examining tropical-extratropical interactions with the greatest discrepancies and largest uncertainties present for the Southern Hemisphere teleconnections. Climate model biases and performance is typically assessed against observational products via systematic comparison of individual metrics, usually focused on the mean climate, over the recent historical period. We demonstrate how Bayesian structure learning can enable a systematic probabilistic framework for process-based model evaluation of both the temporal behavior of individual climate modes but also to identify and assess the teleconnections between those modes. We show that network structures can be fitted simultaneously and feasibly across a representative sample of climate model simulations affording uncertainty estimation of the robustness of differences across models and observations and robustly identify model biases between teleconnections in the climate. Bayesian structure learning is used to quantify uncertainty in estimated network structures describing climate mode teleconnections Dynamic Bayesian networks estimated from reanalyses are compared to CMIP5 model simulations over the historical period Differences in network structures between models and reanalyses quantify complex interacting biases in climate model dynamics