A Bayesian hierarchical model for spatial-temporal assessment of climate model biases

Abstract:

Climate model biases are systematic errors affecting geophysical quantities simulated by coupled general circulation models and Earth system models against an observational target. To this regard, sea-surface temperature (SST) biases are a major concern due to the central role of SST properties for the dynamical coupling between the atmosphere and the ocean, and for the associated variability. Strong SST biases can be detrimental for the overall quality of historical climate simulations, they contribute to uncertainty in simulated future climate scenarios and complicate the initialization and assessment of decadal climate prediction experiments.

Inter-hemispheric and inter-basin connections are apparent between climatological SST biases. They often resemble the imprint on SSTs of dominant large-scale oceanic and atmospheric phenomena. Still, climatological values only provide a static description of a phenomenon, and the hypothesis of teleconnections between regional model biases requires further substantiation. We propose a dynamic linear model developed within a Bayesian hierarchical framework for probabilistic assessment of spatial and temporal characteristics of SST biases in ensemble climate simulations. In our formulation, the statistical model distinguishes between seasonal and longer-term bias components. In the considered Bayesian framework, the conditional estimation of bias components and their evolution accounts for uncertainty in model Gaussian error parameters through sampling of the posterior distribution of associated variances by a Monte Carlo Markov Chain.

In this contribution we illustrate a first application of the model using ensembles of decadal hindcasts initialized over the period 1960-2000 CE and performed with the MiKlip prototype system for decadal climate predictions. We focus on the tropical Atlantic Ocean – a region where climate models are typically affected by a warm SST bias - to demonstrate how our approach allows for a more reliable estimation of model biases, and for a more efficient identification of associated sources of heterogeneity, non-stationarities and propagation pathways.