Towards a conditional representation of heat wave probability in large ensemble climate model data

Research Objective: Single-model initial condition large ensembles provide novel opportunities to study the physical drivers and risks of large-scale climate extremes in a changing climate. The probability of extremes such as weekly heatwaves, here quantified as seven-day maximum temperature (Tx7d), are usually approximated with a general extreme value distribution GEV  that is stationary or accounts for non-stationarity of a warming climate. However, estimating the occurrence probability of very rare climate extremes in the presence of large internal variability further benefits from the integration of process-based covariates characterising the preceding and concurrent climate conditions both at global and local scale.

Data & Methods: We here use more than 6000 years of stationary pre-industrial and 2xCO2 control simulations and an ensemble of 84 transient historical and RCP8.5 simulations performed with the Community Earth System Model CESM1.2 to develop and robustly test methods of quantifying extreme events under a broad range of climatic conditions. The generalised extreme value distribution is parametrised such that it can account for changing environmental circumstances, ranging from large-scale thermodynamic non-stationarity due to climate change, regional-scale dynamic forcing such as atmospheric blocking, or local land-surface conditions such as soil moisture deficits. Fields of covariates are integrated using approaches from statistical learning theory, accounting for the spatio-temporal correlation inherent in climate data.

Preliminary results: Dynamical forcing patterns as simulated by the earth system model compare well with those obtained from reanalysis data and inform the statistical model in a physically traceable fashion. How well the latter generalises is tested with respect to further simulations of the US CLIVAR Working Group on Large Ensembles. The relevance of different covariates can inform both detection and attribution as well as risk assessment how their respective statistical models can be further refined to account for the influence of physical drivers under present and future climate conditions.