A Comparison of Methods on Dynamical Adjustment in a Large Ensemble of Climate Simulations

A key issue in climate change research is distinguishing between anthropogenic forcing and natural (internal) climate variability. Attributing specific climate events to a specific cause is difficult for dynamically driven events, and requires an accurate assessment of the role of internal variability. Here we use the CESM large ensemble to remove internal variability associated with circulation dynamics in the extratropics using various methods. The large ensemble consist of 30 simulations run on the exact same model with the same forcing, and thus any differences within the output are solely due to internal variability. We specifically compare and assess four methods that have been used in previous studies: principal component regression (PCR), maximum covariance analysis (MCA), constructed analogs (CA), and partial least squares (PLS). Furthermore, we also introduce and analyze a method that estimates dynamical adjustment using only data from the temperature field. This is a variation of the cold-ocean-warm-land (COWL) method that does not project onto the radiative warming due to increasing greenhouse gases. Methods are applied to both a historical control run and the large ensemble. Our results show that CA and PLS performed best at narrowing the uncertainty due to internal variability within the large ensemble, bringing simulated temperature trends closer to the radiatively forced signal of the ensemble average. Dynamical adjustment also reduces the number of simulations needed to get an accurate representation of the true forced signal. Applying this to observations shows the historical temperature trends in the context of the spread of internal variability and adjusts for much of the difference between observed and simulated trends.