The importance of comprehensive parameter sampling and multiple observations for robust constraint of aerosol radiative forcing

Observational constraint of simulated aerosol and cloud properties is an essential part of building trustworthy climate models for calculating aerosol radiative forcing. Models are usually tuned to achieve good agreement with observations, but tuning produces just one of many potential variants of a model, so the model uncertainty cannot be determined. Here we estimate the uncertainty in aerosol effective radiative forcing (ERF) in a tuned climate model by constraining 4 million variants of the HadGEM3-UKCA aerosol–climate model to match nine common observations (top-of-atmosphere shortwave flux, aerosol optical depth, PM2.5, cloud condensation nuclei at 0.2 % supersaturation (CCN0.2), and concentrations of sulfate, black carbon and organic carbon, as well as decadal trends in aerosol optical depth and surface shortwave radiation.) The model uncertainty is calculated by using a perturbed parameter ensemble that samples 27 uncertainties in both the aerosol model and the physical climate model, and we use synthetic observations generated from the model itself to determine the potential of each observational type to constrain this uncertainty. Focusing over Europe in July, we show that the aerosol ERF uncertainty can be reduced by about 30 % by constraining it to the nine observations, demonstrating that producing climate models with an observationally plausible “base state” can contribute to narrowing the uncertainty in aerosol ERF. However, the uncertainty in the aerosol ERF after observational constraint is large compared to the typical spread of a multi-model ensemble. Our results therefore raise questions about whether the underlying multi-model uncertainty would be larger if similar approaches as adopted here were applied more widely. The approach presented in this study could be used to identify the most effective observations for model constraint. It is hoped that aerosol ERF uncertainty can be further reduced by introducing process-related constraints; however, any such results will be robust only if the enormous number of potential model variants is explored.