Efficient screening of climate model sensitivity to a large number of perturbed input parameters

Modern climate models contain numerous input parameters, each with a range of possible values. Since the volume of parameter space increases exponentially with the number of parameters N, it is generally impossible to directly evaluate a model throughout this space even if just 2–3 values are chosen for each parameter. Sensitivity screening algorithms, however, can identify input parameters having relatively little effect on a variety of output fields, either individually or in nonlinear combination. This can aid both model development and the uncertainty quantification (UQ) process. Here we report results from a parameter sensitivity screening algorithm hitherto untested in climate modeling, the Morris one‐at‐a‐time (MOAT) method. This algorithm drastically reduces the computational cost of estimating sensitivities in a high dimensional parameter space because the sample size grows linearly rather than exponentially with N. It nevertheless samples over much of the N‐dimensional volume and allows assessment of parameter interactions, unlike traditional elementary one‐at‐a‐time (EOAT) parameter variation. We applied both EOAT and MOAT to the Community Atmosphere Model (CAM), assessing CAM's behavior as a function of 27 uncertain input parameters related to the boundary layer, clouds, and other subgrid scale processes. For radiation balance at the top of the atmosphere, EOAT and MOAT rank most input parameters similarly, but MOAT identifies a sensitivity that EOAT underplays for two convection parameters that operate nonlinearly in the model. MOAT's ranking of input parameters is robust to modest algorithmic variations, and it is qualitatively consistent with model development experience.