Tackling quantitatively large dimensionality problems

A two-step approach to sensitivity analysis of model output in large comput ational models is proposed. A preliminary screening exercise is suggested i n order to identify the subset of the most potentially explanatory factors. Afterwards, a quantitative method is recommended on the subset of preselec ted inputs. The advantage of the proposed procedure is that, very often, am ong a large number of input factors, only a few have a significant effect o n the model output. The approach provides quantitative sensitivity measures while controlling the computational cost of the experiment. The procedure has been tested on a recent version of a chemical kinetics model of the tro pospheric oxidation pathways of dimethylsulphide, including 68 uncertain fa ctors. (C) 1999 Elsevier Science B.V.