[en] In an integrated impact assessment, one has to test several scenarios of the model inputs or/and to identify the effects of model input uncertainties on the model outputs. In both cases, a large number of simulations of the model is necessary. That of course is not feasible with comprehensive Chemistry-Transport Model, due to the need for huge CPU times. Two approaches may be used in order to circumvent these difficulties: The first approach consists in reducing the computational cost of the original model by building a reduced model. Two reduction techniques are used: the first method, POD, is related to the statistical behaviour of the system and is based on a proper orthogonal decomposition of the solutions. The second method, is an efficient representation of the input/output behaviour through look-up tables. It describes the output model as an expansion of finite hierarchical correlated function in terms of the input variables. The second approach is based on reducing the number of models runs required by the standard Monte Carlo methods. It characterizes the probabilistic response of the uncertain model output as an expansion of orthogonal polynomials according to model inputs uncertainties. Then the classical Monte Carlo simulation can easily be used to compute the probability density of the uncertain output. Another key point in an integrated impact assessment is to develop strategies for the reduction of emissions by computing Source/Receptor matrices for several years of simulations. We proposed here an efficient method to calculate these matrices by using the adjoint model and in particular by defining the 'representative chemical day'. All of these methods are applied to POLAIR3D, a Chemistry-Transport model developed in this thesis. (author)