Atmospheric Environment 37 (2003) 3029–3038 A sensitivity analysis study for radm2 mechanism using automatic differentiation Rafik Djouad a , Nicole Audiffren b , Bruno Sportisse a, * a Centre d’Enseignement et de Recherche sur l’Environnement Atmosph ! erique, Ecole Nationale des Ponts et Chauss ! ees (ENPC-CEREA), avenue Blaise Pascal, 77455 Champs sur Marne, France b Laboratoire de M! et ! eorologie Physique CNRS/OPGC, universit ! e Blaise Pascal 24, avenue des Landais, 63177 Aubi " ere Cedex, France Received 28 January 2002; received in revised form 4 April 2003; accepted 14 April 2003 Abstract A sensitivity analysis of an atmospheric multiphase mechanism is performed using an automatic differentiation tool. The sensitivity of some key concentrations is computed with respect to some input parameters (kinetic rates, microphysical parameters). The package odyssee is used in order to obtain the so-called linear tangent model giving the derivatives of outputs with respect to inputs. The direct model takes into account gas-phase reactions, aqueous-phase reactions and interfacial mass transfer and is based on the radm2 mechanism. Local sensitivity coefficients are computed for two different scenarii, rural and sub-urban. We focus in this study on the sensitivity of the gas-phase O 3 – NO x –HO x system with respect to some aqueous phase reactions and we investigate the influence of the reduction in the photolysis rates in the area below the cloud region. This preliminary work illustrates how powerful automatic differentiation tools may be for the study of large chemical mechanisms. We show for instance that the oxidation of trace metals (Fe II ; Fe III ; Cu þ and Cu 2þ ) in the case of low S(IV) polluted area is not always in disfavor of HO x gaseous concentrations, as it is usually claimed. r 2003 Elsevier Science Ltd. All rights reserved. Keywords: Sensitivity analysis; Multiphase chemical mechanism; radm2; Automatic differentiation; Box model 1. Introduction Comprehensive atmospheric multiphase models may contain a large set of chemical species, reactions, thermodynamical and microphysical parameters. In addition to uncertainties introduced by meteorological data such as the wind field and temperature, there are many other sources of uncertainties in such models, either at the structural level (the choice of chemical reactions that have to be taken into account) or for the values of the model parameters (Pielke, 1998; Seinfeld and Pandis, 1998; Carmichael et al., 1996). An assess- ment of the relative importance of each model parameter is therefore a key issue for evaluating the validity of such models. A first approach is the ‘‘global sensitivity analysis’’ that may be performed through Monte Carlo simula- tions (Derwent and Hov, 1988; Moore and Londergan, 2001): the probability density function (PDF) of the outputs is then computed on the basis of the PDFs of the inputs (if known). The drawbacks of this approach are that it is time consuming and that the PDFs of the inputs are usually poorly known. A second approach is the ‘‘local sensitivity analysis’’: it consists in computing the derivative of the outputs with respect to the inputs. Even if the resulting sensitivity coefficients have only a local validity (for nonlinear models), it may improve the knowledge of the model limitations. We focus on this approach in this article. ARTICLE IN PRESS *Corresponding author. Tel.: +33-1-6415-3577; fax: +33-1- 6415-3764. E-mail address: sportiss@cereve.enpc.fr (B. Sportisse). 1352-2310/03/$-see front matter r 2003 Elsevier Science Ltd. All rights reserved. doi:10.1016/S1352-2310(03)00322-4