Automated sensitivity analysis of an atmospheric dispersion model

Abstract Sensitivity analysis is a necessary component in the evaluation of computer models utilized in support of performance assessment activities. Understanding the behavior of predictive models with respect to input data is important for (1) verifying the validity of a model, (2) determining parameters for which it is important to have accurate values, and (3) understanding the behavior of the system being modeled. This paper describes an automated procedure for performing a comprehensive sensitivity analysis using computer calculus. The procedure employs an automated system called GRESS that utilizes a precompiler to enhance a FORTRAN computer code by adding derivative-taking capabilities. From a single run of an enhanced model, GRESS calculates and reports normalized sensitivities of selected results with respect to all input data. GRESS computes a normalized sensitivity by multiplying a derivative by its associated input enhanced model, GRESS calculates and reports normalized sensitivities of selected results with respect to all parameter value and dividing by the associated output value. The automated GRESS procedure is applied to the code AIRDOS-EPA, which estimates radiation doses caused by atmospheric dispersion of radionuclides around nuclear facilities. The sensitivity calculations for a sample problem show good agreement with those obtained by perturbation analysis. Results demonstrate that an automated procedure can be used cheaply and efficiently to perform a comprehensive sensitivity analysis of existing computer models. A methodology for developing enhanced versions of assessment models that include the capability to report routinely sensitivities for important parameters is discussed. Both GRESS and AIRDOS-EPA are available from the Radiation Shielding Information Center at Oak Ridge National Laboratory, P.O. Box 2008.