7.2 Development and Application of the SRSM-ADIFOR

In order to further improve the computational efficiency of the SRSM, this method was coupled with a state of the art automated sensitivity analysis method, ADIFOR (Automatic DIfferentiation of FORtran) [17]. ADIFOR produces fast and accurate estimates of first order partial derivatives of model outputs with respect to model inputs or any intermediate quantities in the model code. ADIFOR accomplishes this by rewriting the model code and producing code that calculates first order partial derivatives. Hence, the coupling of the SRSM and the ADIFOR, where the series expansions of the SRSM are used in conjunction with outputs and partial derivatives from the ADIFOR generated code, provides the potential to substantially reduce the computer demands.

The coupled method, SRSM-ADIFOR, was applied to two of the case studies used for evaluating SRSM: a human PBPK model, and the Reactive Plume Model (RPM). The case studies indicate substantial computer time savings; up to two orders of magnitude reductions in the required number of model simulations compared to ``conventional'' methods were achieved. This demonstrates the advantages of combining sensitivity analysis methods with uncertainty propagation methods. The application of the SRSM-ADIFOR involves processing the original model code using ADIFOR, and subsequent use of the derivative code for uncertainty analysis. For this reason, developing a ``black-box tool'' for using the combined SRSM-ADIFOR approach (as was done for the ``stand-alone'' SRSM) is not possible.