Comparison of global and local sensitivity techniques for rate constants determined using complex reaction mechanisms

Many rate constant measurements, including some "direct" measurements, invo lve fitting a complex reaction mechanism to experimental data. Two techniqu es for estimating the error in such measurements were compared. In the firs t technique, local first-order elementary sensitivities, were used to rapid ly estimate the sensitivity of the fitted rate constants to the remaining m echanism parameters. Our group and others have used this technique for erro r estimation and experimental design. However, the nonlinearity and strong coupling found in reaction mechanisms make verification against globally va lid results desirable. Here, the local results were compared with analogous importance-sampled Monte Carlo calculations in which the parameter values were distributed according to their uncertainties. Two of our published rat e measurements were examined. The local uncertainty estimates were compared with Monte Carlo confidence intervals. The local sensitivity coefficients were compared with coefficients. from first and second-degree polynomial re gressions over the whole parameter space. The first-order uncertainty estim ates were found to be sufficiently accurate for experimental design, but we re subject to error in the presence of higher order sensitivities. In addit ion, global uncertainty estimates were found to narrow when the quality of the fit was used to weight the randomly distributed points. For final resul ts, the global technique was found to provide efficient, accurate values wi thout the assumptions inherent in the local analysis. The rigorous error es timates derived in this way were used to address literature criticism of on e of the studies discussed here. Given its efficiency and the variety of pr oblems it can detect, the global technique could also be used to check loca l results during the experimental design phase-The global routine, coded us ing SENKIN, can easily be extended to different types of data, and therefor e can serve as a valuable tool for assessing error in rate constants determ ined using complex mechanisms. (C) 2001 John Wiley & Sons, inc.