OCR Text |
Show The first step in a round of optimization was identification of active parameters, i.e., the rate coefficients that have the strongest effects on responses (the computed values of ignition delays, concentration profiles and flame speeds). For this purpose screening sensitivity analyses were perfonned. Inspection of the results confirmed as expected that rate coefficients in the trial mechanism could be divided into two classes: active parameters, those which have effects on the responses that are larger than the scatter of the corresponding experimental data, and passive ones, whose effects are comparable to the experimental scatter. From the highest-ranked reactions 29 rate coefficients were chosen, based on sensitivity values and the ranges of uncertainty, as the variables for optimization in the first round The pre-exponential factors of the pressure-independent bimolecular reactions and the rate coefficient multipliers of falloff expressions for the pressuredependent unimolecular reactions were the actual variables optimized. Second-order polynomial approximations to response surfaces 11t = bO+ LbiXi + LbijXiXj ( 1 ) i i<j were developed for the targets. The summations in (1) are taken over all target l1t data values. The variables X in (1), called factorial variables, are defined as X _ In I~~k" (2) , , for each Xi, where k = k kO' k is the central point of the design for computer experiments, k the offset of k' from kO' kO the rate parameter of the trial mechanism, and! the span of parameter k. A search for minima of the objective function comprising the squared residual sum derived from these response surfaces and experimental measurements was done using the computer program ZXMWD of the International Mathematical & Statistical Libraries'[S] which determines the global minimum by finding, using a quasi-Newton method, all local minima within the hyperrectangle of imposed constraints. PERFORMANCE OF GRI-MECH 1.2 The optimization procedure converged to predict the target data set well, in addition to other data not a part of the target set. A representative set of optimization and validation data and model results is shown in Fig. 1-6; a larger set may be found at the World Wide Web site referenced earlier. (Dashed lines where shown are the results of the initial GRIMech 1.1 model.) No systematic trends or deviations with mixture composition or other 4 |