OCR Text |
Show It is important to realize the target results are sensitive to parameters other than the kinetics. Assumptions made in the model of the experiment and the model inputs are both important uncertainties that may disqualify some targets. Temperature, flow rate, and one-dimensionality in flame measurements are one example, stretch corrections to flame speed determinations another. A 20K error in shock tube temperature can produce a 2 0 % difference in ignition delay. Besides kinetics, there are also transport and thermodynamics input parameters to the target modeling. Transport treatments(diffusion and thermal conductivity) and parameters significantly affect the predicted flame speeds, especially those for H, H2, N2, and H 2 0. Accurate flame speed predictions require the use of the thermal diffusion and multi-component transport options and about 500 grid points. Recently revised target values (from removing stretch effects) are n o w consistently about 7 % lower than the optimization can predict, perhaps suggesting some correction to the transport is called for by the optimization. In addition to examining the kinetics sensitivities, in this round of the optimization we also computed the target sensitivity to thermodynamics, which enters the models in the computation of reverse rate constants and partial equilibria where they might happen to occur. These results will also be posted for each target on the W e b site. The largest such sensitivity was that of the H C N oxidation targets to the H C N enthalpy. H C N is an intermediate in the reburn chemistry, and its oxidation is accelerated by a forward equilibrium for H C N + H = C N + H2. As a result, w e reevaluated this surprisingly uncertain value and recommend a lower number than the current JANAF-Sandia polynomial.[5] Other thermodynamic sensitivities also on occasion point to the C H 3 + H = C H 2 + H 2 system, and suggest efforts to reduce the enthalpy uncertainty for CH2. RESULTS A few illustrations from the current optimization will demonstrate some features of the process. First, a run with all variables frozen at their baseline starting values shows the initial misses of the targets by the trial mechanism. In the version 3.0 run, the corresponding value for the objective function, G>, the least squares error residual, is 8.7. If all optimization variables are active, this residual m a y be lowered to 1.07. Even if w e freeze the two generally most sensitive reactions, H + 0 2 and O H + C O , only an increase of 0.02 is seen. M a n y other reactions m a y be removed from the optimization without a large change in either <I> or predictions for individual targets. Freezing 20 additional rate constants of the 70 available in the hydrocarbon part of the mechanism only increases G> by 0.16. At the current stage of the optimization, nearing conclusion, there are 28 remaining active rate variables of the total 98, and <I> has only increased to 2.13. Final release of version 3.0 is expected between the due date for this paper and the conference, pending completion of final optimization runs and validation calculations. The dangers of optimizing to a limited data set are easily demonstrated using the optimization procedure. One can, for example, obtain an excellent fit to the ignition delays alone (0=0.01 instead of the usual O>0.10), but the accumulated O for the remaining hydrocarbon targets (excluding propane) is 5.9. While this is an extreme example in that w e have allowed all sensitive rates to vary in order to get the last percentage of agreement, it nonetheless shows how the consideration or availability of a limited data set can produce a mechanism with high errors ' for other applications. One might also use the optimization to ask whether a proposed rate 6 |