A novel active parameter selection strategy for the efficient optimization of combustion mechanisms
Márton Kovács, Máté Papp, Tamás Turányi, Tibor Nagy
Abstract
Optimization of large combustion mechanisms means that a few dozen parameters (called active parameters) are optimized within their uncertainty limits to achieve a better reproduction of the experimental data, which is usually measured by a mean square error function. In previous studies, the active parameters were selected based either on their local sensitivity coefficients (strategy S) or on the products of local sensitivity coefficient and a corresponding uncertainty parameter (strategy SU). This latter measure is known by various names: optimization potential, sensitivity-uncertainty index or uncertainty-weighted sensitivity coefficient. In this work, we proposed three novel active parameter selection strategies of increasing complexity (PCA-SU, PCA-SUE, PCALIN) and demonstrated their superior performance in the optimization of pre-exponential factors (A) in a methanol/NOx combustion mechanism (562 reaction steps of 70 species) against 2360 data measured in shock tube, JSR and flow reactor experiments. The novel methods are based on the principal component analysis (PCA) of sensitivity matrices scaled by the uncertainties of parameters (U) and the uncertainty of the experimental data (E). These PCA-based methods take into account parameter correlations and designate parameter groups and corresponding relevant subsets of experimental data, thereby a factor of 4–7 savings in optimization time was achieved over the S and SU methods. PCA-SUE method performed better than the PCA-SU as it also considered the uncertainty of the experimental data. The PCALIN strategy is similar to PCA-SUE, but it also considers the linear change (LIN) of the error function, which depends on the simulation error of experimental data, and thereby it could provide the most accurate models as a function of the number of active parameters. Based on the PCALIN strategy, fitting all three Arrhenius parameters resulted in further improvements, however, it provided moderate improvements over simple A-factor tuning and required significantly more computer time.