The Experimental Design Based Uncertainty Analysis Of Combustion Kinetic Models

It is of vital importance for the development of combustion science to improve the predicting ability of combustion kinetic models over a wide range of conditions.Uncertainty quantification(UQ),which is the key method to control the uncertainties of the combustion kinetic models,is to use the mathematical techniques to investigate the sources of and the propagation of model uncertainty and to reduce the parameter uncertainty through model optimization method.However,UQ methods are still hard to be applied on the analysis of complicated combustion kinetic models,partly because the computational efficiency of the current UQ method is still quite low,and the experimental conditions and practical conditions are not matching.The aim of this work is to apply the UQ methods on the analysis of complicated combustion kinetic models and solve the difficulties of this process.The Bayesian model optimization method is improved using the surrogate model of artificial neural network(ANN).The test errors of several popular mathematical surrogate models in UQ of combustion model have been compared,and results show that ANN has higher efficiency than other methods.The ANN-MCMC optimization method is developed based on the method of Markov chain Monte Carlo(MCMC),which improves the efficiency of solving the Bayesian equation.As shown by the example in this work,the ANN-MCMC method needs only 700 samples to achieve convergence,but the traditional MCMC method needs more than 10000 samples to reach a similar accuracy.The experimental design method of surrogate model similarity is developed,which is to measure the potential relationship between the combustion kinetic model predictions at different conditions and provide guidance for the selection of experimental conditions.The similarity coefficient can be calculated using this method,and the model predictions at the conditions with larger similarity coefficients are proved to be influenced by similar reactions.In addition,if the similarity coefficients of two conditions are large,the experiment data measured at one condition can constrain the model predictions at another condition.Based on the work mentioned above,the analysis framework of UQ based on the experimental design method is proposed.The computational platform Combustion UQ is developed,which is applied to the UQ analysis of complicated combustion kinetic models.The basic aims of this method are to reduce the predicting uncertainties of combustion model in actual conditions,quantify the predicting uncertainty and the source of uncertainty,find the conditions which have the most significant constraint on the model parameters,and inversely optimize the model parameters using experimental data.The combustion reaction systems of cyclohexene and 1,5-hexadiene,diacetyl and methanol are employed as examples,which show how to complement the combustion model structure,how to conduct model analysis and experiment design method,and how to constrain the combustion model and finally get the optimized model parameters.