Feature Extraction Based Uncertainty Quantification Of Combustion Kinetic Models

The numerical simulation of the combustion process is of great significance for the development of advanced power devices and the achievement of the“dual carbon”goal,and accurate and reliable combustion kinetic models are the foundation of the numerical simulation.Model analysis is an important technique to reduce the model prediction uncertainty and improve the prediction accuracy.However,there are usually lots of model parameters in real fuels,which cause many problems with directly applying current analysis methods to these large kinetic models,such as high computational cost and inaccurate results.Meanwhile,when using experimental data to optimize models,many experimental data have large uncertainties or are even missing,which requires research and improvement of model analysis-based optimization methods and experimental design methods.This study aims to develop and improve model analysis and experimental design methods based on feature extraction,to improve accuracy and computational efficiency,and to use these newly developed methods to improve the Opt Ex（Optimal Experiments）platform developed by our research group.Reducing the dimension of combustion models can effectively reduce the computational cost of the model analysis.An active subspace-based dimension reduction method is applied to construct a surrogate model for the combustion kinetic model.The active subspace method is used to extract low-dimensional new features of the model,and then construct an artificial neural network（ANN）with the extracted features to conduct the following model analysis.Because the information of the full kinetic model parameters is included in the features,each model parameter’s global sensitivity index can be investigated,thus leads to a more accurate sensitivity analysis result.The effectiveness and efficiency of this method are demonstrated in the cases of hydrogen,methanol,and n-decane systems.To improve the efficiency of experimental design and solve the problem of mismatch between the target condition and the experimental conditions,a similarity analysis method is developed,which uses the physical meanings of the features extracted by the active subspace method.This study uses hydrogen,dimethyl ether,and ethane as examples to verify the effectiveness and efficiency of the active subspace based-similarity analysis method,demonstrating its application in finding alternative experimental measurements or experimental conditions,and it is also compared with the previous similarity method（surrogate model-based similarity analysis）in terms of computational efficiency.Inspired by the above research on feature extraction of input parameters,this study combined model output parameters（corresponding to experimental data）into new features based on experimental measurement principles,and proposed the use of feature experimental data method to reduce the uncertainty of experimental data that are difficult to accurately quantify,such as free radicals.This allows these experimental data to be used quantitatively in model optimization and other analyses,rather than being limited in qualitative analysis.Using the mole fractions of species in the flame or low-temperature oxidation process of CH₃OCH₃ and CH₃OH as examples,and combining with the methods mentioned above,the superiority and inferiority of the feature experimental data compared to species concentrations in model optimization and experimental design are demonstrated.This method with the above methods is also integrated into our OptEx platform.