Study On The Optimization Of Flamelet Tabulated Scalars And Modeling Of Heat Release Rate Based On Machine Learning

With the development of the data driving concept,computational combustion is also becoming combined with data science to solve combustion problems.Therefore,in this dissertation,machine learning algorithm is used as a tool to carry out the research on the optimization of flamelet tabulated scalars for premixed and stratified flames under strain effects and the model of flame front heat release rate.The main contents and conclusions of this dissertation are as follows:（1）Two methods are used to analyze the chemistry data of one-dimensional premixed and stratified CH₄/air flames under strain effects.One is principal component analysis（PCA）based on unsupervised linear dimensionality reduction algorithm,another is partial least squares（PLS）based on supervised linear dimensionality reduction algorithm.Unsupervised PCA method and supervised PLS method with heat release rate as dependent value attain similar results.Tabulated scalars selected by the PCA and PLS algorithm are orthogonal to each other,so it can hold more species information in the mapping of the lower dimension manifold,which is more advantageous than the method of characterizing the strain effects by the mass fraction of species H.（2）PCA and PLS algorithms are optimized to keep mixture fraction Z,which is very important in combustion,as tabulated scalars.The modified algorithms are named PCA*and PLS*.The orthogonal tabulated scalars obtained by the modified algorithm can successfully reconstruct the stratified flame structure under strain effects.Moreover,the modified PLS*algorithm retains passive scalars mixture fraction Z and inherits the advantage of supervised algorithm.It will be a promising method.（3）Using the freely propagating one-dimensional CH₄/air lean premixed flame data as sample,three machine learning algorithms,including artificial neural network,support vector machine regression and multiple linear regression,are used to study the prediction accuracy of heat release rate model under lean combustion regime.The results show that the prediction accuracy of commonly used heat release rate models such as k[CH2O][OH]strongly depends on the equivalence ratio,while machine learning algorithm can have good prediction accuracy under wide range of the equivalence ratio.Comparison between the prediction accuracy and the difficulty of parameter adjustment of three machine learning algorithms suggests that,multiple linear regression algorithm is the best one.The multiple linear regression model based on the molar concentration of species CH3 and O has a good performance in predicting the heat release rate of CH₄/air under a wide range of the equivalence ratio combustion.（4）The heat release rate model including low temperature oxidation reaction and high temperature reaction was studied based on the auto-ignition process data of n-dodecane zero dimensional homogeneous reaction.The results show that the reaction path of high carbon fuel is very complicated.Simple linear models such as k[CH2O][OH]cannot take into account the heat release rate of low temperature and high temperature reaction at the same time,however,the model using multiple linear regression algorithm can take into account the heat release rate of two stages very well.（5）Taking the two-dimensional direct numerical simulation（DNS）data of n-dodecane/air auto-ignition process as a sample,the optimization of the heat release rate model for turbulent spray combustion where complex stratification exists was conducted.The results show that the heat release rate model constructed with multiple linear regression algorithm is much better than the simple linear heat release rate model,and the prediction of heat release reaction zone is more accurate,that is,the heat release rate model constructed by machine learning algorithm has higher accuracy and wider applicable range.