Use of Machine Learning techniques and Neural Network algorithms for the Laminar Burning Speed estimation

This thesis project exploits the learning and computation capacity of Machine Learning algorithms, with the purpose of forecasting the Laminar Burning Speed (S_L) for Compressed Natural Gas (CNG) engines. The analysis has the objective to provide a robust model able to predict, in an acceptable computation time, the S_L. The robustness is fundamental for the CFD computations, also as the saving time that we can gain during the simulating with respect to the requested one for the different flame speed correlation laws. For the creation of predictive model, a dataset is adopted considering the S_L dependence on Temperature (T, [K]), pressure (p, [bar]), EGR rate [%], equivalent air-fuel ratio (Ø, [-]), available percentage of methane (% CH_4) and hydrogen (% H_2). However for an effective result, Rational Quadratic Gaussian Process Regression model is used due to its better performance than the other considered linear regression, regression tree, support vector machine (SVM) models. In addition, the neural networks as well are considered for the forecasting of S_L because these are used for solving the regression problems with a high level of efficiency and accuracy. In the specific for this computation step, the Levenberg-Marquardt training algorithm is considered.