ABSTRACT
Precise predictions of flame types and autoignition locations of jet flames in hot turbulent coflow are crucial for the stable and efficient operations of many advanced combustors. In order to achieve fast and accurate machine learning (ML) based prediction models for the autoignition of fuel jets in turbulent hot coflow, the support vector machine (SVM) and artificial neural network (ANN) models for two problems, flame-type classification and liftoff height regression, were established. The effects of hyperparameters on the models were discussed in the candidates of grid search. Furthermore, the important features for liftoff height regression were identified using Sobol sensitivity analysis. The results show that the grid search is a convenient and effective approach for hyperparameter optimization, especially for important parameters, such as the regularization parameter C and parameter γ of SVM and the initial learning rate (lr0) of ANN. Both SVM and ANN models have an accuracy of over 90% for classification and regression problems, with ANN models outperforming SVM models. This paper successfully predicts the flame type and liftoff height for the first time using ML methods. The predicted discrepancies generally occur in the flame-type transition interval, which may need more experimental data for better training. The sensitivity of each feature on the ANN model is relatively close and much higher than that on the SVM model, with the Reynolds number of fuel jets being the most important feature on the SVM model.
Disclosure statement
No potential conflict of interest was reported by the author(s).