Applying machine learning techniques to predict detonation initiation from hot spots

Abstract

As hot spots can be a source of detonation initiation by autoignition in a reactive mixture, understanding hot spot initiated detonation is significantly important for energy-related applications of detonation. Although the Zel’dovich reactivity gradient theory is extremely useful to predict reaction propagation modes from hot spots, the prediction is limited due to the changes in the unburnt mixture during the induction time. In the current study, machine learning based estimation methods by training the numerical simulation result data set are suggested to avoid the extensive and empirical effort to find initial conditions of hot spot initiated detonation in inhomogeneous mixtures. The data set was obtained from the detailed numerical simulations with various conditions of hot spots and divided into training and test data sets. Some variables were normalized with others to avoid multicollinearity. Three different machine learning techniques, logistic regression, classification and regression trees, and artificial neural network, were utilized to develop prediction models. Using the developed models by machine learning approaches, the modes of hot spot initiated reaction front propagation can be predicted without computationally expensive numerical simulations. The accuracies of the machine learning based prediction models were significantly improved compared to the baseline model simply using the Zel’dovich reactivity gradient theory. These models can be utilized as preliminary prediction methods of hot spot induced detonation initiation conditions for further detailed numerical simulations and dominant input variables.