Prediction of State-to-State Dissociation Rate Coefficients Using Machine-Learning Algorithms

We study the possibility of using machine-learning algorithms to optimize the prediction of state-to-state (STS) dissociation rate coefficients in modeling nonequilibrium air flows. A rigorous but computationally complex theoretical model of reaction-rate coefficients, which considers the electronic and vibrational excitation of all reaction participants (products and reagents), is taken as a basis. Several algorithms are considered for predicting the STS dissociation rate coefficients of air components: k-nearest neighbors (k-NN) and decision tree (DT) regression, as well as neural networks; their accuracy and efficiency are analyzed. It is shown that the use of regression (k-NN and DT) algorithms is inappropriate for our problem, while neural-network algorithms have clear advantages over classical regression algorithms in terms of time and scalability. Validation of the neural-network approach is carried out by considering the example of solving the problem of vibrational-chemical relaxation behind a shock wave. A satisfactory agreement with the experiment and almost complete coincidence of the results with the solution obtained by theoretical methods without the use of machine learning are shown. The approach to data representation and processing proposed in this paper is easily scalable to more complex models taking into account the excitation of internal degrees of freedom. Thus, when taking into account the electronic excitation of a molecule, acceleration of about 1–2 orders is achieved without significant loss of accuracy. As a result, this study demonstrates that the use of neural-network methods makes it possible to predict state-specific reaction-rate coefficients with a high degree of accuracy without performing direct calculations using resource-intensive theoretical formulas directly in the working code. The approach scales as the complexity of the formulation increases (as is shown in the case of taking into account the electronic-vibrational excitation of the reagents), which allows us to reduce the time required to perform the calculations. At the same time, such a result is achieved through serious preliminary work and requires the development of large arrays of preliminary data. If we automate this process using a neural network, we can obtain a computationally efficient tool for systematic predictions of state-to-state reaction-rate coefficients.