The initiation and propagation of gas-phase detonation waves are coupled with various physical phenomena, such as gas dynamics, thermodynamics, chemical reaction dynamics and shock dynamics etc., the flow process is very complex. At present, many fundamental problems associated with the initiation and propagation of detonation waves still lack a detailed, in-depth and accurate understanding. The core physical mechanism of detonation wave initiation and propagation is the interaction between nonlinear wave propagation and chemical reaction zone. The accurate modeling of chemical reaction model is an important part of quantitative investigation on detonation physics. In the numerical simulation of detonation wave propagation, using detailed reaction mechanism causes serious computational time-consuming problem, while the using simplified reaction mechanism or one-step reaction mechanism will greatly reduce the calculation accuracy, so how to improve the calculation efficiency and increase the computational accuracy has become an important issue in the numerical simulation of detonation wave. At the same time, in the experiments of detonation, it is very difficult to accurately measure the distribution information of each component on the chemical reaction zone. It is a very valuable work to combine other information to restore the chemical reaction information. In addition, artificial intelligence techniques develop faster and faster. Many advanced artificial intelligence algorithms have been applied in the study of physical problem to better analyze and model physical phenomena from the perspective of phenomenology. Under such a research background, this paper is based on numerical simulation as the main data source, with GPU acceleration, sensitivity analysis, genetic optimization algorithm and ANN (Artificial Neural Network) method as the main research methods in the process of detonation wave propagation, Relevant researches have been carried out on mechanism modification, chemical reaction information reconstruction and DCSNN(Detonation Chemical Source Neural Network) modeling of chemical reaction source term calculation.

First of all, based on the modular detonation wave numerical simulation program built in C++ language on the CPU, the CUDA language is used to port it to the GPU platform. The GPU acceleration method was verified by simulating the propagation process of a two-dimensional normal detonation wave, and a speedup of about 160 times was obtained. This proves that the developed GPU acceleration method can efficiently simulate detonation wave propagation.

Because of the huge number of components, the number of reactions, and the severe rigidity of the detailed chemical reaction mechanism, the calculation is time-consuming, while the simplified mechanism is used in component concentrations, the physical quantities of final flame temperature, ignition delay time, specific heat ratio, molar mass may have errors up to 10%. For the detonation cell size, which is the most important dynamic parameter in the detonation wave propagation process, these physical quantities have an important influence on the calculation of the cell width. This paper uses a combined modified model to improve the computational accuracy of chemical reactions while maintaining the computational efficiency of the simplified mechanism. After verifying the effectiveness of the simplified mechanism obtained by the sensitivity analysis, the combined modified model is obtained by using the genetic algorithm. Then, this paper analyzes and compares the combined modified model and the simplified mechanism. The component concentration error is reduced from 10% to 0.1%, the final temperature error is reduced from 1% to 0.01%, and the ignition delay error is reduced from 0.7 % is reduced to 0.1%, and the error of molecular weight and specific heat ratio is also greatly reduced. This paper also analyses the influence of  combined modified model on calculating the cell size, a key dynamic parameter of detonation wave propagation.

In the existing detonation numerical simulation, the one-step or two-step reaction model is often used. These reaction models may obtain aerodynamic information consistent with the experimental such as pressure distribution, temperature distribution, and density distribution, but cannot obtain the detail distribution information of species concentration. On the other hand, in the detonation experiment, it is difficult to obtain the precise species concentration distribution information with the current experimental technology, but it is often easier to measure the aerodynamic information. Therefore, this paper proposes a CNN(Convolution Neural Network) method to reconstruct the chemical reaction information based on aerodynamic information during detonation wave propagation. The detailed chemical reaction mechanism is used to calculate the flow field distribution at various times as training samples, the aerodynamic information of the entire flow field is used as input and the species concentration distribution on the chemical reaction surface is used as output. The typical initiation and stable propagation phases are analyzed to verify the effectiveness of the chemical reaction zone reconstruction. In addition, this paper also conducts model training for different combinations of aerodynamic information, and analyzes the prediction results of different combinations of aerodynamic information. With an acceptable error of about 1%, the concentration distribution within the chemical reaction surface can be effectively reconstructed by the combined prediction using density, velocity and pressure, or by the prediction using pressure alone.

In terms of the simplification of the chemical reaction mechanism of the detonation wave numerical simulation, it is necessary to have rich knowledge of chemical reaction kinetics, and select the key reactions from the detailed mechanism according to the characteristics of the detonation wave, and the simplified mechanism obtained may still have rigid problems and thus the calculation is time-consuming. In this paper, a DCSNN/CFD coupling calculation method is proposed to establish a fast calculation method for the chemical reaction source term in the numerical simulation of detonation wave.After verification by simulating the one-dimensional detonation wave propagation, the DCSNN/CFD coupling method is used to calculate the CJ point pressure, temperature, density and wave velocity. At the stable propagation stage of detonation waves, compared with the detailed chemical reaction mechanism, the relative errors of the calculation results by the DCSNN/CFD coupling method are all in the order of 1%, indicating the reliability of this method. At the same time, the acceleration effect of different grid quantities is analyzed, and it is found that the DCSNN model can achieve a speedup ratio of about 13.2 and DCSNN/GPU model can achieve a speedup ratio of about 2000, which proves the efficiency of the DCSNN/CFD coupling calculation method.