Neural Network Approximation of Hydrogen Detonation with Consideration of Physically Reasonable Selection of Training Data

Modeling of chemical reactions accompanying combustion and detonation processes plays a key role in engineering design of rocket and power plants. Currently, approaches based on the use of artificial neural networks capable of approximating the dynamics of chemical reactions with acceptable accuracy under conditions of limited computational resources are being actively developed. Most existing studies in this area focus on the parameters of neural network architecture and methods of their optimization, while the issue of the principles of training sample formation remains unspoken. The choice of training data has a direct impact on the model's generalizability, robustness and interpretability. The use of randomly generated initial conditions can lead to the dominance of trivial or physically insignificant scenarios, which limits the neural network's ability to reproduce critical transient regimes. In this paper, we consider the hypothesis that the inclusion of states corresponding to changing processes in the system in the training sample can increase the sensitivity of the model to significant portions of the phase space. This approach allows not only to reduce the RMS error of approximation, but also to make the behavior of the neural network more physically interpretable, bringing it closer to explainable models in the context of reactive kinetics problems. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.