基于卷积神经网络的地面尾喷焰流场预测

To address the time-consuming nature of traditional computational fluid dynamics (CFD) methods for calculating rocket exhaust plumes, a ground exhaust plume reaction flow intelligent prediction model is proposed. This model utilizes convolutional neural network (CNN) technology to rapidly generate accurate real-time exhaust flow field results based on typical engine spectral parameters. The study focuses on the BEM-II solid rocket engine and employs the latin hypercube sampling (LHS) method to select a sample space within a specific range of engine spectral parameters. A total of 40 steady-state exhaust plume flow field samples, obtained through CFD computations, are used as training data for the model. The model employs an encoder-decoder structure with parallel decoders, deconvolutional layers, and bottleneck layers to achieve reliable training performance with a small sample size. Experimental numerical results are employed to validate and evaluate the performance of the intelligent exhaust plume flow field model, known as CNN-PLUME, under various under-expanded states. The results demonstrate that the developed intelligent prediction model exhibits high regression coefficients close to 1, indicating its robustness. The model is capable of generating real-time exhaust plume reaction flow field results efficiently, and the predicted flow field aligns well with the experimental verification data, illustrating the model's high prediction accuracy. The maximum error in the exhaust plume flow field under different under-expanded states is 14.17%, indicating strong generalization ability. This model provides valuable theoretical support for studying ground exhaust plume flow field characteristics and facilitates the design of rocket engine spectral parameters.