基于物理模型分析与深度神经网络融合的爆炸流场实时模拟方法

To make a rapid and accurate prediction for regional pressure fields produced by Vapor Cloud Explosions (VCEs), this paper proposed a methodology that fuses the physical model-based analysis with a Deep Neural Network (DNN). This fusion method comprises 4 steps, including the construction of a database, dimensionality reduction, data regression, and data generation. For a specific region, terrain parameters such as spatial layout and natural and geographical conditions normally remain the same for a long time or change very little in a short time, presenting static states. Besides, when performing numerical simulations for different types of generic VCEs, the arrangement of observation points would be fixed. Therefore, the parameters of the explosive source and the observation time were defined as explosion scenario-related variables. By randomly changing the explosive source-related parameters of each generic VCE and calculating them in the CFD tool, a large number of blast data could be obtained. After that, the inference algorithm of the Variational Autoencoder with Deep Convolutional Layers (VAEDC) model was used to reduce the dimensions of the regional pressure field and get the latent variables, meanwhile, the generative model was used to map the latent variables to the pressure field data. The role of data regression is to correlate the latent variables with the explosion scenario-related variables by the Multi-layer Forward Neural Network (MFNN) model, making the hybrid model can predict the regional pressure field provided with a combination of explosion scenario-related parameters. Furthermore, a progressive-training approach was adopted to improve the DNN’s learning efficiency. Data on simple explosion scenarios were first fed into the model, and then data on more sophisticated explosion scenarios were provided to the model in turn. VCEs occurring in a typical tank farm were selected as an example for illustration. Through analyzing the effects of the VAEDC’s structure and latent space size, as well as MFNN’s loss function type on the generalization capacity of the model, the hybrid model was finally developed. In addition, the predictive performance of the DNN model was evaluated from 4 dimensions, including the number of input variables, the denseness of obstacles, the comparison of CFD data, and statistics of relative L2 norm errors. The results show that the average relative L2 norm error of prediction results was 12. 8% and the average prediction time was 0. 014 2 s. Therefore, this hybrid model can achieve the balance between the efficiency and accuracy of prediction.