基于“分段-组合”残差神经网络的超声速氢气零维点火计算方法

Calculation Method of Supersonic Hydrogen Zero-Dimensional Ignition based on Segmentation-Combination Residual Neural Network

关键词:  “分段-组合”模型  残差神经网络  零维点火  数值模拟  计算加速

Key words:  Segmentation-combination model  Residual neural network  Zero-dimensional ignition  Numerical simulation  Computational acceleration

摘要:

受限于发动机燃烧数值模拟需要长时间超级计算机运行的问题，发展了一种基于“分段-组合”残差神经网络的氢气零维点火计算方法。以氢气零维点火算例为基础，基于自主研发的高超声速内外流耦合数值模拟软件AHL3D构建数据集。数据集中输入变量为超声速工况下的温度、压强及8种组分质量分数的初始状态值，输出变量为3000个时刻点的温度、压强及8种组分质量分数状态值。构建了一种“分段”训练、“组合”预测的残差神经网络框架。算法首先将高维输入数据进行降维训练，再将“分段”模型预测后的参数冻结形成“组合”模型。与氢燃料直接计算相比，实验结果表明“分段-组合”残差神经网络可显著提升计算效率，对于11组分29反应的反应动力学模型可获得9.13倍的计算加速比，均方根误差降到了7.85×10^-5，氢燃料参数的预测精度都高于98%，计算效率及精度优于现有的神经网络燃烧计算方法。

Abstract:

Due to the problem that the engine combustion numerical simulation requires long-time supercomputer operation， a zero-dimensional hydrogen ignition calculation method based on segmentation-combination residual neural network was developed. For the zero-dimensional hydrogen ignition calculation example， this method establishes a data set in the self-developed hypersonic internal and external flow coupling numerical simulation software AHL3D. The input variables of the data set are the initial state values of temperature， pressure， and mass fractions of eight components under supersonic conditions. The output variables are the temperature， pressure， and state values of mass fractions of eight components at 3000 time points. The framework of residual neural network for “segment” training and “combination” prediction is constructed. The algorithm first reduces the dimensionality of the high-dimensional input data for training， and then freezes the parameters predicted by the “segmented” model to form the “combined” model. Compared with the direct calculation of hydrogen fuel， the experimental results show that the segmentation-combination residual neural network can significantly improve the calculation efficiency. For the reaction kinetic model of 11 components and 29 reactions， the calculation acceleration ratio can be 9.13 times， the root mean square error is reduced to 7.85×10^-5， and the prediction accuracy of hydrogen fuel parameters is higher than 98%. The calculation efficiency and accuracy are better than the existing neural network combustion calculation methods.