EI、Scopus 收录
中文核心期刊

基于深度学习建表的宽域发动机火焰面燃烧模型构建与验证

CONSTRUCTION AND VALIDATION OF A WIDE-DOMAIN ENGINE FLAMELET COMBUSTION MODEL BASED ON DEEP LEARNING TABLE BUILDING

  • 摘要: 以新型宽域发动机为动力的未来新一代飞行器的研发迫切需要CFD方法来进行高效高精度的辅助设计. 文章把传统的火焰面/进度变量燃烧模型与深度学习和神经网络方法相结合, 构建了新的智能化改进的燃烧模型并进行了算例测试与验证, 在保证计算效率的同时提高了预测精度. 首先, 给出了人工神经网络的构建方法, 包括数据库划分、数据归一化以及模型的训练等; 然后, 测试分析了不同函数结构对新建模型的影响, 并讨论了基于CPU和GPU的求解器框架下内存占用优化问题; 最后, 把智能化模型耦合到GPU求解器上对飞行马赫数4 ~ 12的3个发动机算例进行了数值模拟. 结果表明, 智能化改进的模型可代替传统火焰面/进度变量数据库从而实现高维参数建模及模型改进, 并可以成功运行在GPU上; 智能化改进的模型比传统的模型平均误差减小量均超过了50%, 算例误差最大减小值可达57.2%.

     

    Abstract: The development of future next-generation aircraft powered by a new wide domain engine urgently requires CFD methods for efficient and high-precision auxiliary design. This article combines the traditional flamelet/progress variable combustion model with deep learning and neural network methods to construct a new intelligent improved combustion model, which is tested and validated by numerical cases. This improves the prediction accuracy while ensuring computational efficiency. Firstly, the construction method of the artificial neural network in this article were proposed in details, including database partitioning, data normalization, and model training; Then, the impact of different function structures on the new model was tested and analyzed, and the optimization of memory usage under the CPU and GPU based solver framework was discussed; Finally, the intelligent model was coupled to a GPU solver for numerical simulation of three engine cases with flight Mach numbers 4 ~ 12. The results show that the intelligent improved model can replace the traditional flame surface/progress variable database to achieve high-dimensional parameter modeling and model improvement, and can successfully run on GPU; The intelligent improved model has an average error reduction of over 50% compared to traditional models for all mentioned cases, and the maximum reduction in case error can reach 57.2%.

     

/

返回文章
返回