多任务学习方法在神经网络替代火焰面数据库中的应用  

Application of multi-task learning method in replacing flamelet database with neural networks

作　　者：胡昌松 张腾 席玉茹 李井华[1] 颜应文[1] HU Changsong;ZHANG Teng;XI Yuru;LI Jinghua;YAN Yingwen(College of Energy and Power Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China)

机构地区：[1]南京航空航天大学能源与动力学院,南京210016

出　　处：《燃气涡轮试验与研究》2024年第3期53-62,共10页Gas Turbine Experiment and Research

基　　金：航空发动机及燃气轮机基础科学中心项目(P2022-B-Ⅱ-019-001)。

摘　　要：为实现神经网络对火焰面模型数据库的高精度替代,以残差神经网络为主体,采用多任务学习方法进行训练,并结合fmFoam求解器以Sandia D扩散火焰为对象,对该方法训练出的神经网络模型精度进行了验证。结果表明:采用多任务学习方法可有效提高神经网络对火焰面数据库的预测精度。相比于仅使用残差神经网络方法,采用多任务学习方法训练的神经网络可将各物理量预测结果的皮尔森系数由0.9990提升至0.9999,对质量占比前10组分预测结果的平均相对误差至少可降低81.1%;基于OpenFOAM对Sandia D火焰进行数值模拟,其计算结果在中心轴线以及各轴向位置的径向线上与传统方法基本一致,仅在反应进度变量源项上存在小范围差异;以传统查表方法为基准,FGM-MTL计算的温度及主要燃烧产物在中心轴线上的峰值相对误差最大为0.98%,峰值位置相对误差最大为2.37%。To achieve high-precision substitution of flamelet model databases using neural networks,residual neural networks trained by multi-task learning(MTL)method were taken as the subject.Integrated with the fmFoam solver and targeting the Sandia D diffusion flame,the precision of the trained neural network model was validated.The results indicate that the MTL approach significantly improves the prediction accuracy of the neural network for flamelet database.Compared to residual neural networks alone,the MTL-trained neural network increases the Pearson correlation coefficient of predicted physical quantities from 0.9990 to 0.9999,and reduces the average relative error for the top 10 components by mass fraction at least 81.1%.In numerical simulations of the Sandia D flame using OpenFOAM,the results of MTL-trained neural network along the centerline and various radial positions matches those of traditional methods,with only minor differences in the reaction progress variable source term.Using traditional lookup methods as a baseline,the maximum relative error for the peak temperature and main combustion products along the centerline calculated by the FGM-MTL method is 0.98%,with the maximum relative error in the peak position being 2.37%.

关 键 词：火焰面生成流形 多任务学习 残差神经网络 火焰面模型 机器学习 数值模拟 

分 类 号：Q643.21[生物学—生物物理学]