航空学报 > 2021, Vol. 42 Issue (6): 124242-124242   doi: 10.7527/S1000-6893.2020.24242

基于深度学习驱动的L型定向热疏导机理

王泽林1, 籍日添1, 惠心雨1, 丁晨2, 汪辉1, 白俊强1   

  1. 1. 西北工业大学 航空学院, 西安 710072;
    2. 中国运载火箭技术研究院, 北京 100076
  • 收稿日期:2020-05-19 修回日期:2020-06-12 出版日期:2021-06-15 发布日期:1900-01-01
  • 通讯作者: 汪辉 E-mail:wanghui2018@nwpu.edu.cn
  • 基金资助:
    国家自然科学基金(11802245)

L-shaped directional heat transfer based on deep learning

WANG Zelin1, JI Ritian1, HUI Xinyu1, DING Chen2, WANG Hui1, BAI Junqiang1   

  1. 1. School of Aeronautics, Northwestern Polytechnical University, Xi'an 710072, China;
    2. China Academy of Launch Vehicle Technology, Beijing 100076, China
  • Received:2020-05-19 Revised:2020-06-12 Online:2021-06-15 Published:1900-01-01
  • Supported by:
    National Natural Science Foundation of China (11802245)

摘要: 碳/碳(C/C)复合材料具有热导率大、比强度高、耐烧蚀和耐冲刷等优异特性,被广泛应用于飞行器的热防护系统中,其有效导热系数对于实际应用而言是重要的热物理性质,尽管可以通过有效介质理论、对热扩散方程直接求解和玻尔兹曼输运方程等传统方法计算C/C复合材料有效导热系数,但这些数值方法通常十分耗时。本文引入深度学习方法,将格子玻尔兹曼(LBM)的三维格子模型作为三维卷积神经网络(3D-CNN)微观结构,不仅解决了三维微观结构模型难以捕获的问题,还便于实现数值计算模型和CNN模型的同步简化,利用3D-CNN快速精准地预测三维三相C/C复合结构的有效导热系数,基于此对内置L型高导热碳纤维丝的定向热疏C/C复合结构的有效导热系数进行快速预测和研究。研究表明,CNN模型在LBM传热计算上表现出强大的学习能力,但在测试样本结构孔隙率过分超出训练集时预测误差将大幅增加,且当孔隙率变化范围从30%~35%变化到55%~60%时,CNN模型"内插"预测的相对误差较模型"外推"降低了0.93%~30.72%。在C/C复合结构中内置L型高导热碳纤维丝可以将高温区域的热量沿纤维方向定向疏导至低温区域。

关键词: L型定向热疏导, 有效导热系数, 机器学习, 卷积神经网络, 格子玻尔兹曼方法

Abstract: Carbon/Carbon (C/C) composite materials are widely used in thermal protection systems of aircraft for their excellent characteristics such as high thermal conductivity, high specific strength, ablation resistance and erosion resistance, among which the effective thermal conductivity is an important physical property for practical applications. Traditional methods of studying effective thermal conductivity of composite materials such as the effective medium theory, the direct solution of heat diffusion equation or the Boltzmann transport equation are usually time-consuming. This paper introduces a deep learning method with the Lattice Boltzmann Method's (LBM's) three-dimensional lattice model as the microstructure sample of the Three-Dimensional Convolutional Neural Network (3D-CNN). This method not only overcomes the difficulty of capturing the three-dimensional microstructure model, but facilitates simultaneous simplification of the numerical calculation model and the CNN model. In this way, the effective thermal conductivity of the three-dimensional three-phase C/C composite structure can be predicted quickly and accurately by the 3D-CNN method. In addition, we quickly predict and study the effective thermal conductivity of the directional heat transfer C/C composite structure with built-in L-shaped carbon fiber with high thermal conductivity. Results show that the CNN model displays a strong learning ability in the LBM heat transfer calculation; however, when the porosity of the testing sample structure surpasses the training set excessively, the prediction error will increase significantly. When the porosity changes from 30%-35% to 55%-60%, the relative error of the CNN model "interpolation" is 0.93%-30.72% lower than that of the model "extrapolation". The built-in L-shaped carbon fiber with high thermal conductivity in the C/C composite structure can direct the heat in high temperature areas to low temperature areas along the fiber.

Key words: L-shaped directional heat transfer, effective thermal conductivity, machine learning, convolutional neural networks, lattice Boltzmann method

中图分类号: