[论文解读] Deep learning methods based on cross-section images for predicting effective thermal conductivity of composites

Qingyuan Rong, Wei Han+1 more|arXiv (Cornell University)|Apr 12, 2019

Thermal properties of materials参考文献 54被引用 4

一句话总结

本文提出将复合材料的二维截面图像作为二维卷积神经网络（CNN）的输入，以预测有效热导率，其精度与三维CNN相当。通过利用多个截面图像，尤其是与热流方向对齐的截面，二维CNN实现了高性能预测，减少了对昂贵的三维微观结构数据的依赖，同时通过图像的代表性与尺寸保持了高精度。

ABSTRACT

Effective thermal conductivity is an important property of composites for different thermal management applications. Although physics-based methods, such as effective medium theory and solving partial differential equation, dominate the relevant research, there is significant interest to establish the structure-property linkage through the machine learning method. The performance of general machine learning methods is highly dependent on features selected to represent the microstructures. 3D convolutional neural networks (CNNs) can directly extract geometric features of composites, which have been demonstrated to establish structure-property linkages with high accuracy. However, to obtain the 3D microstructure in composite is generally challenging in reality. In this work, we attempt to use 2D cross-section images which can be easier to obtain in real applications as input of 2D CNNs to predict effective thermal conductivity of 3D composites. The results show that by using multiple cross-section images along or perpendicular to the preferred directionality of the fillers, the prediction accuracy of 2D CNNs can be as good as 3D CNNs. Such a result is demonstrated with the particle filled composite and a stochastic complex composite. The prediction accuracy is dependent on the representativeness of cross-section images used. Multiple cross-section images can fully determine the shape and distribution of fillers. The average over multiple images and the use of large-size images can reduce the uncertainty and increase the prediction accuracy. Besides, since cross-section images along the heat flow direction can distinguish between serial structures and parallel structures, they are more representative than cross-section images perpendicular to the heat flow direction.

研究动机与目标

• 利用机器学习建立复合材料有效热导率的结构-性能关联。
• 通过使用二维截面图像作为输入，克服获取三维微观结构数据的挑战。
• 评估仅使用二维图像时，二维CNN是否能达到与三维CNN相当的预测精度。
• 识别最优的截面取向和图像处理策略，以最大化预测的可靠性。
• 证明使用具有代表性的二维图像堆叠可推断颗粒填充及随机复合材料的三维热行为的可行性。

提出的方法

• 在复合材料微观结构的多个二维截面图像上训练二维卷积神经网络（CNN）。
• 采用沿热流方向获取的截面图像，以更好地区分串联与并联热传导路径。
• 通过平均多个截面图像的预测结果，降低不确定性并提升特征表征能力。
• 采用大尺寸图像以增强空间分辨率，捕捉更精细的微观结构细节。
• 将预测结果与基于物理的模拟或解析模型的参考解进行验证。
• 比较二维CNN在截面图像平行与垂直于热流方向时的性能表现。

实验结果

研究问题

• RQ1 在仅使用截面图像训练的情况下，二维CNN能否实现接近三维CNN的热导率预测精度？
• RQ2 截面图像的取向（平行或垂直于热流方向）如何影响预测性能？
• RQ3 使用多个截面图像在多大程度上提升了二维CNN预测的代表性与精度？
• RQ4 图像尺寸与集成平均在降低热导率预测不确定性方面起到何种作用？
• RQ5 二维截面数据能否充分捕捉影响复合材料热传导的关键三维微观结构特征？

主要发现

• 当使用复合材料微观结构的多个二维截面图像时，二维CNN的预测精度可与三维CNN相媲美。
• 与热流方向对齐的截面图像更具代表性，其预测精度显著高于垂直于热流方向的截面图像。
• 通过平均多个截面图像的预测结果，可降低不确定性并提升模型鲁棒性。
• 更大的图像尺寸有助于提升特征提取能力，从而改善预测性能。
• 所用截面图像的代表性是整体预测精度的关键决定因素。
• 该方法可在无需完整三维微观结构数据的情况下实现高精度的热导率预测，显著降低实验与计算成本。