2019/04/12 by Qingyuan Rong, Wei Han, Rong, Qingyuan +3
Engineering · Materials Science · #Composite Material Mechanics #Computational Physics (physics.comp-ph) #Data Analysis #FOS: Physical sciences #Heat Transfer and Optimization #Statistics and Probability (physics.data-an) #Thermal properties of materials #Thermography and Photoacoustic Techniques
paper · pdf · doi:10.48550/arxiv.1904.06104
openalex publication_date 2019/04/12 · openalex created_date 2021/10/11 · openalex updated_date 2026/07/28
Effective thermal conductivity is an important property of composites for\ndifferent thermal management applications. Although physics-based methods, such\nas effective medium theory and solving partial differential equation, dominate\nthe relevant research, there is significant interest to establish the\nstructure-property linkage through the machine learning method. The performance\nof general machine learning methods is highly dependent on features selected to\nrepresent the microstructures. 3D convolutional neural networks (CNNs) can\ndirectly extract geometric features of composites, which have been demonstrated\nto establish structure-property linkages with high accuracy. However, to obtain\nthe 3D microstructure in composite is generally challenging in reality. In this\nwork, we attempt to use 2D cross-section images which can be easier to obtain\nin real applications as input of 2D CNNs to predict effective thermal\nconductivity of 3D composites. The results show that by using multiple\ncross-section images along or perpendicular to the preferred directionality of\nthe fillers, the prediction accuracy of 2D CNNs can be as good as 3D CNNs. Such\na result is demonstrated with the particle filled composite and a stochastic\ncomplex composite. The prediction accuracy is dependent on the\nrepresentativeness of cross-section images used. Multiple cross-section images\ncan fully determine the shape and distribution of fillers. The average over\nmultiple images and the use of large-size images can reduce the uncertainty and\nincrease the prediction accuracy. Besides, since cross-section images along the\nheat flow direction can distinguish between serial structures and parallel\nstructures, they are more representative than cross-section images\nperpendicular to the heat flow direction.\n