Application of Convolutional Neural Network to Predict Anisotropic Effective Thermal Conductivity of Semiconductor Package

With increasing complexity of design patterns in semiconductor package substrates caused by demand for high-power semiconductors, it is necessary to be able to predict the thermal properties according to the pattern. Classifying the patterns is important to predict the effective thermal conductivity (ETC), but it has some difficulties due to the variable setting being labor-intensive and creating human uncertainty. These difficulties are amplified by the complexity of the pattern in the printed circuit board (PCB) substrate. This work presents an automated convolutional neural network (CNN)-based algorithm to infer the anisotropic ETCs of package substrates. This algorithm divides a layer-pattern image of a PCB into local unit-cell images and learns the pattern characteristics of each unit cell to identify the local ETC. The algorithm then builds an ETC map by integrating the local ETCs for the entire layer. The entire process is fully automated to reduce human uncertainty and required workforce. The ETC map from the algorithm was then used in finite element (FE) analysis and compared with three other prediction methods. The proposed algorithm can predict the anisotropic ETCs within 2-3 % errors compared to the reference data while other models lead to at least 16 % error. The FE simulation with the ETC map of the algorithm can reflect the effect of the design pattern on the heat flux and temperature distributions on the package layer, leading to the lowest root mean square error in the temperature distribution compared to other models.