Image-Based Accelerated Prediction of Thermal Properties of Package Substrates Using Combined Deep-Learning and an Enhanced Thermal Network Model

As the need for processing large amounts of data increases, power consumption and the complexity of semiconductor package patterns also rise, making thermal management crucial. Traditional analytical models suffer from accuracy issues when analyzing thermal behaviors of complex patterns in commercial packages. To enable accurate and fast prediction of thermal behavior during the design stage in practical industry applications, this study proposes an image-based accelerated prediction method for the thermal properties of complex patterns in package substrates by using combined deep-learning and an enhanced thermal network model. The proposed method divides the layer-wise image data of package substrates into subdomains to define unit cells, and applies a thermal network with a new structure. The specified thermal networks are then matched with unit cell images and used for deep learning, thus automating the process for quick thermal property assessment. The proposed method is applied to commercialized package substrate designs and validated through experiments and finite element-based models, demonstrating high accuracy with R-squared values over 0.99 and reduction in prediction time exceeding 90%.