The thermo-mechanical properties of System-in-Package (SiP) products play a crucial role in manufacturability and reliability. The multi-material and cross-scale properties of SiP pose challenges for thermo-mechanical modeling and simulation. The industry primarily uses homogenization-based material equivalence methods to avoid dense meshing of microstructures, thereby significantly reducing the scale of Finite Element Analysis (FEA).
To address the issues of insufficient accuracy, high computational resource consumption and poor engineering usability in existing homogenization methods, this study proposes a modeling method based on a convolutional neural network (CNN) for predicting monoclinic homogenized constitutive parameters. This method can accurately characterize the tension-shear coupling and thermal expansion-shear coupling effects caused by the asymmetry of the package’s microstructure. The surrogate model for equivalent mechanical parameter representation, constructed using a CNN based on a multi-channel encoding method of pixel gray values, exhibits strong adaptability to multi-material systems.
This study achieves cross-scale reduced-order modeling of multi-material advanced packaging and complex substrate circuit patterns and it has been preliminarily validated through application in thermal warpage optimization simulations for a product-level SiP. The number of mesh elements constructed is merely 1/3136 of that in traditional direct modeling methods, significantly shortening simulation time while maintaining accuracy.
This study introduces a unified CNN framework for the precise characterization of multiple material systems to predict monoclinic homogenized constitutive parameters. The method accurately captures the tension–shear coupling and thermal expansion–shear coupling effects arising from the asymmetry of the package microstructure, significantly reducing simulation time while maintaining accuracy.
