Deep Learning-Enabled Multiscale Strategy for Rapid and Accurate Warpage Prediction of Interposer Routing Interconnects

In advanced heterogeneous packaging, interposers serve as key components, integrating multiple chiplets onto a single package substrate through high-density signal redistribution. However, incorporating high-CTE metallic interconnects within low-CTE dielectrics and substrates inevitably induces significant warpage during thermal cycling. This deformation poses a severe threat to mechanical reliability, potentially leading to interconnect failure and yield loss, particularly as package dimensions increase while bump pitches and silicon die thicknesses scale down. While accurate warpage prediction is crucial for robust design, conventional finite element analysis (FEA) faces prohibitive computational challenges due to the extreme geometric scale mismatch between fine routing features and the global package structure. To address this challenge, this study proposes a deep learning–accelerated multiscale framework for rapid and accurate warpage prediction of interposer routing interconnects. The proposed approach establishes a warpage prediction pipeline by representing the spatial heterogeneity of complex routing patterns through equivalent properties. Validation against full routing-level simulations confirmed that the proposed method precisely predicts warpage behavior under optimal partitioning conditions. Furthermore, a deep learning surrogate model was employed to replace the iterative FEA-based homogenization process. The trained model enables rapid property estimation while maintaining high physical consistency with routing geometries. Consequently, the integrated methodology achieves a 2,643-fold speedup in analysis time while preserving the accuracy of full routing-level simulations. Ultimately, this framework provides a scalable and resource-efficient simulation strategy that minimizes computational overhead for the precision design of advanced packaging.