Data-driven acceleration of cross-scale thermo-mechanical fatigue assessment for chiplet packages

High-density chiplet packages contain hierarchical micro-interconnects spanning multiple length scales, making fully detailed finite element assessment under coupled electro-thermal board-level loading computationally prohibitive, especially during iterative design updates. This study proposes a machine learning-accelerated cross-scale simulation framework for rapid fatigue life prediction and reliability screening. The framework combines representative volume element homogenization, a surrogate model trained to predict anisotropic effective properties from interconnect geometry and array density, and a global-to-local submodeling strategy. The predicted effective properties are embedded into the package-level model to cmpute temperature and displacement fields, which are then transferred to local submodels to recover stress and strain responses in critical regions. Compared with a fully detailed model, the framework reduces mesh element count by 92% and runtime by 94%, while maintaining key fatigue-relevant stress and strain metrics within 5%. Parametric analyses reveal size- and density-dependent transitions, including a sharp rise in TSV interfacial stress below 50 μm pitch and increased plastic strain in CPBs with diameters of 30–40 μm. Board-level thermal cycling tests with in-situ resistance monitoring and SEM confirm crack initiation at the predicted hotspots. The predicted fatigue life deviates from experiment by 8.3%, demonstrating accuracy and practical value for rapid reliability screening and design optimization.