Answers are grounded in this paper and cite their sources.
Three-dimensional Cu–Cu bonded integration is central to next-generation high-bandwidth and heterogeneous systems, yet the mechanical consequence of increasing interconnect density remains insufficiently resolved at the system level. Here, a validated transient thermo-mechanical framework is developed to examine density-dependent response in direct Cu–Cu bonded stacked dies under JEDEC-aligned thermal cycling, and to convert that response into a rigorous design rule through a supervised machine learning multi-objective desirability-based optimization. Five interconnect densities, spanning 50 × 50 to 70 × 70 arrays, were assessed using coupled thermal-structural simulations and experimental validation of heating temperature and out-of-plane deformation. Beyond conventional metrics, the study introduces two cycle-resolved energetic descriptors: the cycle irreversibility fraction, which measures the share of thermo-mechanical activity converted into irreversible dissipation, and the thermo-plastic accommodation index, which quantifies the plastic accommodation burden imposed on the Cu network. The results show that densification is not monotonically beneficial. Warpage decreased from 0.611 µm at 50 × 50 to 0.562 µm at 65 × 65 before rising at 70 × 70, while maximum principal stress remained restrained through the intermediate regime and then escalated sharply at the highest density. Concurrently, cycle irreversibility increased, whereas thermo-plastic accommodation decreased, revealing a transition from recoverable structural response to dissipation-dominated constraint. Surrogate-assisted desirability analysis identified density 64 as the most robust compromise, establishing that mechanically optimal Cu–Cu architectures emerge from balanced energetic governance rather than maximal interconnect insertion. The current findings are expected to serve as references for advancing interconnection technology.
No takes yet. Share an insight, caveat, or question.
Wong et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: