A Data-Driven Approach to Collaborative Optimization for Enhancing TSV Multiphysics Performance

The through-silicon via (TSV) technology, due to its fine feature dimensions and high operating frequency, introduces complexity in performance analysis. Traditional analytical methods can become unrealistic when confronting the multiple considerations of thermomechanical and signal transmission characteristics. This article introduces a collaborative optimization framework incorporating the multioutput multifidelity surrogate model (MOMFSM) with a multiobjective optimizer. This framework aims to accelerate TSV structure design accounting for the multiphysics effect with low computational demand and uncompromised accuracy. The MOMFSM improves predictive accuracy by capturing the correlations between various physical outputs of TSV and fuses data of different fidelities, reducing reliance on high-fidelity (HF) data. Employing the rapid and precise MOMFSM, thermal stress and S-parameters distributions in TSV were simulated and used as analytical functions for design objectives. The multiobjective optimizer conducts an efficient multiphysics collaborative optimization, yielding a set of nondominated optimization solutions. The technique for order preference by similarity to an ideal solution (TOPSIS) was employed to derive the best compromise solution (BCS) from these options. The results indicate that the thermal stress in TSV interconnection was reduced from 690.3 to 635.4 MPa, the $S₁₁$ at 30 GHz decreased from −26.29 to −39.93 dB, and the $S₂₁$ at 30 GHz increased from −0.646 to −0.443 dB. This study, moreover, provides valuable insights for the comprehensive enhancement of TSV multiphysics performance and offers robust support for integrated circuit (IC) interconnection design.