Low-Fidelity-Guided High-Fidelity Bayesian Optimization for High-Speed Link Signal Integrity

Abstract

To address the tradeoff between the high computational cost of high-fidelity simulations and the limited reliability of machine learning (ML) models in high-speed link signal integrity (SI) optimization, this article proposes a low-fidelity-guided high-fidelity Bayesian optimization (LFGHF-BO) method. Within the conventional Bayesian optimization (BO) framework, low-fidelity information is incorporated as auxiliary guidance, and the predictive distribution is reconstructed to improve sampling decisions, thereby enhancing optimization efficiency while preserving the dominance of high-fidelity evaluations. Unlike conventional approaches, the low-fidelity model is not directly involved in modeling the objective function but instead provides trend information to guide the search process, reducing the dependence on prediction accuracy. Moreover, under the practical assumption that historical high-fidelity simulation data are available, the low-fidelity model can be trained using historical data from simple structures and then reused for optimization tasks involving more complex structures, which effectively reduces the additional modeling cost. Numerical results demonstrate that, under a limited high-fidelity evaluation budget, the proposed method significantly improves sampling efficiency and accelerates convergence while maintaining optimization reliability.