Multi-Fidelity Surrogate-Based Optimization for Electromagnetic Simulation Acceleration

As circuits’ speed and frequency increase, fast and accurate capture of the details of the parasitics in metal structures, such as inductors and clock trees, becomes more critical. However, conducting high-fidelity 3D electromagnetic (EM) simulations within the design loop is very time consuming and computationally expensive. To address this issue, we propose a surrogate-based optimization methodology flow, namely multi-fidelity surrogate-based optimization with candidate search (MFSBO-CS), which integrates the concept of multi-fidelity to reduce the full-wave EM simulation cost in analog/RF simulation-based optimization problems. To do so, a statistical co-kriging model is adapted as the surrogate to model the response surface, and a parallelizable perturbation-based adaptive sampling method is used to find the optima. Within the proposed method, low-fidelity fast RC parasitic extraction tools and high-fidelity full-wave EM solvers are used together to model the target design and then guide the proposed adaptive sample method to achieve the final optimal design parameters. The sampling method in this work not only delivers additional coverage of design space but also helps increase the accuracy of the surrogate model efficiently by updating multiple samples within one iteration. Moreover, a novel modeling technique is developed to further improve the multi-fidelity surrogate model at an acceptable additional computation cost. The effectiveness of the proposed technique is validated by mathematical proofs and numerical test function demonstration. In this article, MFSBO-CS has been applied to two design cases, and the result shows that the proposed methodology offers a cost-efficient solution for analog/RF design problems involving EM simulation. For the two design cases, MFSBO-CS either reaches comparably or outperforms the optimization result from various Bayesian optimization methods with only approximately one- to two-thirds of the computation cost.