An Efficient Surrogate-Based EM-Centric Multiphysics Optimization Incorporating Feature Assistance for Tunable Microwave Filter Design

This article proposes an efficient surrogate-based electromagnetic (EM)-centric multiphysics optimization methodology that incorporates feature information assistance for the design of tunable microwave filters. A neuro-transfer function (neuro-TF) surrogate is adopted to represent the EM-centric multiphysics behavior. The proposed approach tackles the problem of the initial point being considerably distant from the desired design target, as well as the inability to clearly identify and extract feature information in the EM-centric multiphysics response. The proposed technique develops a robust neuro-TF surrogate model considering the coupling effects in the multiphysics domain. Using vector fitting, the effective poles and zeros of the tunable filter are extracted. Neural networks are then used to map design variables to these poles and zeros. Feature parameters, including feature frequencies and responses, are identified. The transfer function of the pole/zero format is employed to help extract feature information directly. Furthermore, new formulations for calculating the first derivative of the response of the EM-centric multiphysics surrogate model relative to the design parameters are derived, so as to conduct gradient-based optimization. A trust region algorithm is applied to improve optimization speed and convergence. A novel objective function incorporating feature parameters is formulated. By leveraging these feature parameters, this approach can enhance the capability to avoid local minima and achieve optimization more rapidly than the surrogate-based multiphysics optimization approach that lacks feature assistance. The efficacy of the proposed approach is demonstrated through two EM-centric multiphysics optimization examples involving tunable microwave filters.