EM-driven miniaturization of high-frequency structures through constrained optimization

Abstract

The trends afoot for miniaturization of high-frequency electronic devices require integration of active and passive high-frequency circuit elements within a single system. This high level of accomplishment not only calls for a cutting-edge integration technology but also necessitates accommodation of the corresponding circuit components within a restricted space in applications such as implantable devices, internet of things (IoT), or 5G communication systems. At the same time, size reduction does not remain the only demand. The performance requirements of the abovementioned systems form a conjugate demand to that of the size reduction, yet with a contrasting nature. A compromise can be achieved through constrained numerical optimization, in which two kinds of constrains may exist: equality and inequality ones. Still, the high cost of electromagnetic-based (EM-based) constraint evaluations remains an obstruction. This issue can be partly mitigated by implicit constraint handling using the penalty function approach. Nevertheless, securing its performance requires expensive guess-work-based identification of the optimum setup of the penalty coefficients. An additional challenge lies in allocating the design within or in the vicinity of a thin feasible region corresponding to equality constraints. Furthermore, multimodal nature of constrained miniaturization problems leads to initial design dependency of the optimization results. Regardless of the constraint type and the corresponding treatment techniques, the computational expenses of the optimization-based size reduction persist as a main challenge. This thesis attempts to address the abovementioned issues specifically pertaining to optimization-driven miniaturization of high frequency structures by developing relevant algorithms in a proper sequence. The first proposed approach with automated adjustment of the penalty functions is based on the concept of sufficient constraint violation improvement, thereby eliminating the costly initial trial-and-error stage for the identification of the optimum setup of the penalty factors. Another introduced approach, i.e., correction-based treatment of the equality constraints alleviates the difficulty of allocating the design within a thin feasible region where designs satisfying the equality constraints reside. The next developed technique allows for global size reduction of high-frequency components. This approach not only eliminates the aforementioned multimodality issues, but also accelerates the overall global optimization process by constructing a dimensionality-reduced surrogate model over a pre-identified feasible region as compared to the complete parameter search space. Further to the latter, an optimization framework employing multi-resolution EM-model management has been proposed to address the high cost issue. The said technique provides nearly 50 percent average acceleration of the optimization-based miniaturization process. The proposed technique pivots upon a newly-defined concept of model-fidelity control based on a combination of algorithmic metrics, namely convergence status and constraint violation level. Numerical validation of the abovementioned algorithms has also been provided using an extensive set of high-frequency benchmark structures. To the best of the author´s knowledge, the presented study is the first investigation of this kind in the literature and can be considered a contribution to the state of the art of automated high-frequency design and miniaturization.