Publications Repository - Gdańsk University of Technology

Design automation has been playing an increasing role in the development of novel antenna structures for various applications. One of its aspects is electromagnetic (EM)-driven design closure, typically applied upon establishing the antenna topology, and aiming at adjustment of geometry parameters to boost the performance figures as much as possible. Parametric optimization is often realized using local methods given usually reasonable quality of the initial designs obtained at the topology evolution stage. The major difficulty here is high computational cost associated with a large number of EM simulations required by conventional methods, both gradient and derivative-free routines. Possible workarounds including surrogate-assisted variable-fidelity methods (e.g., space mapping) face similar problems because the underlying low-fidelity model is often optimized directly. This paper proposes an expedited version of the trust-region (TR) gradient-based algorithm with numerical derivatives. A considerable reduction of the number of EM simulations is achieved by monitoring the behavior of the gradient throughout the algorithm run and omitting the finite-differentiation updates upon detecting stable patterns for particular parameter sensitivities. The proposed approach is benchmarked against the standard TR algorithm as well as the recently reported accelerated TR frameworks. Improved performance is consistently demonstrated for all considered test cases.

Authors

• dr hab. inż. Anna Pietrenko-Dąbrowska link open in new tab ,
• prof. dr inż. Sławomir Kozieł link open in new tab