Automated Space-mapping framework for electromagnetic device optimisation

Wolsky, David William (2019-04)

Thesis (MEng)--Stellenbosch University, 2019.

Thesis

ENGLISH ABSTRACT: A space-mapping (SM) framework that allows an automated approach to solving computer-aided design (CAD) optimisation problems for electromagnetic (EM) devices is presented. Direct optimisation of detailed, high-fidelity/fine EM models can be computationally expensive and can restrict the adoption of optimisation for large systems. SM allows the incorporation of low-fidelity/coarse models that are quick to evaluate, without sacrificing the accuracy of results. The SM framework builds up a surrogate model from a coarse model that is aligned programmatically to the fine model. Optimisation is carried out using the surrogate model. If the surrogate is evaluated far away from where the alignment took place, the results may diverge. A trust-region (TR) is introduced as a method of improving the robustness of the framework. The TR governs the bounds of the optimisation space. Four types of SM are implemented within the automated framework: input, output, implicit and frequency SM. Literature using some of these techniques is investigated, and a detailed analysis on the original SM implementation and a frequency SM approach is included. A basic TR implementation, from literature, is also investigated in detail. The methodology used to develop the automated framework is explained, and Matlab implementation details for each stage are discussed. Model alignment and surrogate building for each of the SM techniques are discussed. The user's interface to the TR enhanced SM optimisation system is detailed. The available high-fidelity solvers are FEKO and CST, while those for low-fidelity are AWR-MWS and Matlab. A microstrip stub example is used to demonstrate input, implicit and frequency SM. FEKO and AWR-MWS are used for these examples. A microstrip double folded stub filter is taken from literature and used to evaluate the system. This bandstop example has three design variables and is required to meet three S-parameter goals. An additive input and implicit SM approach is chosen to solve this problem. Each iteration is analysed and the SM framework successfully meets specification within four fine model evaluations. Finally, improvements to the automated framework are presented. A general mathematical model is suggested for unit-testing, and an object orientated design is suggested.

AFRIKAANSE OPSOMMING: 'n Ruimteafbeelding (SM) raamwerk wat 'n outomatiese benadering vir die oplos van rekenaar gesteunde ontwerp (CAD) optimeringsprobleme vir elektromagnetiese (EM) toestelle toelaat word aangebied. Direkte optimering van gedetailleerde, hoëtrou EM modelle kan bewerkingsintensief wees, en kan die gebruik van optimering in groot stelsels beperk. SM laat die inkorporasie van lae-vertroue/growwe modelle toe wat vinnig is om te evalueer, sonder om die akkuraatheid van die resultate in te boet. Die SM raamwerk bou 'n surrogaatmodel vanaf die growwe model op, wat programmaties belyn word met die hoëtrou/fyn model. Optimering word uitgevoer deur van die surrogaatmodel gebruik te maak. As die surrogaat ver van enige punt waar belyning plaasgevind het geëvalueer word, mag die resultate divergeer. A vertrouegebied (TR) word voorgestel as 'n metode om die robuustheid van die raamwerk te versterk. Die TR beheer die grense van die optimeringsruimte. Vier tipes SM word geïmplementeer binne die outomatiese raamwerk: intree, uittree, implisiete en frekwensie SM. Literatuur wat gebruik maak van party van die tegnieke word bestudeer, en 'n gedetailleerde analise van die oorspronklike SM implementasie en 'n frekwensie SM implementasie word ingesluit. 'n Basiese TR implementasie, van die literatuur, word ook in detail ondersoek. Die metodiek wat gebruik is om die outomatiese raamwerk te ontwikkel word verduidelik, en Matlab implementeringsdetails vir elke stadium word bespreek. Die gebruikerskoppelvlak na die TR-verbeterde SM optimerings stelsel word bespreek. Die beskikbare hoëtrou oplossers is FEKO en CST, terwyl, vir die growwe modelle, AWR-MWS en Matlab gebruik word. 'n Mikrostrook stomplyn voorbeeld is gebruik om die gebruik van intree, implisiete en frekwensie SM toe te lig. FEKO en AWR-MWS word vir hierdie voorbeelde gebruik. 'n Mikrostrook dubbelgevoude stomplyn filter word van die literatuur geneem en gebruik om die stelsel te evalueer. Hierdie bandstop voorbeeld het drie ontwerpsveranderlikes en daar word verwag dat drie S-parameter doelfunksies bereik word. 'n Optellings intree en implisiete SM benadering is gekies om hierdie probleem op te los. Elke iterasie is geanaliseer en die SM raamwerk haal suksesvol die spesifikasie binne vier fyn model evaluasies. Ten slotte word verbeterings aan die outomatiese raamwerk voorgelê. 'n Algemene wiskundige model word voorgestel vir eenheidstoetse, en toekomstige werk wat objek georiënteerde ontwerp voorstel word bespreek. iii

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/105811
This item appears in the following collections: