Yield analysis optimization of microwave devices; antennas using non-linear partial-least-squares; polynomial chaos expansion

Date
2022-12-01
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: Polynomial Chaos Expansion (PCE is introduced as an alternative to classical sensitivity analysis techniques, such as Monte Carlo analysis, for electromagnetic applications. After a su ciently accurate PCE surrogate is constructed, the statistical information estimates can be extracted from the PCE coe cients. The construction of the PCE surrogate typically requires fewer frequency sweeps than statistical information estimation using Monte Carlo analysis. The advantages of PCE are thus two-fold, it provides a computationally inexpensive surrogate and the statistical information estimates can be extracted directly from the PCE coe cients. Di erent coe cient calculation methods are considered to construct the most e cient PCE using the smallest number of samples. A simple inset-fed patch antenna is chosen as a benchmark problem. Cross-polarisation is chosen as a performance characteristic and the position of the feed is identi ed as the most sensitive parameter associated with cross-polarization. PCE models are constructed for each coe cient calculation method and statistical information estimates (mean and variance are extracted and compared to a Monte Carlo analysis, whereafter the most e cient PCE coe cient calculation method is determined. PCE construction for complex structures with a large number of sensitive system parameters is prohibitive since a large number of samples are required. The NLPLS-based PCE method is proposed, where a surrogate model is constructed with a massive reduction in system parameters using NLPLS, and the statistical information is determined simultaneously with the same sample set using PCE. This method is successfully applied to a dual-band patch antenna problem with 8 system parameters and a diplexer problem with 37 system parameters. This allows accurate yield estimates using only 20 and 30 samples for the dual-band patch antenna and the diplexer respectively, whereas a Monte Carlo analysis requires 500 samples. A variance-based global sensitivity analysis seamlessly follows the NLPLSbased PCE surrogate, and global sensitivity analysis of the dual-band patch antenna, the diplexer, and a diplexer optimized for manufacturing is performed. Multiple NLPLS-based PCE derivate yield optimization algorithms are proposed and applied to optimize the yield of the manufactured diplexer, with 38 system parameters, and a 100 GHz lter with 43 system parameters. A performance-guided random walk algorithm improved the yield of the manufactured diplexer from 57.28 % to 100 % after only 4 iterations. The same algorithm improved the yield of the 100 GHz lter from 7.67 % to roughly 90 % after only 8 iterations.
AFRIKAANS OPSOMMING: Polinomiese Chaos Uitbreiding (PCE word bekendgestel as 'n alternatief vir klassieke sensitiwiteitsanalise tegnieke, soos Monte Carlo-analise, vir elektromagnetiese toepassings. Na 'n voldoende akkurate PCE-surrogaat gekonstrueer word, kan die statistiese inligtingskattings uit die PCE-koë siënte onttrek word. Die konstruksie van die PCE-surrogaat vereis gewoonlik minder frekwensie punte as wat statistiese inligtingskatting met behulp van Monte Carlo analise vereis. Die voordele van PCE is dus tweeledig, dit bied 'n kostee ektiewe surrogaat aan en die statistiese inligtingskattings kan direk uit die PCE-koë siënte onttrek word. Verskillende koë siënt berekeningsmetodes word oorweeg om die mees doeltre ende PCE te konstrueer deur die kleinste aantal monsters te gebruik. 'n Eenvoudige inset-gevoede mikrostrook antenna word as 'n maatstafprobleem gekies. Kruispolarisasie word as prestasie kenmerk gekies en die posisie van die voer word geïdenti seer as die mees sensitiewe parameter wat met kruispolarisasie geassosieer word. PCE-modelle word vir elke koë siënt berekeningsmetode saamgestel en statistiese inligtingskattings (gemiddeld en variansie word onttrek en vergelyk met 'n Monte Carlo analise, waarna die mees doeltre ende PCE-koë siënt berekeningsmetode bepaal word. PCE-konstruksie vir komplekse strukture met 'n groot aantal sensitiewe stelselparameters is onaantreklik, aangesien 'n groot aantal monsters benodig word. Die NPLLS-gebaseerde PCE metode word voorgestel, waar 'n surrogaatmodel gekonstrueer word met 'n massiewe vermindering in stelselparameters deur gebruik te maak van NLPLS, en die statistiese inligting gelyktydig met die dieselfde monster-stel onttrek word met behulp van PCE. Hierdie metode word suksesvol toegepas op 'n dubbelband-mikrostrook-antenna probleem met 8 stelselparameters en 'n diplekser probleem met 37 stelselparameters. Dit het akkurate opbrengs moontlik gemaak met slegs 20 en 30 monsters vir die dubbelband-mikrostrook-antenna en die diplekser onderskeidelik, waar 'n Monte Carlo-analise 500 monsters vereis het. 'n Variansie-gebaseerde globale sensitiwiteitsanalise volg direk die NLPLS-gebaseerde PCE surrogaat, en globale sensitiwiteitsanalise van die dubbelband-mikrostrook-antenna, die diplekser en 'n diplekser geoptimeer vir vervaardiging, is suksesvol uitgevoer. Veelvuldige NPLS-gebaseerde PCE-afgeleide opbrengs-optimeringsalgoritmes word voorgestel en toegepas om die opbrengs van die vervaardigde diplekser te optimeer met 38 stelselparameters, en 'n 100 GHz- lter met 43 stelselparameters. 'n Prestasiegeleide ewekansige stap-algoritme het die opbrengs van die vervaardigde diplekser van 57.28 % tot 100 % verbeter na slegs 4 iterasies. Dieselfde algoritme het die opbrengs van die 100 GHz- lter verbeter van 7.67 % tot ongeveer 90 % na slegs 8 iterasies.
Description
Thesis (PhD) -- Stellenbosch University, 2022.
Keywords
Random polynomials, Stochastic analysis, Simulated annealing (Mathematics), Parameter estimation, Electromagnetism, UCTD
Citation