Parametric Yield and Tolerance Optimisation of Electromagnetic Devices using Feature-based Non-Linear Partial Least Squares Polynomial Chaos Expansion Surrogates

Parametric modelling and global optimization techniques are critical in design of microwave components. This dissertation presents research work on efficient yield and tolerance optimisation of microwave components. In the first part of this thesis, we propose a rapid yield and sensitivity analysis approach exploiting feature based non-linear partial least squares polynomial chaos expansion (NLPLS-PCE). In the proposed method, yield is approximated at the level of the feature points since the functional relationship between the feature points and geometrical parameters is much less nonlinear compared to the entire response. Therefore accurate yield estimates are attained using few high-fidelity samples. Furthermore, NLPLS-PCE reduces the complexity of the problem by reducing the number of geometrical parameters used in construction of the surrogate model using NLPLS. This further reduces the number of high-fidelity samples required for accurate yield estimation especially for problems with large geometrical variations. The yield and sensitivity analysis method is successfully applied to a mid-frequency aperture array (MFAA) log periodic dipole antenna (LPDA) with 93 geometrical variables and an 18-variable KA band ridge waveguide filter terminated with a Vivaldi antenna. This allows for accurate yield and sensitivity estimates using 24 and 100 high-fidelity samples, respectively. Compared to Monte Carlo analysis, a minimum of 300 frequency sweeps are required for both the LPDA antenna and the KA band ridge waveguide filter terminated with a Vivaldi antenna to achieve accurate yield estimates. In the second part of this thesis, we propose a surrogate assisted yield optimization method combining feature based NLPLS-PCE and particle swarm algorithm. In the proposed optimization approach, the geometrical design parameters to iteratively change are selected based on sensitivities attained for the nominal design. The proposed yield optimisation method can achieve a higher yield increase with shorter CPU time. The method is successfully applied to an 18 variable KA band ridge waveguide filter terminated with a Vivaldi antenna. For this problem, nine geometrical variable were selected as the design variables. Yield was improved from 61% to 70.12% after only 6 iterations. Furthermore, the yield optimization technique is applied to an X-band square ridge filter characterized by 92 geometrical variables, from which two were selected as design variables. For this structure, yield was improved from 73% to 99% after only 10 iterations. As a further advancement, a surrogate assisted tolerance optimization method combining feature based NLPLS-PCE and particle swarm algorithm is developed. As opposed to optimisation of yield, the objective function is defined as maximizing input geometrical parameter tolerances while maintaining maximum yield. Similar to yield optimization, the design variables to iteratively change are selected based on prior sensitivity analysis of the microwave component at the yield optimized design. The method is applied to KA band ridge waveguide filter terminated with a Vivaldi antenna. The proposed tolerance optimization method improved the standard deviation of input parameters by 20%, from 1 𝜇𝑚 to 1.205 𝜇𝑚. The computational cost of yielding the final design corresponds to 288 full-wave EM simulations. The proposed algorithm outperforms Monte Carlo approach since a minimum of 300 full-wave EM simulations are required to estimate yield through Monte Carlo analysis.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Nfanyana, Ketshabile

Advisors dc:contributor.advisor

• Bodenstein, Leanne
• Meyer, Petrie

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/136101

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/136101